diff --git a/.comfyignore b/.comfyignore
index 46e2048..e8f24d9 100644
--- a/.comfyignore
+++ b/.comfyignore
@@ -7,6 +7,8 @@ __pycache__/
 .git/
 .github/
 docs/
+tests/
+GITHUB_RELEASE_CHECKLIST.md
 temp/
 tmp/
 
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
new file mode 100644
index 0000000..67ac0a6
--- /dev/null
+++ b/.github/workflows/ci.yml
@@ -0,0 +1,75 @@
+name: CI
+
+on:
+  push:
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  unit:
+    name: Python unit tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v6
+
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Install unit-test dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install --index-url https://download.pytorch.org/whl/cpu torch
+          python -m pip install "jsonschema>=4.26.0"
+
+      - name: Run unit tests
+        run: python -m unittest discover -s tests -p "test_*.py" -v
+
+  schema-readiness:
+    name: Schema readiness
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    env:
+      DG_PROGRESS: "0"
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v6
+
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Install schema-test dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install "jsonschema>=4.26.0"
+
+      - name: Run readiness and schema smoke
+        run: python scripts/smoke_prompt_builder.py
+
+  javascript:
+    name: JavaScript tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v6
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          package-manager-cache: false
+
+      - name: Check shipped JavaScript syntax
+        run: node --check web/js/model_loader.js
+
+      - name: Run JavaScript tests
+        run: node --test tests/*.mjs
diff --git a/.github/workflows/publish_action.yml b/.github/workflows/publish_action.yml
index 8d7d111..5f2867a 100644
--- a/.github/workflows/publish_action.yml
+++ b/.github/workflows/publish_action.yml
@@ -3,13 +3,16 @@ name: Publish to Comfy Registry
 on:
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   publish-node:
     name: Publish custom node to registry
     runs-on: ubuntu-latest
     steps:
       - name: Check out code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Publish custom node
         uses: Comfy-Org/publish-node-action@main
diff --git a/GITHUB_RELEASE_CHECKLIST.md b/GITHUB_RELEASE_CHECKLIST.md
index 9d52add..e31b04c 100644
--- a/GITHUB_RELEASE_CHECKLIST.md
+++ b/GITHUB_RELEASE_CHECKLIST.md
@@ -1,38 +1,69 @@
-# GitHub Release Checklist
+# GitHub and Comfy Registry Release Checklist
 
-Use this checklist before pushing the node pack to a standalone GitHub repo.
+Use this checklist from the standalone `ComfyUI-DiffusionGemmaPromptBuilder` repository before pushing or publishing a release.
 
-## Package Scope
+## Product Boundary
 
-- Exported DiffusionGemma nodes are limited to the nodes used by `examples/07_ltx23_character_motion_transfer.json`.
-- The example assets under `examples/assets/character_motion_transfer` are intended to ship with the repo.
-- The NVFP4 bridge is source code inside `nodes.py`; do not bundle model weights or generated tensor dumps.
+- Keep the stable Director Assistant surface at seven exports:
+  - five-node core: Model Loader, Context Hub, Target Profile, CoT Generator, JSON Splitter
+  - two companions: H3 Reference Context and Generation Gate
+- Keep the two SplatStage planners visibly experimental. Their mapping ids stay compatible, their display names end in `(Experimental)`, and their category is `prompt/diffusiongemma/experimental-motion-planning`.
+- Confirm `NODE_CLASS_MAPPINGS` and `NODE_DISPLAY_NAME_MAPPINGS` contain the same nine ids.
+- Treat `ltx_reframe.py` and the reframe workflows as unexported prototypes. Do not describe them as installed nodes unless the package entry point and frontend are deliberately restored together.
+- Do not bundle model weights or generated tensor dumps. The NVFP4 bridge remains source code in `nodes.py`.
 
-## Before Upload
+## Repository Preflight
 
-1. Confirm the GitHub repo URL in `pyproject.toml`.
-2. Confirm the Comfy Registry publisher id in `pyproject.toml` if you plan to publish through the Registry.
-3. Confirm `LICENSE` and the Python copyright headers still match the intended copyright holder and release policy.
-4. Run `C:\ComfyUI\.venv\Scripts\python.exe scripts\smoke_prompt_builder.py`.
-5. Run `C:\ComfyUI\.venv\Scripts\python.exe proof_gates.py --model-path C:\ComfyUI\models\LLM\diffusiongemma-26B-A4B-it-NVFP4`.
-6. Fresh-install the repo in a clean ComfyUI custom-nodes folder and confirm the five nodes appear.
-7. Copy `examples/assets/character_motion_transfer` to `ComfyUI/input/character_motion_transfer` and load the example workflow.
-8. Confirm `.comfyignore` excludes `*.mp4` and `*.gif` before publishing to the Comfy Registry.
+1. Run `git rev-parse --show-toplevel` and confirm it resolves to this package directory, not the parent ComfyUI checkout.
+2. Confirm `origin` is `https://github.com/exportAnything/ComfyUI-DiffusionGemmaPromptBuilder` and that the branch is based on the public repository history.
+3. Confirm the worktree is clean and the intended release commit is pushed.
+4. Confirm the repository URLs in `pyproject.toml` are current.
+5. Keep `[tool.comfy].PublisherId` exactly `exportanything`; this is the immutable Registry id, while `exportAnything` is its display name.
+6. Confirm `project.version` is newer than the latest Registry version. Version `0.1.0` is already published, so this release begins at `0.2.0`.
+7. Confirm `requires-comfyui = ">=0.30.1"` and record ComfyUI `0.30.1` as the tested baseline. The Registry field is a compatibility range, not a test-report field.
+8. Confirm `LICENSE` and the Python copyright headers match the intended release policy.
 
-## Files To Include
+## Automated Gates
 
-- `__init__.py`
-- `nodes.py`
-- `proof_gates.py`
-- `README.md`
-- `requirements.txt`
-- `pyproject.toml`
-- `.gitignore`
-- `.comfyignore`
-- `.github/workflows/publish_action.yml`
-- `web/`
-- `schemas/`
-- `scripts/`
-- `examples/07_ltx23_character_motion_transfer.json`
-- `examples/README.md`
-- `examples/assets/character_motion_transfer/`
+Run the same three gates enforced by `.github/workflows/ci.yml`:
+
+```text
+python -m unittest discover -s tests -p "test_*.py" -v
+python scripts/smoke_prompt_builder.py
+node --check web/js/model_loader.js
+node --test tests/*.mjs
+```
+
+- The unit suite must validate the exact nine-node surface, the Director/experimental category split, both bundled SplatStage schemas, and retained reframe helpers.
+- The readiness/schema smoke must finish with `Prompt Builder readiness/schema smoke passed`.
+- The JavaScript suite must exercise the shipped `model_loader.js`. No test may import the retired `ltx_reframe_layout.js`.
+- Require all three CI jobs to pass on the release commit.
+
+## Manual Runtime Gates
+
+1. Run the NVFP4 proof gate on a CUDA workstation with the real model:
+
+   ```text
+   python proof_gates.py --model-path C:\ComfyUI\models\LLM\diffusiongemma-26B-A4B-it-NVFP4
+   ```
+
+2. Fresh-install the repository in a clean ComfyUI `0.30.1` custom-nodes folder and confirm all nine nodes appear: seven Director nodes at the stable category and two SplatStage planners only under the experimental motion-planning category.
+3. Copy `examples/assets/character_motion_transfer` to `ComfyUI/input/character_motion_transfer` and load `examples/07_ltx23_character_motion_transfer.json`.
+4. Load the MiniMax H3 T2VA and Ref2VA examples and confirm the Generation Gate blocks invalid or empty prompt packets.
+5. Treat longer experimental SplatStage and reframe workflows as separate opt-in validation, not as evidence for the stable Director compatibility promise.
+
+## Files Tracked on GitHub
+
+- Package entry point and source: `__init__.py`, `nodes.py`, `ltx_reframe.py`, `proof_gates.py`
+- Metadata and policy: `README.md`, `LICENSE`, `requirements.txt`, `pyproject.toml`, `.gitignore`, `.comfyignore`, this checklist
+- Runtime contracts: `schemas/`, `web/`, `scripts/`
+- Verification: `tests/`, `.github/workflows/ci.yml`
+- Publishing: `.github/workflows/publish_action.yml`
+- Documentation, current examples, and GitHub demo assets: `docs/`, `examples/`
+
+## Registry Archive
+
+- Keep `.github/`, `docs/`, `tests/`, caches, temporary files, and large demo video/GIF files out of the Registry archive through `.comfyignore`.
+- Confirm both SplatStage schemas remain in the Registry archive because the experimental planners load them at runtime.
+- Inspect the archive before publishing and confirm it contains no model weights, local paths, logs, caches, or generated outputs.
+- Publish only after the GitHub release commit and version metadata agree.
diff --git a/README.md b/README.md
index 51a0d19..80932cb 100644
--- a/README.md
+++ b/README.md
@@ -4,21 +4,39 @@
 
 # ComfyUI DiffusionGemma Prompt Builder
 
-Five DiffusionGemma Director Assistant nodes for ComfyUI workflows that use DiffusionGemma as a multimodal prompt writer, not as the image or video sampler.
+DiffusionGemma Director provides a stable five-node prompt-authoring core for ComfyUI, with two stable Director companion nodes. The package also exports two experimental SplatStage motion-planning nodes. DiffusionGemma writes and validates prompts or blueprints; it is not the downstream image or video sampler.
 
-This release is packaged around the working LTX 2.3 character motion-transfer workflow in `examples/07_ltx23_character_motion_transfer.json`.
+The primary stable workflow is the working LTX 2.3 character motion-transfer graph in `examples/07_ltx23_character_motion_transfer.json`. Additional Director examples cover native MiniMax H3 T2VA and Ref2VA paths.
 
 ## Exported Nodes
 
-Only the nodes used by the packaged workflow are registered in ComfyUI:
+Nine nodes are registered in ComfyUI. Their stable and experimental surfaces are intentionally separate.
+
+### Stable Director Assistant
+
+The five-node core path used by the primary LTX workflow is:
 
 - `DiffusionGemmaModelLoader` - loads the local Transformers/NVFP4 DiffusionGemma runtime.
 - `DiffusionGemmaContextHub` - combines user text, optional image/video inputs, and optional `visual_description` into one `DG_CONTEXT`.
-- `DiffusionGemmaTargetProfile` - stores target generation settings such as LTX profile, audio mode, duration, aspect ratio, and negative prompt policy.
+- `DiffusionGemmaTargetProfile` - stores the LTX, Ideogram4, or MiniMax H3 target plus audio, duration, aspect ratio, and negative prompt settings.
 - `DiffusionGemmaCoTGenerator` - asks DiffusionGemma for the final prompt packet and returns final JSON plus debug outputs.
-- `DiffusionGemmaJSONSplitter` - turns the final JSON into LTX prompt, negative prompt, metadata, resolution, segment, and readiness outputs.
+- `DiffusionGemmaJSONSplitter` - turns the final JSON into LTX, Ideogram4, or MiniMax H3 prompt output plus metadata, resolution, segment, and readiness outputs.
 
-Older compatibility and utility classes remain in source where the final nodes still reuse helper behavior internally, but they are not exported as ComfyUI nodes in this package.
+Two stable companions extend that core without changing it:
+
+- `DiffusionGemmaH3ReferenceContext` - binds an ordered H3 asset-role manifest and optional picture/video evidence into a Ref2VA-aware `DG_CONTEXT`.
+- `DiffusionGemmaGenerationGate` - fails closed when validation says a prompt is not ready, preventing an invalid prompt from reaching a downstream generator.
+
+### Experimental Motion Planning
+
+These larger planners are exported for opt-in SplatStage work, but they are not part of the stable Director surface or the primary LTX workflow:
+
+- `DiffusionGemmaSplatStagePlanner` - writes and validates the strict SplatStage show blueprint.
+- `DiffusionGemmaSplatStageEfficientPlanner` - writes and validates the higher-variation efficient SplatStage blueprint.
+
+Both are labeled `(Experimental)` and appear under `prompt/diffusiongemma/experimental-motion-planning` in ComfyUI. Their schemas ship with this repository, while executing a resulting show still requires the appropriate downstream SplatStage workflow and nodes.
+
+Older compatibility and utility classes remain in source where exported nodes reuse their behavior internally, but they are not registered. The `ltx_reframe.py` module and LTX reframe workflows are retained as unexported prototypes; the retired `ltx_reframe_layout.js` frontend is not part of this release.
 
 ## Packaged Workflow
 
@@ -74,6 +92,89 @@ DiffusionGemma JSON Splitter.ready_for_generation -> generation gate or manual c
 
 Use `image_identity_video_control` on the Context Hub when the reference image should define the output subject while the video provides motion, pose, camera, depth, Canny/edge layout, blocking, and timing.
 
+For MiniMax H3, select `minimax_h3` on the same Target Profile node and wire that same `target_profile_config` to both the CoT Generator and JSON Splitter:
+
+```text
+DiffusionGemma Target Profile.target_profile_config -> DiffusionGemma CoT Generator.target_profile_config
+DiffusionGemma Target Profile.target_profile_config -> DiffusionGemma JSON Splitter.target_profile_config
+DiffusionGemma CoT Generator.final_json -> DiffusionGemma JSON Splitter.final_json
+DiffusionGemma JSON Splitter.minimax_h3_prompt -> MiniMax H3 text encoder prompt
+```
+
+The splitter appends `minimax_h3_prompt` after its existing outputs so saved LTX and Ideogram4 workflow output indices remain unchanged.
+
+## MiniMax H3 T2VA Prompt Profile
+
+With `minimax_h3_mode=t2va`, the MiniMax H3 profile turns a normal creative brief into H3's canonical text-to-video/audio representation. T2VA does not invent keyframe or full-reference labels when Director has not been given explicit reference roles. Its output preserves section breaks and uses exactly these three fields in order:
+
+```text
+integrated_multimodal_description: [Shot 1] ...
+
+overall_soundscape: ...
+
+non_diegetic_music: ...
+```
+
+The first shot has no timestamp. Later cuts use increasing instants such as `[Shot 2] At 00:03.500, the camera cuts to...`, all inside the selected duration. Hard cuts are the default; cross-dissolves, fades, and wipes are used only when the brief requests them. Synchronized visible action, speech, and scene sounds stay in `integrated_multimodal_description`; ambient sound and Foley belong in `overall_soundscape`; audience-only score belongs in `non_diegetic_music`, where instrumentation, tempo/rhythm, and dynamic development are described concretely instead of as an abstract mood.
+
+`target_duration_seconds` remains user controlled and has no H3-specific upper clamp. MiniMax documents 4-15 seconds as the standard H3 generation range, so that range is useful quality guidance, but Director does not clamp longer requests. A value such as 20 seconds is retained in the prompt contract; the connected H3 workflow's frame count, sampler settings, VRAM, and generation time determine whether it can be rendered successfully. `0` uses source-media duration when available, then an explicit duration written in the creative brief, and otherwise a 5-second H3 planning default.
+
+H3 does not expose a separate negative-conditioning prompt through its local text encoder. When MiniMax H3 is selected, Director keeps `negative_prompt` empty and expresses necessary exclusions concisely inside `integrated_multimodal_description` for T2VA or `detailed_description` for Ref2VA instead of inventing an unsupported negative channel.
+
+See [`examples/09_minimax_h3_t2va_director.json`](examples/09_minimax_h3_t2va_director.json) for a native MiniMax-H3 T2VA workflow. It shares one user-editable duration between Director and H3, routes the dedicated H3 prompt plus computed resolution into the generator, previews both the generated prompt and any validation block reasons, unloads DiffusionGemma before sampling, and is preset to the proven 20-second, 25-step configuration.
+
+## MiniMax H3 Ref2VA Prompt Profile
+
+Set `target_profile=minimax_h3`, `minimax_h3_mode=ref2va`, and build the context with `DiffusionGemmaH3ReferenceContext`. Its manifest is the asset-role contract for the native reference sockets, for example:
+
+```text
+<Picture 1>: protagonist identity, face, hair, body, wardrobe, and distinguishing traits; preserve fully.
+<Picture 2>: environment, palette, lighting, and 2D animation style; do not transfer pictured subject identity.
+<Audio 1>: synchronized soundtrack timing and impact-texture reference for <Video 1>; do not copy speech.
+<Video 1>: running motion, camera path, hard-cut rhythm, and landing mechanics; do not transfer performer identity.
+```
+
+Tags are one-based and numbered independently by type. Keep the manifest aligned with native H3: all pictures are presented first; a paired video soundtrack receives an Audio label immediately before its Video label; standalone audio follows the videos. Director supports the native local limits of nine pictures, three videos, three paired soundtracks, and three standalone audio inputs. It deliberately blocks an audio-only manifest because its Ref2VA compiler requires at least one visible reference source.
+
+Ref2VA output follows MiniMax's six-section Context-IR contract exactly:
+
+```text
+subject_definitions:
+...
+
+summary:
+...
+
+retention_analysis:
+...
+
+detailed_description:
+...
+
+overall_soundscape:
+...
+
+non_diegetic_music:
+...
+```
+
+`<Subject N>` means reusable visible content abstracted from one or more assets; it is not an attachment number. Pictures used only for identity, environment, costume, or style are cited inside Subject definitions. A standalone Picture definition is reserved for a concrete composition/storyboard anchor. In the local native Ref2VA path, an identity/style picture is not a guaranteed first- or last-frame pixel lock; use the separate FL2VA path when that is the intended control.
+
+The compiler requires official summary task types and retention markers, exact declared asset tags, consecutive Subject labels, a medium/style opening before `[Shot 1]`, a detailed chronological timeline, camera direction, reference use at the relevant moments, audio-layer fidelity, and a resolved final state. Invalid, missing, duplicate, or undeclared references block the splitter output instead of reaching H3.
+
+See [`examples/10_minimax_h3_ref2va_director.json`](examples/10_minimax_h3_ref2va_director.json) for the native two-picture sample. It wires both original pictures to `MiniMaxH3ReferenceToVideo`, sends an ordered analysis batch to Director, uses a full 2048-token prompt budget, and retains the user-selected 20-second, 481-frame, 25-step configuration without running a render automatically.
+
+Official references:
+
+- [MiniMax H3 repository and model overview](https://github.com/MiniMax-AI/MiniMax-H3)
+- [MiniMax H3 prompt-writing skill](https://github.com/MiniMax-AI/MiniMax-H3/blob/main/skills/h3-prompt-writing/SKILL.md)
+- [MiniMax H3 Base prompt guide](https://github.com/MiniMax-AI/MiniMax-H3/blob/main/skills/h3-prompt-writing/references/base-en.txt)
+- [MiniMax H3 full-reference prompt guide](https://github.com/MiniMax-AI/MiniMax-H3/blob/main/skills/h3-prompt-writing/references/ref-en.txt)
+
+## Compatibility
+
+The verified baseline is ComfyUI `0.30.1` with Python 3.11. Registry metadata declares `requires-comfyui = ">=0.30.1"`: this is the Registry's minimum-compatible-version constraint, while `0.30.1` remains the version exercised by the compatibility and release checks. Raise that floor when an exported node begins to require a newer ComfyUI API.
+
 ## Runtime Dependencies
 
 This pack assumes a working ComfyUI install with CUDA-capable PyTorch. The Python dependencies declared in `requirements.txt` are:
@@ -186,7 +287,7 @@ A good NVFP4 result reports:
 
 ## Packaging Notes
 
-`pyproject.toml` is included for GitHub and Comfy Registry readiness. The Comfy Registry requires a real publisher id; if `exportAnything` is not the exact Registry publisher id, update `[tool.comfy].PublisherId` before publishing.
+`pyproject.toml` is included for GitHub and Comfy Registry publishing. The immutable Registry publisher id is lowercase `exportanything` (display name `exportAnything`); keep that id and casing unchanged. Increment `project.version` for every Registry release.
 
 The current GitHub repository URL in metadata is:
 
diff --git a/docs/LTX_REFRAME_LAYOUT.md b/docs/LTX_REFRAME_LAYOUT.md
new file mode 100644
index 0000000..69c3461
--- /dev/null
+++ b/docs/LTX_REFRAME_LAYOUT.md
@@ -0,0 +1,29 @@
+# LTX Reframe Layout
+
+`LTX Reframe Layout` is a preparation node for LTX video outpainting. It does not load a model or sample a new video.
+
+## What the node does
+
+1. Select or upload a source video.
+2. Choose the final canvas aspect ratio and exact LTX-compatible dimensions.
+3. Drag the red source rectangle to reposition the video.
+4. Drag any red corner to resize it without changing the source aspect ratio. Scale below `1.0` zooms out; scale above `1.0` zooms in and crops the source at the target boundary.
+5. Execute the node to decode the video and produce padded frames, an outpaint mask, a source mask, the original frames, audio, FPS, and placement metadata.
+
+Target dimensions are snapped to multiples of 64. The source is never stretched. `source_scale` ranges from `0.1` to `4.0`: the fitted source is `1.0`, smaller values create outpaint space, and larger values crop/zoom into the source. In the preview, video beyond the target boundary is dimmed while the red resize handles remain visible. `position_x`, `position_y`, and `source_scale` are ordinary serialized ComfyUI widgets, so the visual placement is reproducible in saved workflows and API prompts.
+
+## LTX 2.3 connection
+
+- Connect `padded_frames` to `LTXVInpaintPreprocess.images`.
+- Connect `outpaint_mask` to `LTXVInpaintPreprocess.mask`.
+- Feed the preprocessed image into the official LTX 2.3 IC-LoRA outpaint guide path.
+- Use `source_mask` with `padded_frames` for a final source-preserving composite if the generated result changes the original region.
+- Connect `audio` and `fps` to the final video assembly nodes.
+
+The official local reference workflow is:
+
+`ComfyUI-LTXVideo/example_workflows/2.3/LTX-2.3_ICLoRA_Outpaint_Two_Stage_Distilled.json`
+
+## Current boundary
+
+The first version supports one static placement for the entire clip. Animated reframing/keyframes, rotation, perspective warping, per-shot layouts, and automatic subject-aware positioning are intentionally left for later nodes or workflow stages.
diff --git a/docs/PLATE_REPAIR_WORKBENCH_PROTOTYPE_PLAN.md b/docs/PLATE_REPAIR_WORKBENCH_PROTOTYPE_PLAN.md
new file mode 100644
index 0000000..e9b101d
--- /dev/null
+++ b/docs/PLATE_REPAIR_WORKBENCH_PROTOTYPE_PLAN.md
@@ -0,0 +1,484 @@
+# DiffusionGemma Plate Repair Workbench Prototype Plan
+
+Status: research-backed prototype plan, not implementation proof.
+
+Implementation target: `C:\ComfyUI\app\custom_nodes\ComfyUI-DiffusionGemma-LTX-Motion-Planner`.
+
+Prompt Builder role: keep the stable Director Assistant surface. The Prompt Builder should supply the local DiffusionGemma NVFP4 runtime and structured decision packets, but the video repair workbench belongs in the sibling Motion Planner pack.
+
+Research method: this plan was informed by three focused subagent probes:
+
+- LTX/Lightricks lane: confirmed LTX-2.3 IC-LoRA In/Outpainting and mask-aware LTXVideo nodes are the best LTX-native repair primitives; Motion Brush is timeline/control support, not the raw plate-mask bridge.
+- Wan/VACE lane: confirmed `WanVaceToVideo` and official VACE semantics fit masked repair plates; local VACE model availability still needs verification.
+- ComfyUI glue lane: confirmed local IO, SAM3, crop/uncrop, batch composite, alpha, and pyramid-blend primitives already cover most non-model work; missing pieces are thin deterministic adapters and QA.
+
+## Thesis
+
+The repair loop should not behave like a serial full-video retake agent. It should behave like a small VFX plate system:
+
+```text
+Immutable source plate
+-> DiffusionGemma issue inventory
+-> plate plan
+-> per-issue temporal masks
+-> per-issue repair plates from LTX or Wan
+-> deterministic masked composite over source
+-> critic verifies the composite and updates issue status
+```
+
+This is scientifically plausible because it converts an underconstrained generative loop into a sequence of explicit hypotheses:
+
+1. A defect exists at a named time window and layer.
+2. The defect has a maskable target.
+3. A mask covers the intended damaged pixels and not stable pixels.
+4. A backend can generate a candidate for only that plate.
+5. The compositor preserves unmasked source pixels.
+6. The critic evaluates the composite, not a full-video regeneration chain.
+
+The architecture should prefer visible schema-bound nodes over hidden autonomous loops. DiffusionGemma is the orchestrator; ComfyUI nodes are the subagents.
+
+## Evidence Summary
+
+### LTX/Lightricks
+
+LTX provides relevant primitives for this architecture, especially through IC-LoRA in/outpainting and ComfyUI-LTXVideo utility nodes.
+
+External evidence:
+
+- `https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-In-Outpainting`
+  - The model card describes inpainting and outpainting on top of LTX-2.3-22B.
+  - It names the training type as video-to-video, mask-conditioned IC-LoRA.
+  - It uses a reference video plus binary mask, generating the masked region while leaving the unmasked area intact.
+  - It describes a two-stage pipeline with boundary refinement and Laplacian blending.
+  - It warns to describe only the masked/outpainted region, not the full scene.
+- `https://docs.ltx.video/open-source-model/usage-guides/ic-lo-ra`
+  - IC-LoRAs are task-specific, reference-driven adapters for structural control, VFX, restoration, and creative transforms.
+  - The guide names `LTX Add Video IC-LoRA Guide` and `LTX Add Video IC-LoRA Guide Advanced`; the advanced path includes mask-aware conditioning.
+- `https://docs.ltx.video/open-source-model/integration-tools/ltx-comfy-ui-nodes`
+  - The official LTX ComfyUI node documentation names the same inpainting and blend primitives used by the local node pack.
+- `https://github.com/Lightricks/ComfyUI-LTXVideo`
+  - The repository hosts LTX-2 ComfyUI custom nodes and workflows, including IC-LoRA workflows and LTX-2.3 adapter assets.
+
+Local evidence:
+
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\vanish_nodes.py:8` registers `LTXVDilateVideoMask`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\vanish_nodes.py:96` registers `LTXVInpaintPreprocess`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\masks.py:11` registers `LTXVPreprocessMasks`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\latents.py:209` registers `LTXVSetVideoLatentNoiseMasks`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\pyramid_blending.py:178` registers `LTXVLaplacianPyramidBlend`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\guide.py:258` registers `LTXVAddGuideAdvancedAttention`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\iclora.py:261` registers `LTXAddVideoICLoRAGuideAdvanced`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\example_workflows\2.3\LTX-2.3_ICLoRA_Inpaint_Two_Stage_Distilled.json` exists locally.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\example_workflows\2.3\LTX-2.3_ICLoRA_Outpaint_Two_Stage_Distilled.json` exists locally.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\example_workflows\2.3\LTX-2.3_ICLoRA_Motion_Track_Distilled.json` exists locally.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-LTXVideo\example_workflows\2.3\LTX-2.3_ICLoRA_Motion_Brush_Distilled.json` exists locally.
+
+Important limitation:
+
+`ComfyUI-LTX-Director-Motion-Brush` is not the raw mask compositor. Motion Brush is useful for timeline and guide payloads, but the direct mask seam is in LTXVideo guide, IC-LoRA, mask, latent-mask, and blend nodes.
+
+Keep three mask concepts separate:
+
+- `attention_mask`: controls guide/conditioning influence.
+- latent or denoise/noise mask: controls where generation may rewrite latent content.
+- composite or blend mask: controls final pixel replacement over the source plate.
+
+### Wan/VACE
+
+Wan VACE is also a strong backend candidate for repair plates. It may be better than LTX for some masked object replacement and region-regeneration tasks.
+
+External evidence:
+
+- `https://github.com/ali-vilab/VACE`
+  - VACE describes itself as all-in-one video creation and editing.
+  - It explicitly includes reference-to-video generation, video-to-video editing, and masked video-to-video editing.
+  - Its preprocessing path produces `src_video`, `src_mask`, and `src_ref_images`, which maps cleanly onto a plate manifest.
+- `https://docs.comfy.org/built-in-nodes/conditioning/video-models/wan-vace-to-video`
+  - The ComfyUI built-in `WanVaceToVideo` node accepts `control_video`, `control_masks`, and `reference_image`.
+- `https://docs.comfy.org/tutorials/video/wan/vace`
+  - ComfyUI's Wan VACE page describes VACE as supporting text, images, video, masks, and control signals; it names local replacement through masks.
+- `https://github.com/stuttlepress/ComfyUI-Wan-VACE-Prep`
+  - Its `Wan VACE Inpaint` node prepares a control video and mask for `WanVaceToVideo.control_video` and `control_masks`, replacing masked pixels with neutral gray and preserving the mask polarity.
+
+Local evidence:
+
+- `C:\ComfyUI\app\comfy_extras\nodes_wan.py:285` defines `WanVaceToVideo`.
+- `C:\ComfyUI\app\comfy_extras\nodes_wan.py:301` exposes `control_video`.
+- `C:\ComfyUI\app\comfy_extras\nodes_wan.py:302` exposes `control_masks`.
+- `C:\ComfyUI\app\comfy_extras\nodes_wan.py:303` exposes `reference_image`.
+- `C:\ComfyUI\app\comfy_extras\nodes_wan.py:365` and `:366` attach `vace_frames`, `vace_mask`, and `vace_strength` into conditioning.
+- `C:\ComfyUI\app\custom_nodes\comfyui_fill-nodes\nodes\wip\FL_WanVaceToVideoMultiRef.py:17` provides a local multi-reference adapter.
+
+Important limitation:
+
+Wan VACE should be treated as a repair backend, not as the overall autonomous planner. The plate contract should be independent of the backend so LTX and Wan can be compared on the same issue packet.
+
+Local model status:
+
+- Current local scan found Wan diffusion models in `C:\ComfyUI\app\models\diffusion_models`, specifically `wan2.2_bernini_r_low_noise_mxfp8.safetensors` and `wan2.2_bernini_r_high_noise_mxfp8.safetensors`.
+- The same scan did not find an obvious VACE diffusion model. Wan/VACE should therefore be designed into the prototype, but LTX is the likely first runnable backend unless VACE weights are added.
+- `ComfyUI-Wan-VACE-Prep` is useful prior art, but it is not installed locally. A tiny Wan prep node may be justified if we want to avoid an external custom-node dependency.
+
+### ComfyUI Glue
+
+The non-model pieces also mostly exist.
+
+Local evidence:
+
+- `C:\ComfyUI\app\custom_nodes\comfyui-videohelpersuite\videohelpersuite\load_video_nodes.py:476` provides local video loading into frame batches.
+- `C:\ComfyUI\app\custom_nodes\comfyui-videohelpersuite\videohelpersuite\nodes.py:235` provides `VHS_VideoCombine`.
+- `C:\ComfyUI\app\custom_nodes\comfyui-kjnodes\nodes\image_nodes.py:1962` provides `GetImageRangeFromBatch`.
+- `C:\ComfyUI\app\comfy_extras\nodes_sam3.py:88` provides native SAM3 node definitions.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-DiffusionGemma-LTX-Motion-Planner\nodes.py:3471` defines `DiffusionGemmaLTXQualityCritic`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-DiffusionGemma-LTX-Motion-Planner\nodes.py:3666` defines `DiffusionGemmaLTXIssueScheduler`.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-DiffusionGemma-LTX-Motion-Planner\nodes.py:3771` rejects scheduled issues unless they are both time-localized and maskable.
+- `C:\ComfyUI\app\custom_nodes\ComfyUI-DiffusionGemma-LTX-Motion-Planner\nodes.py:3933` defines `DiffusionGemmaLTXTimelineMaskPad`.
+- `C:\ComfyUI\app\custom_nodes\comfyui_essentials\image.py:206` defines `ImageCompositeFromMaskBatch`.
+- `C:\ComfyUI\app\custom_nodes\comfyui_fill-nodes\nodes\utility\FL_PasteByMask.py:65` defines `FL_PasteByMask`.
+- `C:\ComfyUI\app\custom_nodes\comfyui_fill-nodes\nodes\utility\FL_VideoCropNStitch.py:6` defines `FL_VideoCropMask`.
+- `C:\ComfyUI\app\custom_nodes\comfyui_fill-nodes\nodes\utility\FL_VideoCropNStitch.py:135` defines `FL_VideoRecompose`.
+- `C:\ComfyUI\app\custom_nodes\comfyui-kjnodes\nodes\batchcrop_nodes.py:26` defines `BatchCropFromMask`.
+- `C:\ComfyUI\app\custom_nodes\comfyui-kjnodes\nodes\batchcrop_nodes.py:156` defines `BatchUncrop`.
+- `C:\ComfyUI\app\custom_nodes\comfyui-kjnodes\nodes\batchcrop_nodes.py:253` defines `BatchCropFromMaskAdvanced`.
+- `C:\ComfyUI\app\custom_nodes\comfyui-kjnodes\nodes\batchcrop_nodes.py:535` defines `BatchUncropAdvanced`.
+
+Composite caution:
+
+`FL_VideoRecompose` is useful, but it replaces the whole crop rectangle. It should not be the scientific preservation proof. The final proof should still use an explicit mask composite over the immutable source plate.
+
+## Proposed Node Pack Shape
+
+### 1. DiffusionGemma Plate Critic
+
+Can start as an extension of `DiffusionGemmaLTXQualityCritic`.
+
+Inputs:
+
+- original source evidence
+- generated/composited candidate video
+- user intent
+- optional current plate manifest
+- optional previous issue statuses
+
+Outputs:
+
+- `issue_inventory_json`
+- `plate_plan_json`
+- `blocked_or_advisory_reasons`
+- `verification_level`
+
+Each issue should include:
+
+- `issue_id`
+- `category`
+- `severity`
+- `confidence`
+- `time_window`
+- `target_layer`
+- `z_order`
+- `maskable_target`
+- `sam_prompt`
+- `bbox_hint`
+- `repair_instruction`
+- `preserve_prompt`
+- `backend_preference`
+- `evidence_notes`
+- `status`
+
+### 2. Plate Plan Compiler
+
+New thin deterministic node.
+
+Purpose:
+
+Convert issue inventory into executable plate jobs. This is the place to enforce scientific constraints before any generator runs.
+
+Required invariants:
+
+- every executable job has `time_localized=true`
+- every executable job has `maskable_target=true`
+- every executable job references the immutable `base_plate_id`
+- every executable job declares mask polarity
+- every executable job declares backend and prompt scope
+- no job may use a previous generated/composited video as `base_plate`
+
+Outputs:
+
+- `plate_manifest_json`
+- `next_plate_job_json`
+- `remaining_plate_jobs_json`
+- `source_plate_file`
+- `repair_prompt`
+- `preserve_prompt`
+- `backend_kind`
+- `z_order`
+- `composite_mode`
+
+### 3. Mask Acquisition Lane
+
+Reuse existing nodes:
+
+```text
+Plate job
+-> Issue Scheduler or Plate Plan Compiler lane output
+-> SAM3 Detect / SAM3 Video Track / SAM3 Track to Mask
+-> DiffusionGemma LTX Timeline Mask Pad
+-> mask processors or LTXVDilateVideoMask
+```
+
+This lane should produce:
+
+- `window_mask`
+- `timeline_mask`
+- `mask_debug_json`
+- `nonzero_frame_count`
+- `mask_quality_flags`
+
+### 4. Repair Backend Adapter
+
+New thin routing node or two separate backend-specific adapter nodes.
+
+Backend A: LTX IC-LoRA in/outpainting.
+
+Likely node chain:
+
+```text
+base source frames + timeline mask
+-> LTXVInpaintPreprocess
+-> LTX IC-LoRA loader / Add Video IC-LoRA Guide Advanced
+-> sampler/decode
+-> repair_plate_frames
+```
+
+Alternative LTX guide path:
+
+```text
+timeline_mask
+-> LTXVAddGuideAdvancedAttention.attention_mask
+or LTXAddVideoICLoRAGuideAdvanced.attention_mask
+```
+
+Backend B: Wan VACE.
+
+Likely node chain:
+
+```text
+base source frames + timeline mask
+-> VACE control video prep
+-> WanVaceToVideo.control_video
+-> WanVaceToVideo.control_masks
+-> Wan sampler/decode
+-> repair_plate_frames
+```
+
+If `ComfyUI-Wan-VACE-Prep` is not installed, the first Wan adapter can be tiny:
+
+- input `IMAGE` source frames and `MASK`
+- output `control_video` where masked pixels become neutral 0.5 gray
+- output `control_masks` with white indicating regenerate
+- output width, height, length
+
+Backend C: no-generator deterministic proof.
+
+Before either LTX or Wan generation, the router should support a synthetic or passthrough repair plate. This lets tests prove mask polarity, frame alignment, and unmasked preservation without confusing generator errors with orchestration errors.
+
+### 5. Plate Composite
+
+Start with existing nodes:
+
+- `ImageCompositeFromMaskBatch`
+- `LTXVLaplacianPyramidBlend`
+- `FL_PasteByMask`
+- `FL_VideoRecompose`
+
+Add a custom node only if existing nodes fail frame-count, z-order, or holdout semantics.
+
+Potential new node:
+
+`DiffusionGemmaPlateComposite`
+
+Inputs:
+
+- `base_frames`
+- `repair_frames`
+- `timeline_mask`
+- `plate_manifest_json`
+- `z_order`
+- `feather_or_blend_mode`
+
+Outputs:
+
+- `composited_frames`
+- `composite_debug_json`
+- `unmasked_pixel_delta`
+- `mask_coverage_stats`
+
+Non-negotiable test:
+
+Pixels outside the active mask must match the immutable source plate exactly, or within a declared tolerance if a blend operation intentionally softens boundaries.
+
+Recommended convention:
+
+- project mask polarity: `1 = repair/generate`, `0 = preserve source`
+- exact composite: `ImageCompositeFromMaskBatch(image_from=source, image_to=repair, mask=timeline_mask)`
+- LTX Laplacian blend: `image_a=repair`, `image_b=source`, `mask=timeline_mask`
+- feathered/pyramid modes may allow tolerance only inside a declared dilated boundary band
+
+### 6. Plate Verifier
+
+Can start as a mode of `DiffusionGemmaLTXQualityCritic`.
+
+It should evaluate:
+
+- issue fixed or not fixed
+- new artifacts introduced or not
+- unmasked preservation respected or not
+- mask leakage at boundary
+- temporal consistency inside the repaired window
+- whether remaining issues should proceed
+
+Outputs:
+
+- `issue_status_json`
+- `fixed_issue_ids`
+- `failed_issue_ids`
+- `regression_issue_ids`
+- `continue_next_plate`
+- `verification_level`
+
+## Minimal Prototype Workflow
+
+### Phase 0: Deterministic Composite Proof
+
+No generator.
+
+Use synthetic tensors or a short loaded clip:
+
+1. Load base video frames.
+2. Create a simple moving mask or use SAM3 output.
+3. Create a synthetic repair plate, for example solid color or blurred patch.
+4. Composite repair over source.
+5. Assert unmasked pixels are preserved.
+6. Assert frame counts and mask windows match.
+
+This proves the plate invariant before any model gets blamed.
+
+### Phase 1: One-Issue LTX Plate
+
+1. Source evidence from original video.
+2. DiffusionGemma emits one issue with time window, layer, mask prompt, and repair instruction.
+3. SAM3 generates the window mask.
+4. Timeline Mask Pad expands it to full frame count.
+5. LTX in/outpainting or IC-LoRA guide generates one repair plate.
+6. Composite repair over immutable source.
+7. Critic verifies only that issue.
+
+Success condition:
+
+The final output is source pixels plus masked repair, not generation two of generation one.
+
+### Phase 2: Three Manual Sequential Lanes
+
+Expose three lanes for `issue_index=0`, `1`, and `2`.
+
+This keeps debugging visible and avoids pretending the automatic loop is solved.
+
+Each lane writes:
+
+- selected issue
+- mask preview
+- repair plate preview
+- composite preview
+- issue status
+
+### Phase 3: LTX vs Wan Backend Comparison
+
+For the same issue packet and same mask:
+
+1. Generate repair plate with LTX.
+2. Generate repair plate with Wan VACE.
+3. Composite both over the same immutable source.
+4. Critic compares composites using the same criteria.
+
+This creates a scientific backend-selection loop. The orchestrator can later learn that LTX is better for some classes and Wan is better for others.
+
+## Backend Selection Hypotheses
+
+Start with these priors and let evidence update them:
+
+- LTX IC-LoRA In-Outpainting: best first choice for masked fill, boundary-aware repair, canvas extension, and repairs that should preserve LTX timing/style.
+- LTX attention-mask guide path: useful when the repair should bias conditioning but not fully replace pixels.
+- Wan VACE: strong alternate for local replacement, background replacement, and larger masked video-to-video edits.
+- Motion Brush: useful for timeline, retake, and motion-control payloads, but not the raw mask-to-pixel compositor.
+- Generic image/composite nodes: should own final pixel preservation and plate merge whenever possible.
+
+## Scientific Acceptance Criteria
+
+The prototype is not accepted until these are true:
+
+1. Source plate immutability is explicit in every plate manifest.
+2. Every executable repair job has a time window and maskable target.
+3. Timeline masks fail fast on frame-count mismatch.
+4. Unmasked source pixels are preserved by deterministic composite.
+5. Prompts for repair backends describe only the repaired region.
+6. The critic updates per-issue status rather than one overall retake score.
+7. A generated or composited output is never reused as the next base plate.
+8. LTX and Wan backends can be tested against the same plate job.
+9. Fallback or metadata-only critic reports cannot green-light execution.
+10. The graph exposes intermediate plate, mask, and composite previews.
+
+## What Not To Build
+
+Do not build:
+
+- a hidden all-in-one retake agent
+- a new SAM implementation
+- a custom compositor before testing `ImageCompositeFromMaskBatch`, `LTXVLaplacianPyramidBlend`, and `FL_PasteByMask`
+- a custom Wan VACE inpaint implementation if a small control-video prep adapter is enough
+- a prompt-only multi-issue repair chain that collapses layers into one dense retake prompt
+- a loop where pass two treats pass one's defects as new source truth
+
+## Required New Artifacts
+
+Likely needed:
+
+- `schemas/plate_plan.schema.json`
+- `schemas/plate_job.schema.json`
+- `schemas/plate_result.schema.json`
+- `schemas/issue_status.schema.json`
+- smoke test for synthetic composite preservation
+- smoke test for scheduler sequential lanes
+- smoke test for vague issue refusal
+- smoke test for strict mask/window mismatch
+- optional example workflow: `examples/layered_repair_workbench_poc.json`
+
+Likely new nodes:
+
+- `DiffusionGemma Source Plate Manifest`
+- `DiffusionGemma LTX Plate Plan Compiler`
+- `DiffusionGemma LTX Plate Job Router`
+- `DiffusionGemma Plate Composite QA`
+- `DiffusionGemma Plate Composite` only if existing composites do not satisfy z-order/holdout/testing needs
+- `DiffusionGemma Wan VACE Prep` only if installing or relying on external Wan prep nodes is undesirable
+
+## Open Verification Tasks
+
+Before implementation:
+
+1. Load the official LTX in/outpainting workflow locally and record exact node sequence.
+2. Confirm local `ComfyUI-LTXVideo` in/outpainting workflows run with the installed model files, not only that the JSON exists.
+3. Confirm whether local ComfyUI has native SAM3 model files available, not just node classes.
+4. Confirm whether Wan VACE model files are installed and runnable in this ComfyUI environment.
+5. Decide whether the first prototype should target LTX only, or LTX plus a Wan A/B backend.
+6. Build a no-generator composite smoke before touching generation.
+
+## Handoff Summary
+
+The architecture is not vibes:
+
+- LTX/Lightricks provides mask-conditioned in/outpainting and blend primitives.
+- Wan VACE provides masked video-to-video/control-mask primitives.
+- ComfyUI already provides SAM3 masks, crop/recompose, batch compositing, and LTX/Wan sockets.
+- DiffusionGemma's unique role is not to invent pixels. Its role is to produce verified issue inventory, plate plans, backend routing, and critic status updates.
+
+The next implementation should be a visible plate workbench, not a closed-loop full-video retake chain.
diff --git a/examples/08_ltx23_reframe_iclora_outpaint_two_stage.json b/examples/08_ltx23_reframe_iclora_outpaint_two_stage.json
new file mode 100644
index 0000000..1b033ae
--- /dev/null
+++ b/examples/08_ltx23_reframe_iclora_outpaint_two_stage.json
@@ -0,0 +1,4484 @@
+{
+  "id": "90172662-3bf3-57bf-bd6d-0a13de3dc8e4",
+  "revision": 0,
+  "last_node_id": 5401,
+  "last_link_id": 14450,
+  "nodes": [
+    {
+      "id": 5023,
+      "type": "LTXAVTextEncoderLoader",
+      "pos": [
+        -3517.2747393631416,
+        3351.2683784179994
+      ],
+      "size": [
+        318.444921875,
+        122
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            13459,
+            13460
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "LTXAVTextEncoderLoader"
+      },
+      "widgets_values": [
+        "gemma_3_12B_it_fp4_mixed.safetensors",
+        "ltx-2.3-22b-dev.safetensors",
+        "default"
+      ]
+    },
+    {
+      "id": 5022,
+      "type": "PrimitiveString",
+      "pos": [
+        -3501.28471605786,
+        3031.3030061080112
+      ],
+      "size": [
+        270,
+        58
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 4,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [
+            13457,
+            13458
+          ]
+        }
+      ],
+      "title": "LTX API KEY",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "PrimitiveString"
+      },
+      "widgets_values": [
+        ""
+      ]
+    },
+    {
+      "id": 3940,
+      "type": "CheckpointLoaderSimple",
+      "pos": [
+        -4599.826307240758,
+        2753.3368885519267
+      ],
+      "size": [
+        383.4002295473938,
+        108.51948797992713
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            13217
+          ]
+        },
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": []
+        },
+        {
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            13279,
+            13348,
+            13699,
+            14392,
+            14393,
+            14413
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.56",
+        "Node name for S&R": "CheckpointLoaderSimple"
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-dev.safetensors"
+      ]
+    },
+    {
+      "id": 5193,
+      "type": "MarkdownNote",
+      "pos": [
+        -4564.234267550487,
+        2919.780044411716
+      ],
+      "size": [
+        323.71820098153694,
+        102.23102769520756
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "The `latent_downscale_factor` output from the IC-LoRA Loader is intentionally ignored. The inpainting LoRA was trained with `reference_downscale_factor = 1` (full resolution), so no spatial downscaling of the reference is needed. For reference on handling other values, see `LTX-2.3_ICLoRA_Union_Control_Distilled.json`"
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 1241,
+      "type": "LTXVConditioning",
+      "pos": [
+        -2809.701994778083,
+        3351.4001060906317
+      ],
+      "size": [
+        210,
+        78
+      ],
+      "flags": {},
+      "order": 29,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 7065
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 7652
+        },
+        {
+          "name": "frame_rate",
+          "type": "FLOAT",
+          "widget": {
+            "name": "frame_rate"
+          },
+          "link": 13889
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            13701
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "slot_index": 1,
+          "links": [
+            13700
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "LTXVConditioning"
+      },
+      "widgets_values": [
+        24
+      ]
+    },
+    {
+      "id": 2004,
+      "type": "LoadImage",
+      "pos": [
+        -4221.497658697313,
+        3166.988989590151
+      ],
+      "size": [
+        274.080078125,
+        314.000244140625
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            13851
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.56",
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "example.png",
+        "image"
+      ]
+    },
+    {
+      "id": 5360,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        -3137.729014735421,
+        3802.617334027078
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 32,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14364
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14360,
+            14365,
+            14370
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale to multiple",
+        64,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5365,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        -3133.6568807035846,
+        3970.919441086199
+      ],
+      "size": [
+        270,
+        126
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 33,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14369
+        },
+        {
+          "label": "match",
+          "name": "resize_type.match",
+          "type": "IMAGE,MASK",
+          "link": 14370
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "MASK",
+          "links": [
+            14366,
+            14368,
+            14376
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "match size",
+        "center",
+        "area"
+      ]
+    },
+    {
+      "id": 5361,
+      "type": "LTXVInpaintPreprocess",
+      "pos": [
+        -2836.4256900607797,
+        3878.7764040782017
+      ],
+      "size": [
+        248.4576171875,
+        46
+      ],
+      "flags": {},
+      "order": 34,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14365
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14366
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14374
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVInpaintPreprocess"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5357,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        -2528.116627857796,
+        3808.6039390018286
+      ],
+      "size": [
+        270,
+        126
+      ],
+      "flags": {},
+      "order": 36,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14360
+        },
+        {
+          "label": "match",
+          "name": "resize_type.match",
+          "type": "IMAGE,MASK",
+          "link": 14361
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14362
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "match size",
+        "center",
+        "area"
+      ]
+    },
+    {
+      "id": 2483,
+      "type": "CLIPTextEncode",
+      "pos": [
+        -3148.5185252516953,
+        3345.9814910205873
+      ],
+      "size": [
+        298.24996904371346,
+        88
+      ],
+      "flags": {},
+      "order": 18,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 13459
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            7065
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Positive Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "A cinematic live-action scene continuing naturally beyond the source frame, with a coherent environment, realistic lighting, consistent subjects and camera motion, seamless edges, and stable fine detail."
+      ],
+      "color": "#232",
+      "bgcolor": "#353"
+    },
+    {
+      "id": 2612,
+      "type": "CLIPTextEncode",
+      "pos": [
+        -3157.621893338516,
+        3500.304316455244
+      ],
+      "size": [
+        309.5560766015442,
+        120.13641397900165
+      ],
+      "flags": {},
+      "order": 19,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 13460
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            7652
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Negative Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "pc game, console game, video game, cartoon, childish, ugly"
+      ],
+      "color": "#322",
+      "bgcolor": "#533"
+    },
+    {
+      "id": 4832,
+      "type": "RandomNoise",
+      "pos": [
+        -1282.2004731850486,
+        3136.4885401533693
+      ],
+      "size": [
+        279.7478557810682,
+        82
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [
+            13631
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        42,
+        "fixed"
+      ]
+    },
+    {
+      "id": 4851,
+      "type": "VAEDecodeTiled",
+      "pos": [
+        -275.13453421434303,
+        3220.9811784851136
+      ],
+      "size": [
+        270,
+        150
+      ],
+      "flags": {},
+      "order": 50,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 13412
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 13348
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14193
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "VAEDecodeTiled"
+      },
+      "widgets_values": [
+        512,
+        64,
+        512,
+        64
+      ]
+    },
+    {
+      "id": 5088,
+      "type": "PrimitiveBoolean",
+      "pos": [
+        -3900.329183769558,
+        3193.4571415039886
+      ],
+      "size": [
+        270,
+        58
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "BOOLEAN",
+          "type": "BOOLEAN",
+          "links": [
+            14318,
+            14386
+          ]
+        }
+      ],
+      "title": "bypass_i2v",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "PrimitiveBoolean"
+      },
+      "widgets_values": [
+        true
+      ]
+    },
+    {
+      "id": 5188,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        -3900.2419246429245,
+        3326.7561203909813
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 23,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 13851
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14387,
+            14416
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.16.0",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale longer dimension",
+        1024,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5025,
+      "type": "ManualSigmas",
+      "pos": [
+        -1275.2858877128047,
+        3539.075254238973
+      ],
+      "size": [
+        499.94633283206986,
+        67.3573576655076
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            13664
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "ManualSigmas"
+      },
+      "widgets_values": [
+        "1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0"
+      ]
+    },
+    {
+      "id": 4828,
+      "type": "CFGGuider",
+      "pos": [
+        -1275.2377178837455,
+        3258.169012259166
+      ],
+      "size": [
+        270,
+        114
+      ],
+      "flags": {},
+      "order": 42,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 13401
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 13786
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 13787
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            13632
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ]
+    },
+    {
+      "id": 5356,
+      "type": "MarkdownNote",
+      "pos": [
+        -3503.5333421952055,
+        3992.0569650957063
+      ],
+      "size": [
+        330.26254735838165,
+        121.31629268413144
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "Note",
+      "properties": {},
+      "widgets_values": [
+        "### Reframe Setup\n\nUse **LTX Reframe Layout** to choose the final canvas and place the source. `source_scale < 1` zooms out for outpainting; `source_scale > 1` zooms in and crops. The example starts at **576 \u00d7 1024**, 121 frames, with audio, so it is suitable for a first end-to-end test."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 5054,
+      "type": "GetImageSize",
+      "pos": [
+        -2537.9363511080323,
+        3413.0334463354366
+      ],
+      "size": [
+        163.19199846392712,
+        124
+      ],
+      "flags": {},
+      "order": 38,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14371
+        }
+      ],
+      "outputs": [
+        {
+          "name": "width",
+          "type": "INT",
+          "links": [
+            13531
+          ]
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "links": [
+            13532
+          ]
+        },
+        {
+          "name": "batch_size",
+          "type": "INT",
+          "links": [
+            13627,
+            14447,
+            14448
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.16.0",
+        "Node name for S&R": "GetImageSize"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5164,
+      "type": "MarkdownNote",
+      "pos": [
+        -4703.860696244165,
+        3109.620866704518
+      ],
+      "size": [
+        429.8300275342899,
+        372.1472188722282
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "# How to Use Outpainting\n\nThis workflow supports **outpainting** (extending the canvas beyond the original frame). You provide a mask telling the model where to generate, and an optional prompt describing what should be there.\n\n## 1. No prompt = the simple case\n\n- without a prompt \u2192 just extends the existing canvas. The model fills the new area to match the surrounding scene.\n\n\n## 2. Prompts describe the scene, not the edit\n\nIt's not an edit tool. Write a regular T2V-style prompt describing the out-painting region.\n\n\n\n## 3. Green leftovers\n\nThe pipeline composites green under the mask before diffusion \u2014 sometimes traces leak through at mask edges. Try a bigger dilation, a different seed, or re-encode the mask losslessly or the blending parameter is very important."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 5013,
+      "type": "LTXVCropGuides",
+      "pos": [
+        -548.5261857816255,
+        3180.975716856467
+      ],
+      "size": [
+        187.5,
+        100.83333333333334
+      ],
+      "flags": {},
+      "order": 47,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 13694
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 13696
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14425
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            14262,
+            14443
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            14263,
+            14442
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            13412
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.75",
+        "Node name for S&R": "LTXVCropGuides"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5209,
+      "type": "CFGGuider",
+      "pos": [
+        749.917968641874,
+        3161.2856570924046
+      ],
+      "size": [
+        270,
+        98
+      ],
+      "flags": {},
+      "order": 48,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 13911
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14262
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14263
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            13909
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ]
+    },
+    {
+      "id": 5210,
+      "type": "RandomNoise",
+      "pos": [
+        750.7124305557006,
+        3037.1803832063797
+      ],
+      "size": [
+        270,
+        82
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 10,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [
+            13908
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.56",
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        42,
+        "fixed"
+      ]
+    },
+    {
+      "id": 5266,
+      "type": "LTXVLaplacianPyramidBlend",
+      "pos": [
+        -275.5732882228261,
+        3431.880734673949
+      ],
+      "size": [
+        337.65390625,
+        122
+      ],
+      "flags": {},
+      "order": 51,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_a",
+          "type": "IMAGE",
+          "link": 14193
+        },
+        {
+          "name": "image_b",
+          "type": "IMAGE",
+          "link": 14439
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14450
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14408
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVLaplacianPyramidBlend"
+      },
+      "widgets_values": [
+        true,
+        5
+      ]
+    },
+    {
+      "id": 5211,
+      "type": "ManualSigmas",
+      "pos": [
+        788.7312466477023,
+        3434.9579610502183
+      ],
+      "size": [
+        274.0120536645418,
+        58
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            13907
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "ManualSigmas"
+      },
+      "widgets_values": [
+        "0.7250, 0.4219, 0.0"
+      ]
+    },
+    {
+      "id": 5379,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        139.7628663326855,
+        3544.2029223199124
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 52,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14408
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14409
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale by multiplier",
+        2,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5267,
+      "type": "VAEEncodeTiled",
+      "pos": [
+        424.32499173298754,
+        3546.493495511004
+      ],
+      "size": [
+        270,
+        150
+      ],
+      "flags": {},
+      "order": 53,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "pixels",
+          "type": "IMAGE",
+          "link": 14409
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14392
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14410
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "VAEEncodeTiled"
+      },
+      "widgets_values": [
+        512,
+        64,
+        512,
+        64
+      ]
+    },
+    {
+      "id": 5215,
+      "type": "KSamplerSelect",
+      "pos": [
+        777.0958692052785,
+        3310.003437337445
+      ],
+      "size": [
+        270,
+        74
+      ],
+      "flags": {},
+      "order": 13,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            13910
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler_cfg_pp"
+      ]
+    },
+    {
+      "id": 4831,
+      "type": "KSamplerSelect",
+      "pos": [
+        -1274.908203992836,
+        3414.0866355538983
+      ],
+      "size": [
+        270,
+        74
+      ],
+      "flags": {},
+      "order": 14,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            13633
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler_ancestral_cfg_pp"
+      ]
+    },
+    {
+      "id": 5380,
+      "type": "LTXVTiledVAEDecode",
+      "pos": [
+        1390.5442945783254,
+        3012.070498452684
+      ],
+      "size": [
+        267.6441352715875,
+        198
+      ],
+      "flags": {},
+      "order": 58,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14413
+        },
+        {
+          "name": "latents",
+          "type": "LATENT",
+          "link": 14430
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14415
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVTiledVAEDecode"
+      },
+      "widgets_values": [
+        2,
+        2,
+        6,
+        false,
+        "auto",
+        "auto"
+      ]
+    },
+    {
+      "id": 5226,
+      "type": "LTXVLaplacianPyramidBlend",
+      "pos": [
+        1387.8563040360793,
+        3261.377584603743
+      ],
+      "size": [
+        337.65390625,
+        122
+      ],
+      "flags": {},
+      "order": 60,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_a",
+          "type": "IMAGE",
+          "link": 14415
+        },
+        {
+          "name": "image_b",
+          "type": "IMAGE",
+          "link": 14374
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14449
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            13934
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVLaplacianPyramidBlend"
+      },
+      "widgets_values": [
+        true,
+        2
+      ]
+    },
+    {
+      "id": 5227,
+      "type": "CreateVideo",
+      "pos": [
+        1757.9136242280626,
+        3260.538739618814
+      ],
+      "size": [
+        210,
+        78
+      ],
+      "flags": {},
+      "order": 61,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 13934
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 14433
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "widget": {
+            "name": "fps"
+          },
+          "link": 13968
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [
+            13935
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CreateVideo"
+      },
+      "widgets_values": [
+        30
+      ]
+    },
+    {
+      "id": 5344,
+      "type": "MarkdownNote",
+      "pos": [
+        1697.2107701774198,
+        3013.68865701351
+      ],
+      "size": [
+        485.490094009273,
+        133.8106738488218
+      ],
+      "flags": {},
+      "order": 15,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "# Pyramid Blending Note:\n### The blending dilation parameter is crucial and can significantly change results. We recommend experimenting with it to find the best setting for your use case.\n\n#### * Tip: If you have low frequencies in the blending area, increase the dilation value."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 5364,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        -2522.4254843800945,
+        3979.718653146065
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 35,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14368
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "MASK",
+          "links": [
+            14361,
+            14363,
+            14377
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale by multiplier",
+        0.5,
+        "area"
+      ]
+    },
+    {
+      "id": 3159,
+      "type": "LTXVImgToVideoConditionOnly",
+      "pos": [
+        -2098.1288523308262,
+        3352.0601221410543
+      ],
+      "size": [
+        313.2623046875,
+        142.859507731343
+      ],
+      "flags": {},
+      "order": 40,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 13279
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14417
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 11401
+        },
+        {
+          "name": "bypass",
+          "shape": 7,
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "bypass"
+          },
+          "link": 14318
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            13698
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVImgToVideoConditionOnly"
+      },
+      "widgets_values": [
+        0.7,
+        false
+      ]
+    },
+    {
+      "id": 5381,
+      "type": "LTXVPreprocess",
+      "pos": [
+        -2437.8952223512865,
+        3289.1648736936368
+      ],
+      "size": [
+        270,
+        58
+      ],
+      "flags": {},
+      "order": 26,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14416
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output_image",
+          "type": "IMAGE",
+          "links": [
+            14417
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LTXVPreprocess"
+      },
+      "widgets_values": [
+        18
+      ]
+    },
+    {
+      "id": 5093,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        -792.3920655382041,
+        3201.813714736499
+      ],
+      "size": [
+        212.3638671875,
+        106
+      ],
+      "flags": {},
+      "order": 45,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 13631
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 13632
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 13633
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 13664
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 14423
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": []
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": [
+            14424
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5213,
+      "type": "LTXVImgToVideoConditionOnly",
+      "pos": [
+        755.8808030041615,
+        3554.9547610713958
+      ],
+      "size": [
+        313.2623046875,
+        122
+      ],
+      "flags": {},
+      "order": 54,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14393
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14387
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14410
+        },
+        {
+          "name": "bypass",
+          "shape": 7,
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "bypass"
+          },
+          "link": 14386
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14426
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVImgToVideoConditionOnly"
+      },
+      "widgets_values": [
+        1,
+        false
+      ]
+    },
+    {
+      "id": 5387,
+      "type": "LTXVConcatAVLatent",
+      "pos": [
+        1119.3879862370936,
+        3495.7141909980437
+      ],
+      "size": [
+        181.308203125,
+        46
+      ],
+      "flags": {},
+      "order": 55,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "link": 14426
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14441
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14428
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVConcatAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5214,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        1086.4164988833702,
+        3258.7388797889157
+      ],
+      "size": [
+        212.3638671875,
+        106
+      ],
+      "flags": {},
+      "order": 56,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 13908
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 13909
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 13910
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 13907
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 14428
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": [
+            14429
+          ]
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.71",
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5388,
+      "type": "LTXVSeparateAVLatent",
+      "pos": [
+        1134.7614571145912,
+        3094.4250109360846
+      ],
+      "size": [
+        193.2916015625,
+        46
+      ],
+      "flags": {},
+      "order": 57,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "av_latent",
+          "type": "LATENT",
+          "link": 14429
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "links": [
+            14430
+          ]
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "links": [
+            14431
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVSeparateAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5114,
+      "type": "LTXAddVideoICLoRAGuideAdvanced",
+      "pos": [
+        -1738.5746947606647,
+        3289.73850629993
+      ],
+      "size": [
+        387.788671875,
+        326
+      ],
+      "flags": {},
+      "order": 41,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 13701
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 13700
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 13699
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 13698
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14372
+        },
+        {
+          "name": "attention_mask",
+          "shape": 7,
+          "type": "MASK",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            13694,
+            13786,
+            14437
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            13696,
+            13787,
+            14438
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14419
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXAddVideoICLoRAGuideAdvanced"
+      },
+      "widgets_values": [
+        0,
+        1,
+        1,
+        "disabled",
+        false,
+        256,
+        64,
+        1
+      ]
+    },
+    {
+      "id": 5385,
+      "type": "LTXVAudioVAELoader",
+      "pos": [
+        -4593.26243582442,
+        2610.667720674385
+      ],
+      "size": [
+        384.6576773713696,
+        79.30435270514249
+      ],
+      "flags": {},
+      "order": 16,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "Audio VAE",
+          "type": "VAE",
+          "links": [
+            14421,
+            14444
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.64",
+        "Node name for S&R": "LTXVAudioVAELoader"
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-dev.safetensors"
+      ]
+    },
+    {
+      "id": 4922,
+      "type": "LoraLoaderModelOnly",
+      "pos": [
+        -4135.6637869126125,
+        2704.8442961524884
+      ],
+      "size": [
+        454.14193097539646,
+        128.44311555561626
+      ],
+      "flags": {},
+      "order": 22,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 13217
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            13400
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "LoraLoaderModelOnly"
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-distilled-lora-384-1.1.safetensors",
+        0.5
+      ]
+    },
+    {
+      "id": 5383,
+      "type": "LTXVConcatAVLatent",
+      "pos": [
+        -990.7200189453805,
+        3692.9211897884597
+      ],
+      "size": [
+        181.308203125,
+        46
+      ],
+      "flags": {},
+      "order": 44,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "link": 14419
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14436
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14423
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVConcatAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 3059,
+      "type": "EmptyLTXVLatentVideo",
+      "pos": [
+        -2341.3250113842996,
+        3391.2726818340766
+      ],
+      "size": [
+        210,
+        130
+      ],
+      "flags": {},
+      "order": 39,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "width",
+          "type": "INT",
+          "widget": {
+            "name": "width"
+          },
+          "link": 13531
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "widget": {
+            "name": "height"
+          },
+          "link": 13532
+        },
+        {
+          "name": "length",
+          "type": "INT",
+          "widget": {
+            "name": "length"
+          },
+          "link": 13627
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            11401
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.43",
+        "Node name for S&R": "EmptyLTXVLatentVideo"
+      },
+      "widgets_values": [
+        960,
+        544,
+        121,
+        1
+      ]
+    },
+    {
+      "id": 5386,
+      "type": "LTXVSeparateAVLatent",
+      "pos": [
+        -769.7261825014932,
+        3357.1443502214884
+      ],
+      "size": [
+        193.2916015625,
+        46
+      ],
+      "flags": {},
+      "order": 46,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "av_latent",
+          "type": "LATENT",
+          "link": 14424
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "links": [
+            14425
+          ]
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "links": [
+            14440
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVSeparateAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5358,
+      "type": "LTXVInpaintPreprocess",
+      "pos": [
+        -2230.412581503991,
+        3899.2256168057274
+      ],
+      "size": [
+        248.4576171875,
+        46
+      ],
+      "flags": {},
+      "order": 37,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14362
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14363
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14371,
+            14372,
+            14439
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVInpaintPreprocess"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5382,
+      "type": "VAEEncodeAudio",
+      "pos": [
+        -1261.610854315966,
+        3651.8024206056803
+      ],
+      "size": [
+        199.7151612405418,
+        46
+      ],
+      "flags": {},
+      "order": 28,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14422
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14421
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14435
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "VAEEncodeAudio"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5391,
+      "type": "LTXVSetAudioRefTokens",
+      "pos": [
+        830.493941178539,
+        3732.732676853854
+      ],
+      "size": [
+        227.3482421875,
+        66
+      ],
+      "flags": {},
+      "order": 49,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14443
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14442
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14440
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "frozen_audio",
+          "type": "LATENT",
+          "links": [
+            14441
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVSetAudioRefTokens"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5389,
+      "type": "LTXVAudioVAEDecode",
+      "pos": [
+        1520.9967443048356,
+        3439.3878698866783
+      ],
+      "size": [
+        202.991796875,
+        46
+      ],
+      "flags": {},
+      "order": 59,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14431
+        },
+        {
+          "label": "Audio VAE",
+          "name": "audio_vae",
+          "type": "VAE",
+          "link": 14444
+        }
+      ],
+      "outputs": [
+        {
+          "name": "Audio",
+          "type": "AUDIO",
+          "links": [
+            14433
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LTXVAudioVAEDecode"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5390,
+      "type": "LTXVSetAudioRefTokens",
+      "pos": [
+        -1276.4815389923103,
+        3752.536917757552
+      ],
+      "size": [
+        227.3482421875,
+        66
+      ],
+      "flags": {},
+      "order": 43,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14437
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14438
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14435
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "frozen_audio",
+          "type": "LATENT",
+          "links": [
+            14436
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXVSetAudioRefTokens"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5392,
+      "type": "Note",
+      "pos": [
+        -770.2245251700002,
+        3656.869431993711
+      ],
+      "size": [
+        382.08295365380104,
+        112.03481164049845
+      ],
+      "flags": {},
+      "order": 17,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "Frozen audio helps guiding outpainting to be consistent with the sounds in the video.\nIf you don't want this guidance (e.g. you use this flow to replace a section of the frame or you prefer the off-frame sound to stay off frame) remove audio stream from the generation altogether and/or adjust positive/negative conditioning."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 5228,
+      "type": "SaveVideo",
+      "pos": [
+        2020.2332780449372,
+        3216.7642556077935
+      ],
+      "size": [
+        835.0556530433919,
+        571.1982033912554
+      ],
+      "flags": {},
+      "order": 62,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 13935
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "video/LTX_Reframe_Outpaint",
+        "auto",
+        "auto"
+      ]
+    },
+    {
+      "id": 5020,
+      "type": "GemmaAPITextEncode",
+      "pos": [
+        -2936.926436998826,
+        3142.334368848411
+      ],
+      "size": [
+        337.49481491927054,
+        160
+      ],
+      "flags": {},
+      "order": 20,
+      "mode": 4,
+      "inputs": [
+        {
+          "name": "api_key",
+          "type": "STRING",
+          "widget": {
+            "name": "api_key"
+          },
+          "link": 13457
+        }
+      ],
+      "outputs": [
+        {
+          "name": "conditioning",
+          "type": "CONDITIONING",
+          "links": []
+        }
+      ],
+      "title": "\ud83c\udd5b\ud83c\udd63\ud83c\udd67 Gemma API Text Encode - NEGATIVE",
+      "properties": {
+        "Node name for S&R": "GemmaAPITextEncode"
+      },
+      "widgets_values": [
+        "",
+        "pc game, console game, video game, cartoon, childish, ugly",
+        false,
+        "ltx-2.3-22b-dev.safetensors"
+      ]
+    },
+    {
+      "id": 5021,
+      "type": "GemmaAPITextEncode",
+      "pos": [
+        -3300.7291881866963,
+        3141.7377574886063
+      ],
+      "size": [
+        347.3886205089311,
+        160
+      ],
+      "flags": {},
+      "order": 21,
+      "mode": 4,
+      "inputs": [
+        {
+          "name": "api_key",
+          "type": "STRING",
+          "widget": {
+            "name": "api_key"
+          },
+          "link": 13458
+        }
+      ],
+      "outputs": [
+        {
+          "name": "conditioning",
+          "type": "CONDITIONING",
+          "links": []
+        }
+      ],
+      "title": "\ud83c\udd5b\ud83c\udd63\ud83c\udd67 Gemma API Text Encode - POSITIVE",
+      "properties": {
+        "Node name for S&R": "GemmaAPITextEncode"
+      },
+      "widgets_values": [
+        "",
+        "",
+        "ltx-2.3-22b-dev.safetensors",
+        "ltx-2.3-22b-dev.safetensors"
+      ]
+    },
+    {
+      "id": 5011,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        -4189.54959372093,
+        2922.366756915213
+      ],
+      "size": [
+        570.2243956508886,
+        105.50935843305933
+      ],
+      "flags": {},
+      "order": 25,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 13400
+        }
+      ],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            13401,
+            13911
+          ]
+        },
+        {
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXICLoRALoaderModelOnly"
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors",
+        1
+      ]
+    },
+    {
+      "id": 5368,
+      "type": "LTXReframeLayout",
+      "pos": [
+        -4225.854240441995,
+        3547.938320763212
+      ],
+      "size": [
+        600,
+        760
+      ],
+      "flags": {},
+      "order": 12,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "padded_frames",
+          "type": "IMAGE",
+          "links": [
+            14364
+          ]
+        },
+        {
+          "name": "outpaint_mask",
+          "type": "MASK",
+          "links": [
+            14369
+          ]
+        },
+        {
+          "name": "source_mask",
+          "type": "MASK",
+          "links": []
+        },
+        {
+          "name": "source_frames",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            14422
+          ]
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "links": [
+            13889,
+            13968
+          ]
+        },
+        {
+          "name": "source_video",
+          "type": "VIDEO",
+          "links": []
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "links": []
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "links": []
+        },
+        {
+          "name": "placement_json",
+          "type": "STRING",
+          "links": []
+        }
+      ],
+      "title": "1. LTX Reframe Layout \u00b7 Source & Framing",
+      "properties": {
+        "Node name for S&R": "LTXReframeLayout"
+      },
+      "widgets_values": [
+        "LTX-2_00180_.mp4",
+        "9:16",
+        576,
+        1024,
+        0.5,
+        0.5,
+        0.7,
+        32
+      ],
+      "color": "#1f4d52",
+      "bgcolor": "#16383c"
+    },
+    {
+      "id": 5400,
+      "type": "MaskExpandBatch+",
+      "pos": [
+        1050.0,
+        3261.0
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 63,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14376
+        },
+        {
+          "name": "size",
+          "type": "INT",
+          "widget": {
+            "name": "size"
+          },
+          "link": 14447
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14449
+          ]
+        }
+      ],
+      "title": "Full-Res Mask \u00b7 Match Video Frame Count",
+      "properties": {
+        "cnr_id": "comfyui_essentials",
+        "Node name for S&R": "MaskExpandBatch+"
+      },
+      "widgets_values": [
+        121,
+        "expand"
+      ],
+      "color": "#5b4636",
+      "bgcolor": "#3c3027"
+    },
+    {
+      "id": 5401,
+      "type": "MaskExpandBatch+",
+      "pos": [
+        -660.0,
+        3432.0
+      ],
+      "size": [
+        270,
+        106
+      ],
+      "flags": {},
+      "order": 64,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14377
+        },
+        {
+          "name": "size",
+          "type": "INT",
+          "widget": {
+            "name": "size"
+          },
+          "link": 14448
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14450
+          ]
+        }
+      ],
+      "title": "Half-Res Mask \u00b7 Match Video Frame Count",
+      "properties": {
+        "cnr_id": "comfyui_essentials",
+        "Node name for S&R": "MaskExpandBatch+"
+      },
+      "widgets_values": [
+        121,
+        "expand"
+      ],
+      "color": "#5b4636",
+      "bgcolor": "#3c3027"
+    }
+  ],
+  "links": [
+    [
+      7065,
+      2483,
+      0,
+      1241,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      7652,
+      2612,
+      0,
+      1241,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      11401,
+      3059,
+      0,
+      3159,
+      2,
+      "LATENT"
+    ],
+    [
+      13217,
+      3940,
+      0,
+      4922,
+      0,
+      "MODEL"
+    ],
+    [
+      13279,
+      3940,
+      2,
+      3159,
+      0,
+      "VAE"
+    ],
+    [
+      13348,
+      3940,
+      2,
+      4851,
+      1,
+      "VAE"
+    ],
+    [
+      13400,
+      4922,
+      0,
+      5011,
+      0,
+      "MODEL"
+    ],
+    [
+      13401,
+      5011,
+      0,
+      4828,
+      0,
+      "MODEL"
+    ],
+    [
+      13412,
+      5013,
+      2,
+      4851,
+      0,
+      "LATENT"
+    ],
+    [
+      13457,
+      5022,
+      0,
+      5020,
+      0,
+      "STRING"
+    ],
+    [
+      13458,
+      5022,
+      0,
+      5021,
+      0,
+      "STRING"
+    ],
+    [
+      13459,
+      5023,
+      0,
+      2483,
+      0,
+      "CLIP"
+    ],
+    [
+      13460,
+      5023,
+      0,
+      2612,
+      0,
+      "CLIP"
+    ],
+    [
+      13531,
+      5054,
+      0,
+      3059,
+      0,
+      "INT"
+    ],
+    [
+      13532,
+      5054,
+      1,
+      3059,
+      1,
+      "INT"
+    ],
+    [
+      13627,
+      5054,
+      2,
+      3059,
+      2,
+      "INT"
+    ],
+    [
+      13631,
+      4832,
+      0,
+      5093,
+      0,
+      "NOISE"
+    ],
+    [
+      13632,
+      4828,
+      0,
+      5093,
+      1,
+      "GUIDER"
+    ],
+    [
+      13633,
+      4831,
+      0,
+      5093,
+      2,
+      "SAMPLER"
+    ],
+    [
+      13664,
+      5025,
+      0,
+      5093,
+      3,
+      "SIGMAS"
+    ],
+    [
+      13694,
+      5114,
+      0,
+      5013,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      13696,
+      5114,
+      1,
+      5013,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      13698,
+      3159,
+      0,
+      5114,
+      3,
+      "LATENT"
+    ],
+    [
+      13699,
+      3940,
+      2,
+      5114,
+      2,
+      "VAE"
+    ],
+    [
+      13700,
+      1241,
+      1,
+      5114,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      13701,
+      1241,
+      0,
+      5114,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      13786,
+      5114,
+      0,
+      4828,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      13787,
+      5114,
+      1,
+      4828,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      13851,
+      2004,
+      0,
+      5188,
+      0,
+      "IMAGE"
+    ],
+    [
+      13889,
+      5368,
+      5,
+      1241,
+      2,
+      "FLOAT"
+    ],
+    [
+      13907,
+      5211,
+      0,
+      5214,
+      3,
+      "SIGMAS"
+    ],
+    [
+      13908,
+      5210,
+      0,
+      5214,
+      0,
+      "NOISE"
+    ],
+    [
+      13909,
+      5209,
+      0,
+      5214,
+      1,
+      "GUIDER"
+    ],
+    [
+      13910,
+      5215,
+      0,
+      5214,
+      2,
+      "SAMPLER"
+    ],
+    [
+      13911,
+      5011,
+      0,
+      5209,
+      0,
+      "MODEL"
+    ],
+    [
+      13934,
+      5226,
+      0,
+      5227,
+      0,
+      "IMAGE"
+    ],
+    [
+      13935,
+      5227,
+      0,
+      5228,
+      0,
+      "VIDEO"
+    ],
+    [
+      13968,
+      5368,
+      5,
+      5227,
+      2,
+      "FLOAT"
+    ],
+    [
+      14193,
+      4851,
+      0,
+      5266,
+      0,
+      "IMAGE"
+    ],
+    [
+      14262,
+      5013,
+      0,
+      5209,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14263,
+      5013,
+      1,
+      5209,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      14318,
+      5088,
+      0,
+      3159,
+      3,
+      "BOOLEAN"
+    ],
+    [
+      14360,
+      5360,
+      0,
+      5357,
+      0,
+      "IMAGE"
+    ],
+    [
+      14361,
+      5364,
+      0,
+      5357,
+      1,
+      "IMAGE,MASK"
+    ],
+    [
+      14362,
+      5357,
+      0,
+      5358,
+      0,
+      "IMAGE"
+    ],
+    [
+      14363,
+      5364,
+      0,
+      5358,
+      1,
+      "MASK"
+    ],
+    [
+      14364,
+      5368,
+      0,
+      5360,
+      0,
+      "IMAGE"
+    ],
+    [
+      14365,
+      5360,
+      0,
+      5361,
+      0,
+      "IMAGE"
+    ],
+    [
+      14366,
+      5365,
+      0,
+      5361,
+      1,
+      "MASK"
+    ],
+    [
+      14368,
+      5365,
+      0,
+      5364,
+      0,
+      "MASK"
+    ],
+    [
+      14369,
+      5368,
+      1,
+      5365,
+      0,
+      "MASK"
+    ],
+    [
+      14370,
+      5360,
+      0,
+      5365,
+      1,
+      "IMAGE"
+    ],
+    [
+      14371,
+      5358,
+      0,
+      5054,
+      0,
+      "IMAGE"
+    ],
+    [
+      14372,
+      5358,
+      0,
+      5114,
+      4,
+      "IMAGE"
+    ],
+    [
+      14374,
+      5361,
+      0,
+      5226,
+      1,
+      "IMAGE"
+    ],
+    [
+      14376,
+      5365,
+      0,
+      5400,
+      0,
+      "MASK"
+    ],
+    [
+      14377,
+      5364,
+      0,
+      5401,
+      0,
+      "MASK"
+    ],
+    [
+      14386,
+      5088,
+      0,
+      5213,
+      3,
+      "BOOLEAN"
+    ],
+    [
+      14387,
+      5188,
+      0,
+      5213,
+      1,
+      "IMAGE"
+    ],
+    [
+      14392,
+      3940,
+      2,
+      5267,
+      1,
+      "VAE"
+    ],
+    [
+      14393,
+      3940,
+      2,
+      5213,
+      0,
+      "VAE"
+    ],
+    [
+      14408,
+      5266,
+      0,
+      5379,
+      0,
+      "IMAGE"
+    ],
+    [
+      14409,
+      5379,
+      0,
+      5267,
+      0,
+      "IMAGE"
+    ],
+    [
+      14410,
+      5267,
+      0,
+      5213,
+      2,
+      "LATENT"
+    ],
+    [
+      14413,
+      3940,
+      2,
+      5380,
+      0,
+      "VAE"
+    ],
+    [
+      14415,
+      5380,
+      0,
+      5226,
+      0,
+      "IMAGE"
+    ],
+    [
+      14416,
+      5188,
+      0,
+      5381,
+      0,
+      "IMAGE"
+    ],
+    [
+      14417,
+      5381,
+      0,
+      3159,
+      1,
+      "IMAGE"
+    ],
+    [
+      14419,
+      5114,
+      2,
+      5383,
+      0,
+      "LATENT"
+    ],
+    [
+      14421,
+      5385,
+      0,
+      5382,
+      1,
+      "VAE"
+    ],
+    [
+      14422,
+      5368,
+      4,
+      5382,
+      0,
+      "AUDIO"
+    ],
+    [
+      14423,
+      5383,
+      0,
+      5093,
+      4,
+      "LATENT"
+    ],
+    [
+      14424,
+      5093,
+      1,
+      5386,
+      0,
+      "LATENT"
+    ],
+    [
+      14425,
+      5386,
+      0,
+      5013,
+      2,
+      "LATENT"
+    ],
+    [
+      14426,
+      5213,
+      0,
+      5387,
+      0,
+      "LATENT"
+    ],
+    [
+      14428,
+      5387,
+      0,
+      5214,
+      4,
+      "LATENT"
+    ],
+    [
+      14429,
+      5214,
+      0,
+      5388,
+      0,
+      "LATENT"
+    ],
+    [
+      14430,
+      5388,
+      0,
+      5380,
+      1,
+      "LATENT"
+    ],
+    [
+      14431,
+      5388,
+      1,
+      5389,
+      0,
+      "LATENT"
+    ],
+    [
+      14433,
+      5389,
+      0,
+      5227,
+      1,
+      "AUDIO"
+    ],
+    [
+      14435,
+      5382,
+      0,
+      5390,
+      2,
+      "LATENT"
+    ],
+    [
+      14436,
+      5390,
+      2,
+      5383,
+      1,
+      "LATENT"
+    ],
+    [
+      14437,
+      5114,
+      0,
+      5390,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14438,
+      5114,
+      1,
+      5390,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14439,
+      5358,
+      0,
+      5266,
+      1,
+      "IMAGE"
+    ],
+    [
+      14440,
+      5386,
+      1,
+      5391,
+      2,
+      "LATENT"
+    ],
+    [
+      14441,
+      5391,
+      2,
+      5387,
+      1,
+      "LATENT"
+    ],
+    [
+      14442,
+      5013,
+      1,
+      5391,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14443,
+      5013,
+      0,
+      5391,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14444,
+      5385,
+      0,
+      5389,
+      1,
+      "VAE"
+    ],
+    [
+      14447,
+      5054,
+      2,
+      5400,
+      1,
+      "INT"
+    ],
+    [
+      14448,
+      5054,
+      2,
+      5401,
+      1,
+      "INT"
+    ],
+    [
+      14449,
+      5400,
+      0,
+      5226,
+      2,
+      "MASK"
+    ],
+    [
+      14450,
+      5401,
+      0,
+      5266,
+      2,
+      "MASK"
+    ]
+  ],
+  "floatingLinks": [
+    {
+      "id": 2,
+      "origin_id": 3940,
+      "origin_slot": 2,
+      "target_id": -1,
+      "target_slot": -1,
+      "type": "VAE",
+      "parentId": 155
+    }
+  ],
+  "groups": [
+    {
+      "id": 54,
+      "title": "Set prompts     -> use API nodes to save on memory",
+      "bounding": [
+        -3527.2747393631416,
+        2961.3030061080112,
+        937.5727445850584,
+        669.1377243262345
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        5023,
+        5022,
+        1241,
+        2483,
+        2612,
+        5020,
+        5021
+      ]
+    },
+    {
+      "id": 106,
+      "title": "Preprocess",
+      "bounding": [
+        -2547.9363511080323,
+        3219.73850629993,
+        1207.1503282223675,
+        406
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        5054,
+        3159,
+        5381,
+        5114,
+        3059
+      ]
+    },
+    {
+      "id": 110,
+      "title": "Load Models",
+      "bounding": [
+        -4609.826307240758,
+        2540.667720674385,
+        1000.501109170717,
+        497.2083946738876
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        3940,
+        5193,
+        5385,
+        4922,
+        5011
+      ]
+    },
+    {
+      "id": 119,
+      "title": "Load Image",
+      "bounding": [
+        -4231.497658697313,
+        3093.3889895901516,
+        626.0190879868719,
+        397.600244140625
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        2004,
+        5088,
+        5188
+      ]
+    },
+    {
+      "id": 147,
+      "title": "Generate Low Resolution",
+      "bounding": [
+        -1292.2004731850486,
+        3066.4885401533693,
+        953.426390181316,
+        762.537672464814
+      ],
+      "color": "#b58b2a",
+      "flags": {},
+      "profilingGroup": false,
+      "nodes": [
+        4832,
+        5025,
+        4828,
+        5013,
+        4831,
+        5093,
+        5383,
+        5386,
+        5382,
+        5390,
+        5392
+      ]
+    },
+    {
+      "id": 150,
+      "title": "Decode and Blend",
+      "bounding": [
+        -285.57328822282614,
+        3150.9811784851136,
+        357.65390625,
+        412.8995561888355
+      ],
+      "color": "#b58b2a",
+      "flags": {},
+      "nodes": [
+        4851,
+        5266
+      ]
+    },
+    {
+      "id": 164,
+      "title": "1. Reframe Source Video",
+      "bounding": [
+        -4235.854240441995,
+        3507.938320763212,
+        620,
+        800
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        5368
+      ]
+    },
+    {
+      "id": 168,
+      "title": "2. Prepare Two-Stage Outpaint Conditions",
+      "bounding": [
+        -3519.702251971042,
+        3700.708147137584,
+        1557.7472876545512,
+        422.66511064225415
+      ],
+      "color": "#3f789e",
+      "flags": {},
+      "nodes": [
+        5360,
+        5365,
+        5361,
+        5357,
+        5356,
+        5364,
+        5358
+      ]
+    },
+    {
+      "id": 174,
+      "title": "Generate High Resolution",
+      "bounding": [
+        128.07073714387283,
+        2963.58038320638,
+        1209.4420412977106,
+        842.1418679941612
+      ],
+      "color": "#b58b2a",
+      "flags": {},
+      "profilingGroup": false,
+      "nodes": [
+        5209,
+        5210,
+        5211,
+        5379,
+        5267,
+        5215,
+        5213,
+        5387,
+        5214,
+        5388,
+        5391
+      ]
+    },
+    {
+      "id": 175,
+      "title": "Decode and Blend",
+      "bounding": [
+        1363.6521166758394,
+        2933.2015609870045,
+        1501.6368144124888,
+        864.7608980120444
+      ],
+      "color": "#b58b2a",
+      "flags": {},
+      "nodes": [
+        5380,
+        5226,
+        5227,
+        5344,
+        5389,
+        5228
+      ]
+    },
+    {
+      "id": 178,
+      "title": "Stage 2 Full-Resolution Mask",
+      "bounding": [
+        -3519.702251971042,
+        3731.071210260153,
+        941.7341790977625,
+        392.3020475196854
+      ],
+      "color": "#b06634",
+      "flags": {},
+      "nodes": [
+        5360,
+        5365,
+        5361,
+        5356
+      ]
+    },
+    {
+      "id": 179,
+      "title": "Stage 1 Half-Resolution Mask",
+      "bounding": [
+        -2544.3352235060156,
+        3740.708147137584,
+        572.3802591895246,
+        360.2386318034096
+      ],
+      "color": "#8AA",
+      "flags": {},
+      "nodes": [
+        5357,
+        5364,
+        5358
+      ]
+    }
+  ],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.8013358768403982,
+      "offset": [
+        4795.347100706883,
+        -2751.771636181878
+      ]
+    },
+    "frontendVersion": "1.44.19",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true,
+    "workspace_info": {
+      "id": "elBQFQknIoLYTEwIloQuw"
+    },
+    "node_versions": {
+      "comfy-core": "0.3.30",
+      "comfyui-kjnodes": "1.0.8"
+    },
+    "workflowRendererVersion": "LG",
+    "reroutes": [
+      {
+        "id": 2,
+        "pos": [
+          -3590.6270518552647,
+          2904.0409790822855
+        ],
+        "linkIds": [
+          13401,
+          13911
+        ]
+      },
+      {
+        "id": 3,
+        "parentId": 2,
+        "pos": [
+          -1345.751529515756,
+          2890.936829220813
+        ],
+        "linkIds": [
+          13401,
+          13911
+        ]
+      },
+      {
+        "id": 15,
+        "pos": [
+          -4194.169315694357,
+          2864.3530863704896
+        ],
+        "linkIds": [
+          13279,
+          13348,
+          13699,
+          14392,
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 16,
+        "parentId": 15,
+        "pos": [
+          -2113.9545303856066,
+          2855.5838284457577
+        ],
+        "linkIds": [
+          13279,
+          13348,
+          13699,
+          14392,
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 33,
+        "pos": [
+          -1301.8023808193664,
+          3708.981681527194
+        ],
+        "linkIds": [
+          14419
+        ]
+      },
+      {
+        "id": 50,
+        "parentId": 58,
+        "pos": [
+          -386.5494676604432,
+          2848.2582402119815
+        ],
+        "linkIds": [
+          13348,
+          14392,
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 58,
+        "parentId": 16,
+        "pos": [
+          -1817.0178416961805,
+          2860.2505773848466
+        ],
+        "linkIds": [
+          13348,
+          13699,
+          14392,
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 61,
+        "pos": [
+          -1315.211190063696,
+          3009.587340206122
+        ],
+        "linkIds": [
+          13694
+        ]
+      },
+      {
+        "id": 63,
+        "pos": [
+          -1297.1256414042925,
+          3036.022647952905
+        ],
+        "linkIds": [
+          13696
+        ]
+      },
+      {
+        "id": 64,
+        "parentId": 61,
+        "pos": [
+          -578.2753913432492,
+          3006.2921398060853
+        ],
+        "linkIds": [
+          13694
+        ]
+      },
+      {
+        "id": 65,
+        "parentId": 63,
+        "pos": [
+          -588.34055850083,
+          3036.2805015181625
+        ],
+        "linkIds": [
+          13696
+        ]
+      },
+      {
+        "id": 102,
+        "parentId": 16,
+        "pos": [
+          156.21662258384063,
+          1922.9418953645031
+        ],
+        "linkIds": [
+          13348
+        ]
+      },
+      {
+        "id": 108,
+        "parentId": 102,
+        "pos": [
+          1527.502764513066,
+          1929.93778577031
+        ],
+        "linkIds": [
+          13348
+        ]
+      },
+      {
+        "id": 137,
+        "pos": [
+          -2817.3764805292653,
+          3672.8607371859407
+        ],
+        "linkIds": [
+          13889,
+          13968
+        ]
+      },
+      {
+        "id": 147,
+        "parentId": 211,
+        "pos": [
+          129.4131565164567,
+          3836.5654119170695
+        ],
+        "linkIds": [
+          13968
+        ]
+      },
+      {
+        "id": 151,
+        "pos": [
+          -332.54814239009386,
+          2932.684066637595
+        ],
+        "linkIds": [
+          14262,
+          14443
+        ]
+      },
+      {
+        "id": 152,
+        "parentId": 151,
+        "pos": [
+          704.2657235180052,
+          2920.393040909325
+        ],
+        "linkIds": [
+          14262,
+          14443
+        ]
+      },
+      {
+        "id": 153,
+        "pos": [
+          -312.56638993921615,
+          2952.5427681952415
+        ],
+        "linkIds": [
+          14263,
+          14442
+        ]
+      },
+      {
+        "id": 154,
+        "parentId": 153,
+        "pos": [
+          680.1015071708348,
+          2945.400140257721
+        ],
+        "linkIds": [
+          14263,
+          14442
+        ]
+      },
+      {
+        "id": 155,
+        "parentId": 205,
+        "pos": [
+          1318.749983351773,
+          2847.4611906688683
+        ],
+        "linkIds": [
+          14413
+        ],
+        "floating": {
+          "slotType": "output"
+        }
+      },
+      {
+        "id": 157,
+        "parentId": 3,
+        "pos": [
+          732.9823615083058,
+          2887.9073219003294
+        ],
+        "linkIds": [
+          13911
+        ]
+      },
+      {
+        "id": 164,
+        "parentId": 147,
+        "pos": [
+          1709.2982241765972,
+          3839.5390229057557
+        ],
+        "linkIds": [
+          13968
+        ]
+      },
+      {
+        "id": 181,
+        "pos": [
+          -1804.0506779305658,
+          3653.22543111722
+        ],
+        "linkIds": [
+          14371,
+          14372
+        ]
+      },
+      {
+        "id": 182,
+        "parentId": 212,
+        "pos": [
+          -361.3396521175539,
+          3908.40232811846
+        ],
+        "linkIds": [
+          14439
+        ]
+      },
+      {
+        "id": 183,
+        "pos": [
+          -2556.4095785712134,
+          4154.12832933469
+        ],
+        "linkIds": [
+          14374
+        ]
+      },
+      {
+        "id": 184,
+        "parentId": 183,
+        "pos": [
+          1293.5707952714795,
+          4145.010399529478
+        ],
+        "linkIds": [
+          14374
+        ]
+      },
+      {
+        "id": 185,
+        "parentId": 189,
+        "pos": [
+          1261.497771098751,
+          4183.692406781076
+        ],
+        "linkIds": [
+          14376
+        ]
+      },
+      {
+        "id": 187,
+        "pos": [
+          -337.5323369674491,
+          3985.9354727247196
+        ],
+        "linkIds": [
+          14377
+        ]
+      },
+      {
+        "id": 189,
+        "pos": [
+          -2848.799531654894,
+          4193.862441364728
+        ],
+        "linkIds": [
+          14376
+        ]
+      },
+      {
+        "id": 190,
+        "pos": [
+          -3545.3374007324874,
+          2836.2377036689645
+        ],
+        "linkIds": [
+          14387,
+          14416
+        ]
+      },
+      {
+        "id": 191,
+        "parentId": 190,
+        "pos": [
+          -2497.853857066724,
+          2834.5541740100625
+        ],
+        "linkIds": [
+          14387,
+          14416
+        ]
+      },
+      {
+        "id": 192,
+        "pos": [
+          -3562.8828163755775,
+          4239.244444751012
+        ],
+        "linkIds": [
+          14422
+        ]
+      },
+      {
+        "id": 193,
+        "pos": [
+          -3575.9231334331366,
+          2941.741293671788
+        ],
+        "linkIds": [
+          14318,
+          14386
+        ]
+      },
+      {
+        "id": 194,
+        "pos": [
+          -2577.5687441827413,
+          3202.523972318653
+        ],
+        "linkIds": [
+          13701
+        ]
+      },
+      {
+        "id": 195,
+        "pos": [
+          -2562.7026800037997,
+          3169.9109288866416
+        ],
+        "linkIds": [
+          13700
+        ]
+      },
+      {
+        "id": 196,
+        "parentId": 195,
+        "pos": [
+          -1775.6366139134614,
+          3168.8300997642004
+        ],
+        "linkIds": [
+          13700
+        ]
+      },
+      {
+        "id": 197,
+        "parentId": 194,
+        "pos": [
+          -1756.8446250046516,
+          3187.986882616346
+        ],
+        "linkIds": [
+          13701
+        ]
+      },
+      {
+        "id": 198,
+        "parentId": 193,
+        "pos": [
+          -2132.239420587815,
+          2931.5292677593325
+        ],
+        "linkIds": [
+          14318,
+          14386
+        ]
+      },
+      {
+        "id": 199,
+        "parentId": 198,
+        "pos": [
+          523.5774899379476,
+          2929.2130266535046
+        ],
+        "linkIds": [
+          14386
+        ]
+      },
+      {
+        "id": 200,
+        "parentId": 191,
+        "pos": [
+          552.0037649556203,
+          2805.801934024303
+        ],
+        "linkIds": [
+          14387
+        ]
+      },
+      {
+        "id": 201,
+        "parentId": 181,
+        "pos": [
+          -2560.1889677068675,
+          3653.471174393283
+        ],
+        "linkIds": [
+          14371
+        ]
+      },
+      {
+        "id": 204,
+        "parentId": 50,
+        "pos": [
+          323.1928722399215,
+          2840.372502491493
+        ],
+        "linkIds": [
+          14392,
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 205,
+        "parentId": 204,
+        "pos": [
+          600.8814556987629,
+          2841.8071406777667
+        ],
+        "linkIds": [
+          14393,
+          14413
+        ]
+      },
+      {
+        "id": 206,
+        "parentId": 192,
+        "pos": [
+          -1339.664655053101,
+          4247.736434825181
+        ],
+        "linkIds": [
+          14422
+        ]
+      },
+      {
+        "id": 207,
+        "pos": [
+          -1368.8218458556153,
+          2628.289001654905
+        ],
+        "linkIds": [
+          14421,
+          14444
+        ]
+      },
+      {
+        "id": 209,
+        "pos": [
+          -531.2587270015435,
+          3785.0234281876337
+        ],
+        "linkIds": [
+          14440
+        ]
+      },
+      {
+        "id": 210,
+        "parentId": 137,
+        "pos": [
+          -1937.6672230023778,
+          3679.06553253562
+        ],
+        "linkIds": [
+          13968
+        ]
+      },
+      {
+        "id": 211,
+        "parentId": 210,
+        "pos": [
+          -1919.9830666698938,
+          3846.22463810917
+        ],
+        "linkIds": [
+          13968
+        ]
+      },
+      {
+        "id": 212,
+        "pos": [
+          -1755.0500983419108,
+          3912.7077856699543
+        ],
+        "linkIds": [
+          14439
+        ]
+      },
+      {
+        "id": 214,
+        "parentId": 207,
+        "pos": [
+          1347.8728831463682,
+          2631.5394893165935
+        ],
+        "linkIds": [
+          14444
+        ]
+      }
+    ],
+    "linkExtensions": [
+      {
+        "id": 13279,
+        "parentId": 16
+      },
+      {
+        "id": 13348,
+        "parentId": 50
+      },
+      {
+        "id": 13401,
+        "parentId": 3
+      },
+      {
+        "id": 13694,
+        "parentId": 64
+      },
+      {
+        "id": 13696,
+        "parentId": 65
+      },
+      {
+        "id": 13699,
+        "parentId": 58
+      },
+      {
+        "id": 13700,
+        "parentId": 196
+      },
+      {
+        "id": 13701,
+        "parentId": 197
+      },
+      {
+        "id": 13889,
+        "parentId": 137
+      },
+      {
+        "id": 13911,
+        "parentId": 157
+      },
+      {
+        "id": 13968,
+        "parentId": 164
+      },
+      {
+        "id": 14262,
+        "parentId": 152
+      },
+      {
+        "id": 14263,
+        "parentId": 154
+      },
+      {
+        "id": 14318,
+        "parentId": 198
+      },
+      {
+        "id": 14371,
+        "parentId": 201
+      },
+      {
+        "id": 14372,
+        "parentId": 181
+      },
+      {
+        "id": 14374,
+        "parentId": 184
+      },
+      {
+        "id": 14376,
+        "parentId": 185
+      },
+      {
+        "id": 14377,
+        "parentId": 187
+      },
+      {
+        "id": 14386,
+        "parentId": 199
+      },
+      {
+        "id": 14387,
+        "parentId": 200
+      },
+      {
+        "id": 14392,
+        "parentId": 204
+      },
+      {
+        "id": 14393,
+        "parentId": 205
+      },
+      {
+        "id": 14413,
+        "parentId": 155
+      },
+      {
+        "id": 14416,
+        "parentId": 191
+      },
+      {
+        "id": 14419,
+        "parentId": 33
+      },
+      {
+        "id": 14421,
+        "parentId": 207
+      },
+      {
+        "id": 14422,
+        "parentId": 206
+      },
+      {
+        "id": 14439,
+        "parentId": 182
+      },
+      {
+        "id": 14440,
+        "parentId": 209
+      },
+      {
+        "id": 14442,
+        "parentId": 154
+      },
+      {
+        "id": 14443,
+        "parentId": 152
+      },
+      {
+        "id": 14444,
+        "parentId": 214
+      }
+    ],
+    "ltx_reframe_workflow": {
+      "schema": "ltx_reframe_iclora_outpaint/v1",
+      "based_on": "LTX-2.3_ICLoRA_Outpaint_Two_Stage_Distilled.json",
+      "source_node": "LTXReframeLayout",
+      "default_source": "LTX-2_00180_.mp4",
+      "default_target": [
+        576,
+        1024
+      ],
+      "required_lora": "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors",
+      "mask_batch_safety": "MaskExpandBatch+ before both Laplacian blends"
+    }
+  },
+  "version": 0.4
+}
diff --git a/examples/09_minimax_h3_t2va_director.json b/examples/09_minimax_h3_t2va_director.json
new file mode 100644
index 0000000..0711231
--- /dev/null
+++ b/examples/09_minimax_h3_t2va_director.json
@@ -0,0 +1,2251 @@
+{
+    "id": "7fb07ba2-cbd8-4a52-b924-59cab6d2aeb3",
+    "revision": 0,
+    "last_node_id": 327,
+    "last_link_id": 479,
+    "nodes": [
+        {
+            "id": 312,
+            "type": "PrimitiveStringMultiline",
+            "pos": [
+                0,
+                0
+            ],
+            "size": [
+                500,
+                320
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [
+                {
+                    "name": "STRING",
+                    "type": "STRING",
+                    "links": [
+                        454
+                    ]
+                }
+            ],
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.24.0",
+                "Node name for S&R": "PrimitiveStringMultiline"
+            },
+            "widgets_values": [
+                "Create a 20-second realistic live-action action-trailer sequence at post-rain dusk. A coat-wearing protagonist escapes pursuers by sprinting and leaping across wet skyscraper rooftops; make every jump fluid and dangerous. Use hard cuts on musical beats, distinct camera angles, wind, footsteps, city ambience, and a restrained cinematic score. Preserve premium practical-film texture, anamorphic depth, volumetric city fog, and natural body motion. No text, subtitles, logos, animation, or overly-CG rendering."
+            ],
+            "title": "1. Creative brief"
+        },
+        {
+            "id": 325,
+            "type": "PrimitiveFloat",
+            "pos": [
+                0,
+                355
+            ],
+            "size": [
+                320,
+                80
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "value",
+                    "name": "value",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "value"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "localized_name": "FLOAT",
+                    "name": "FLOAT",
+                    "type": "FLOAT",
+                    "links": [
+                        473,
+                        474
+                    ]
+                }
+            ],
+            "title": "2. Duration seconds (shared)",
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.7.0",
+                "Node name for S&R": "PrimitiveFloat",
+                "enableTabs": false,
+                "tabWidth": 65,
+                "tabXOffset": 10,
+                "hasSecondTab": false,
+                "secondTabText": "Send Back",
+                "secondTabOffset": 80,
+                "secondTabWidth": 65
+            },
+            "widgets_values": [
+                20
+            ],
+            "color": "#223322",
+            "bgcolor": "#355335"
+        },
+        {
+            "id": 320,
+            "type": "DiffusionGemmaModelLoader",
+            "pos": [
+                0,
+                480
+            ],
+            "size": [
+                500,
+                320
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [
+                {
+                    "name": "model_config",
+                    "type": "DG_MODEL_CONFIG",
+                    "links": [
+                        457
+                    ]
+                },
+                {
+                    "name": "runtime_status_json",
+                    "type": "STRING",
+                    "links": []
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "DiffusionGemmaModelLoader"
+            },
+            "widgets_values": [
+                "C:\\ComfyUI\\app\\models\\LLM\\diffusiongemma-26B-A4B-it-NVFP4",
+                "transformers_inprocess",
+                "auto",
+                "modelopt_nvfp4",
+                true,
+                "unload_after_run",
+                "20.0 GB - safe practical",
+                0.35
+            ],
+            "title": "3. Load Director (unloads before H3)"
+        },
+        {
+            "id": 313,
+            "type": "DiffusionGemmaContextHub",
+            "pos": [
+                560,
+                0
+            ],
+            "size": [
+                420,
+                344
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "user_prompt",
+                    "type": "STRING",
+                    "link": 454
+                },
+                {
+                    "name": "image",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": null
+                },
+                {
+                    "name": "video",
+                    "shape": 7,
+                    "type": "VIDEO",
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "gemma_context",
+                    "type": "DG_CONTEXT",
+                    "links": [
+                        458,
+                        461
+                    ]
+                },
+                {
+                    "name": "context_json",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "context_preview",
+                    "type": "STRING",
+                    "links": []
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "DiffusionGemmaContextHub"
+            },
+            "widgets_values": [
+                1,
+                60,
+                60,
+                "trim",
+                false,
+                "",
+                "video_recreation"
+            ],
+            "title": "4. Prompt context"
+        },
+        {
+            "id": 314,
+            "type": "DiffusionGemmaTargetProfile",
+            "pos": [
+                560,
+                400
+            ],
+            "size": [
+                420,
+                560
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+                {
+                    "localized_name": "target_duration_seconds",
+                    "name": "target_duration_seconds",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "target_duration_seconds"
+                    },
+                    "link": 473
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "target_profile_config",
+                    "type": "DG_TARGET_PROFILE_CONFIG",
+                    "links": [
+                        459,
+                        462
+                    ]
+                },
+                {
+                    "name": "target_profile_json",
+                    "type": "STRING",
+                    "links": null
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "DiffusionGemmaTargetProfile"
+            },
+            "widgets_values": [
+                "minimax_h3",
+                "auto_scene_audio",
+                "",
+                20,
+                "",
+                "16:9",
+                "",
+                "",
+                false,
+                "auto",
+                ""
+            ],
+            "title": "5. MiniMax-H3 target profile"
+        },
+        {
+            "id": 315,
+            "type": "DiffusionGemmaCoTGenerator",
+            "pos": [
+                1040,
+                0
+            ],
+            "size": [
+                430,
+                304
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "model_config",
+                    "type": "DG_MODEL_CONFIG",
+                    "link": 457
+                },
+                {
+                    "name": "gemma_context",
+                    "type": "DG_CONTEXT",
+                    "link": 458
+                },
+                {
+                    "name": "target_profile_config",
+                    "type": "DG_TARGET_PROFILE_CONFIG",
+                    "link": 459
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "final_json",
+                    "type": "STRING",
+                    "links": [
+                        460
+                    ]
+                },
+                {
+                    "name": "reasoning_text",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "raw_response",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "metadata_json",
+                    "type": "STRING",
+                    "links": [
+                        479
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "DiffusionGemmaCoTGenerator"
+            },
+            "widgets_values": [
+                0.45,
+                "cinematic",
+                0.7,
+                "on",
+                1280
+            ],
+            "title": "6. Write strict H3 prompt"
+        },
+        {
+            "id": 316,
+            "type": "DiffusionGemmaJSONSplitter",
+            "pos": [
+                1040,
+                350
+            ],
+            "size": [
+                430,
+                510
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "final_json",
+                    "type": "STRING",
+                    "link": 460
+                },
+                {
+                    "name": "gemma_context",
+                    "shape": 7,
+                    "type": "DG_CONTEXT",
+                    "link": 461
+                },
+                {
+                    "name": "target_profile_config",
+                    "shape": 7,
+                    "type": "DG_TARGET_PROFILE_CONFIG",
+                    "link": 462
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "ltx_prompt",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "ideogram_prompt",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "negative_prompt",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "aspect_ratio",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "metadata_json",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "scene_segments_json",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "local_prompts",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "segment_lengths",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "is_valid",
+                    "type": "BOOLEAN",
+                    "links": null
+                },
+                {
+                    "name": "resolution_selector_preset",
+                    "type": "STRING",
+                    "links": null
+                },
+                {
+                    "name": "resolution_width",
+                    "type": "INT",
+                    "links": [
+                        475
+                    ]
+                },
+                {
+                    "name": "resolution_height",
+                    "type": "INT",
+                    "links": [
+                        476
+                    ]
+                },
+                {
+                    "name": "ready_for_generation",
+                    "type": "BOOLEAN",
+                    "links": null
+                },
+                {
+                    "name": "minimax_h3_prompt",
+                    "type": "STRING",
+                    "links": [
+                        477,
+                        478
+                    ]
+                }
+            ],
+            "properties": {
+                "Node name for S&R": "DiffusionGemmaJSONSplitter"
+            },
+            "widgets_values": [
+                "",
+                0.4,
+                32
+            ],
+            "title": "7. Route H3 prompt + resolution"
+        },
+        {
+            "id": 105,
+            "type": "4f4334ed-1aa4-5860-b237-56c863ab42d6",
+            "pos": [
+                1550,
+                0
+            ],
+            "size": [
+                515,
+                880
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "first_frame",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": null
+                },
+                {
+                    "name": "last_frame",
+                    "shape": 7,
+                    "type": "IMAGE",
+                    "link": null
+                },
+                {
+                    "name": "prompt",
+                    "type": "STRING",
+                    "widget": {
+                        "name": "prompt"
+                    },
+                    "link": 477
+                },
+                {
+                    "name": "width",
+                    "type": "INT",
+                    "widget": {
+                        "name": "width"
+                    },
+                    "link": 475
+                },
+                {
+                    "name": "height",
+                    "type": "INT",
+                    "widget": {
+                        "name": "height"
+                    },
+                    "link": 476
+                },
+                {
+                    "label": "duration",
+                    "name": "value_1",
+                    "type": "FLOAT",
+                    "widget": {
+                        "name": "value_1"
+                    },
+                    "link": 474
+                },
+                {
+                    "label": "audio_vae",
+                    "name": "vae_name_1",
+                    "type": "COMBO",
+                    "widget": {
+                        "name": "vae_name_1"
+                    },
+                    "link": null
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "VIDEO",
+                    "type": "VIDEO",
+                    "links": [
+                        194
+                    ]
+                }
+            ],
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.30.1",
+                "previewExposures": []
+            },
+            "widgets_values": [
+                "",
+                864,
+                480,
+                20,
+                42,
+                "minimax_h3_fl2va_pruned_int8_convrot.safetensors",
+                "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+                "minimax_h3_video_vae_fp16.safetensors",
+                "minimax_h3_audio_vae_fp32.safetensors"
+            ],
+            "title": "8. Native MiniMax-H3 T2VA \u2014 25 steps"
+        },
+        {
+            "id": 92,
+            "type": "SaveVideo",
+            "pos": [
+                2110,
+                0
+            ],
+            "size": [
+                550,
+                420
+            ],
+            "flags": {},
+            "order": 8,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "video",
+                    "type": "VIDEO",
+                    "link": 194
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "video",
+                    "type": "VIDEO",
+                    "links": null
+                }
+            ],
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.30.1",
+                "Node name for S&R": "SaveVideo"
+            },
+            "widgets_values": [
+                "video/minimax_h3/diffusiongemma_director_t2va",
+                "mp4",
+                "h264"
+            ],
+            "title": "9. Save generated video"
+        },
+        {
+            "id": 326,
+            "type": "PreviewAny",
+            "pos": [
+                2110,
+                460
+            ],
+            "size": [
+                550,
+                340
+            ],
+            "flags": {},
+            "order": 9,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "source",
+                    "type": "*",
+                    "link": 478
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "STRING",
+                    "type": "STRING",
+                    "links": null
+                }
+            ],
+            "title": "Generated MiniMax-H3 prompt",
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.15.1",
+                "Node name for S&R": "PreviewAny",
+                "enableTabs": false,
+                "tabWidth": 65,
+                "tabXOffset": 10,
+                "hasSecondTab": false,
+                "secondTabText": "Send Back",
+                "secondTabOffset": 80,
+                "secondTabWidth": 65
+            },
+            "widgets_values": [
+                null,
+                null,
+                null
+            ],
+            "color": "#2a363b",
+            "bgcolor": "#3f5159"
+        },
+        {
+            "id": 327,
+            "type": "PreviewAny",
+            "pos": [
+                2110,
+                830
+            ],
+            "size": [
+                550,
+                300
+            ],
+            "flags": {},
+            "order": 10,
+            "mode": 0,
+            "inputs": [
+                {
+                    "name": "source",
+                    "type": "*",
+                    "link": 479
+                }
+            ],
+            "outputs": [
+                {
+                    "name": "STRING",
+                    "type": "STRING",
+                    "links": null
+                }
+            ],
+            "title": "H3 validation / blocked reasons",
+            "properties": {
+                "cnr_id": "comfy-core",
+                "ver": "0.15.1",
+                "Node name for S&R": "PreviewAny",
+                "enableTabs": false,
+                "tabWidth": 65,
+                "tabXOffset": 10,
+                "hasSecondTab": false,
+                "secondTabText": "Send Back",
+                "secondTabOffset": 80,
+                "secondTabWidth": 65
+            },
+            "widgets_values": [
+                null,
+                null,
+                null
+            ],
+            "color": "#2a363b",
+            "bgcolor": "#3f5159"
+        },
+        {
+            "id": 118,
+            "type": "MarkdownNote",
+            "pos": [
+                0,
+                1220
+            ],
+            "size": [
+                1600,
+                390
+            ],
+            "flags": {},
+            "order": 11,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [],
+            "properties": {},
+            "widgets_values": [
+                "# DiffusionGemma Director \u2192 native MiniMax-H3 (T2VA)\n\n1. Replace the creative brief at upper left.\n2. Set Duration seconds. The same value feeds the Director profile and native H3 frame calculation.\n3. Queue the workflow. DiffusionGemma unloads after writing the H3 prompt, freeing VRAM before video generation.\n\nThis sample is set to 20 seconds and 25 sampling steps. At 24 fps, 20 seconds resolves to 481 valid H3 frames (about 20.04 seconds) because H3 uses a 17k+5 temporal grid.\n\nThe Director's dedicated minimax_h3_prompt output feeds the native MiniMaxH3ImageToVideo prompt directly and branches to Generated MiniMax-H3 prompt for inspection. H3 validation / blocked reasons displays the splitter metadata when a response cannot safely be routed. Its resolution outputs feed width and height at 0.4 MP, multiple 32, 16:9.\n\nShort clips in the 4\u201315 second range remain a useful creative default, but this native path is not artificially clamped to that range. Its node accepts up to 3600 frames. Longer clips require proportionally more generation time and VRAM."
+            ],
+            "color": "#222",
+            "bgcolor": "#000",
+            "title": "How this sample works"
+        }
+    ],
+    "links": [
+        [
+            194,
+            105,
+            0,
+            92,
+            0,
+            "VIDEO"
+        ],
+        [
+            454,
+            312,
+            0,
+            313,
+            0,
+            "STRING"
+        ],
+        [
+            457,
+            320,
+            0,
+            315,
+            0,
+            "DG_MODEL_CONFIG"
+        ],
+        [
+            458,
+            313,
+            0,
+            315,
+            1,
+            "DG_CONTEXT"
+        ],
+        [
+            459,
+            314,
+            0,
+            315,
+            2,
+            "DG_TARGET_PROFILE_CONFIG"
+        ],
+        [
+            460,
+            315,
+            0,
+            316,
+            0,
+            "STRING"
+        ],
+        [
+            461,
+            313,
+            0,
+            316,
+            1,
+            "DG_CONTEXT"
+        ],
+        [
+            462,
+            314,
+            0,
+            316,
+            2,
+            "DG_TARGET_PROFILE_CONFIG"
+        ],
+        [
+            473,
+            325,
+            0,
+            314,
+            0,
+            "FLOAT"
+        ],
+        [
+            474,
+            325,
+            0,
+            105,
+            5,
+            "FLOAT"
+        ],
+        [
+            475,
+            316,
+            10,
+            105,
+            3,
+            "INT"
+        ],
+        [
+            476,
+            316,
+            11,
+            105,
+            4,
+            "INT"
+        ],
+        [
+            477,
+            316,
+            13,
+            105,
+            2,
+            "STRING"
+        ],
+        [
+            478,
+            316,
+            13,
+            326,
+            0,
+            "STRING"
+        ],
+        [
+            479,
+            316,
+            4,
+            327,
+            0,
+            "STRING"
+        ]
+    ],
+    "groups": [
+        {
+            "id": 1,
+            "title": "INPUTS",
+            "bounding": [
+                -40,
+                -40,
+                570,
+                880
+            ],
+            "color": "#3f789e",
+            "font_size": 24,
+            "flags": {}
+        },
+        {
+            "id": 2,
+            "title": "DIFFUSIONGEMMA DIRECTOR \u2014 MINIMAX-H3",
+            "bounding": [
+                520,
+                -40,
+                990,
+                1040
+            ],
+            "color": "#8b5a2b",
+            "font_size": 24,
+            "flags": {}
+        },
+        {
+            "id": 3,
+            "title": "NATIVE MINIMAX-H3 GENERATION",
+            "bounding": [
+                1510,
+                -40,
+                1190,
+                1200
+            ],
+            "color": "#557755",
+            "font_size": 24,
+            "flags": {}
+        }
+    ],
+    "definitions": {
+        "subgraphs": [
+            {
+                "id": "4f4334ed-1aa4-5860-b237-56c863ab42d6",
+                "version": 1,
+                "state": {
+                    "lastGroupId": 6,
+                    "lastNodeId": 324,
+                    "lastLinkId": 472,
+                    "lastRerouteId": 0
+                },
+                "revision": 2,
+                "config": {},
+                "name": "Text to Video + Audio (MiniMax H3, SageAttention)",
+                "inputNode": {
+                    "id": -10,
+                    "bounding": [
+                        -2560,
+                        4720,
+                        128,
+                        268
+                    ]
+                },
+                "outputNode": {
+                    "id": -20,
+                    "bounding": [
+                        670,
+                        4780,
+                        128,
+                        68
+                    ]
+                },
+                "inputs": [
+                    {
+                        "id": "d6eaa195-2266-4016-b7ed-b2d17a3e53c2",
+                        "name": "first_frame",
+                        "type": "IMAGE",
+                        "linkIds": [
+                            195
+                        ],
+                        "pos": [
+                            -2456,
+                            4744
+                        ]
+                    },
+                    {
+                        "id": "03b95567-f496-4279-9d38-989dd34fa882",
+                        "name": "last_frame",
+                        "type": "IMAGE",
+                        "linkIds": [
+                            196
+                        ],
+                        "pos": [
+                            -2456,
+                            4764
+                        ]
+                    },
+                    {
+                        "id": "d7302ca7-24ed-44c4-8ac9-736540dab7fb",
+                        "name": "prompt",
+                        "type": "STRING",
+                        "linkIds": [
+                            197
+                        ],
+                        "pos": [
+                            -2456,
+                            4784
+                        ]
+                    },
+                    {
+                        "id": "709e2d94-3172-496f-ad86-7d672f3df568",
+                        "name": "width",
+                        "type": "INT",
+                        "linkIds": [
+                            200
+                        ],
+                        "pos": [
+                            -2456,
+                            4804
+                        ]
+                    },
+                    {
+                        "id": "8205b85b-19bb-47f6-8418-89cd42940c1d",
+                        "name": "height",
+                        "type": "INT",
+                        "linkIds": [
+                            201
+                        ],
+                        "pos": [
+                            -2456,
+                            4824
+                        ]
+                    },
+                    {
+                        "id": "a40e5e96-4307-4a8e-a4f8-13a27d4bc9d3",
+                        "name": "value_1",
+                        "type": "FLOAT",
+                        "linkIds": [
+                            206
+                        ],
+                        "label": "duration",
+                        "pos": [
+                            -2456,
+                            4844
+                        ]
+                    },
+                    {
+                        "id": "9f8734bc-c2b3-43b0-911e-e89560e8b777",
+                        "name": "noise_seed",
+                        "type": "INT",
+                        "linkIds": [
+                            207
+                        ],
+                        "pos": [
+                            -2456,
+                            4864
+                        ]
+                    },
+                    {
+                        "id": "9a6d2811-673c-47e0-aa94-dda7c83621e4",
+                        "name": "unet_name",
+                        "type": "COMBO",
+                        "linkIds": [
+                            221
+                        ],
+                        "pos": [
+                            -2456,
+                            4884
+                        ]
+                    },
+                    {
+                        "id": "7100a787-5536-4002-840b-1d2fd071010a",
+                        "name": "clip_name",
+                        "type": "COMBO",
+                        "linkIds": [
+                            227
+                        ],
+                        "pos": [
+                            -2456,
+                            4904
+                        ]
+                    },
+                    {
+                        "id": "d40632fd-e0ae-4319-b85d-5479db4d6e11",
+                        "name": "vae_name",
+                        "type": "COMBO",
+                        "linkIds": [
+                            223
+                        ],
+                        "pos": [
+                            -2456,
+                            4924
+                        ]
+                    },
+                    {
+                        "id": "2992852c-4b20-439a-8771-c866ec1996e6",
+                        "name": "vae_name_1",
+                        "type": "COMBO",
+                        "linkIds": [
+                            224
+                        ],
+                        "label": "audio_vae",
+                        "pos": [
+                            -2456,
+                            4944
+                        ]
+                    }
+                ],
+                "outputs": [
+                    {
+                        "id": "adb2611c-e490-4fca-8067-b718c992f8cc",
+                        "name": "VIDEO",
+                        "type": "VIDEO",
+                        "linkIds": [
+                            168
+                        ],
+                        "localized_name": "VIDEO",
+                        "pos": [
+                            694,
+                            4804
+                        ]
+                    }
+                ],
+                "widgets": [],
+                "nodes": [
+                    {
+                        "id": 11,
+                        "type": "VAELoader",
+                        "pos": [
+                            -2020,
+                            4970
+                        ],
+                        "size": [
+                            639.984375,
+                            70
+                        ],
+                        "flags": {},
+                        "order": 4,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "vae_name",
+                                "name": "vae_name",
+                                "type": "COMBO",
+                                "widget": {
+                                    "name": "vae_name"
+                                },
+                                "link": 223
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "VAE",
+                                "name": "VAE",
+                                "type": "VAE",
+                                "links": [
+                                    8,
+                                    190
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "VAELoader",
+                            "models": [
+                                {
+                                    "name": "minimax_h3_video_vae_fp16.safetensors",
+                                    "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors",
+                                    "directory": "vae"
+                                }
+                            ]
+                        },
+                        "widgets_values": [
+                            "minimax_h3_video_vae_fp16.safetensors"
+                        ]
+                    },
+                    {
+                        "id": 24,
+                        "type": "VAELoader",
+                        "pos": [
+                            -2020,
+                            5100
+                        ],
+                        "size": [
+                            650,
+                            70
+                        ],
+                        "flags": {},
+                        "order": 10,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "vae_name",
+                                "name": "vae_name",
+                                "type": "COMBO",
+                                "widget": {
+                                    "name": "vae_name"
+                                },
+                                "link": 224
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "VAE",
+                                "name": "VAE",
+                                "type": "VAE",
+                                "links": [
+                                    23
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "VAELoader",
+                            "models": [
+                                {
+                                    "name": "minimax_h3_audio_vae_fp32.safetensors",
+                                    "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors",
+                                    "directory": "vae"
+                                }
+                            ]
+                        },
+                        "widgets_values": [
+                            "minimax_h3_audio_vae_fp32.safetensors"
+                        ]
+                    },
+                    {
+                        "id": 23,
+                        "type": "VAEDecodeAudio",
+                        "pos": [
+                            -50,
+                            4880
+                        ],
+                        "size": [
+                            229.984375,
+                            59.984375
+                        ],
+                        "flags": {
+                            "collapsed": false
+                        },
+                        "order": 9,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "samples",
+                                "name": "samples",
+                                "type": "LATENT",
+                                "link": 226
+                            },
+                            {
+                                "localized_name": "vae",
+                                "name": "vae",
+                                "type": "VAE",
+                                "link": 23
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "AUDIO",
+                                "name": "AUDIO",
+                                "type": "AUDIO",
+                                "links": [
+                                    166
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "VAEDecodeAudio"
+                        },
+                        "widgets_values": []
+                    },
+                    {
+                        "id": 10,
+                        "type": "VAEDecode",
+                        "pos": [
+                            -50,
+                            4760
+                        ],
+                        "size": [
+                            229.984375,
+                            59.984375
+                        ],
+                        "flags": {
+                            "collapsed": false
+                        },
+                        "order": 3,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "samples",
+                                "name": "samples",
+                                "type": "LATENT",
+                                "link": 225
+                            },
+                            {
+                                "localized_name": "vae",
+                                "name": "vae",
+                                "type": "VAE",
+                                "link": 8
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "IMAGE",
+                                "name": "IMAGE",
+                                "type": "IMAGE",
+                                "links": [
+                                    167
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "VAEDecode"
+                        },
+                        "widgets_values": []
+                    },
+                    {
+                        "id": 17,
+                        "type": "KSamplerSelect",
+                        "pos": [
+                            -790,
+                            4910
+                        ],
+                        "size": [
+                            369.984375,
+                            70
+                        ],
+                        "flags": {},
+                        "order": 0,
+                        "mode": 0,
+                        "inputs": [],
+                        "outputs": [
+                            {
+                                "localized_name": "SAMPLER",
+                                "name": "SAMPLER",
+                                "type": "SAMPLER",
+                                "links": [
+                                    16
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "KSamplerSelect"
+                        },
+                        "widgets_values": [
+                            "res_multistep"
+                        ]
+                    },
+                    {
+                        "id": 9,
+                        "type": "BasicScheduler",
+                        "pos": [
+                            -790,
+                            5030
+                        ],
+                        "size": [
+                            369.984375,
+                            129.984375
+                        ],
+                        "flags": {},
+                        "order": 2,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "model",
+                                "name": "model",
+                                "type": "MODEL",
+                                "link": 229
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "SIGMAS",
+                                "name": "SIGMAS",
+                                "type": "SIGMAS",
+                                "links": [
+                                    18
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "BasicScheduler"
+                        },
+                        "widgets_values": [
+                            "simple",
+                            25,
+                            1.0
+                        ]
+                    },
+                    {
+                        "id": 14,
+                        "type": "SamplerCustomAdvanced",
+                        "pos": [
+                            -360,
+                            4820
+                        ],
+                        "size": [
+                            229.984375,
+                            139.984375
+                        ],
+                        "flags": {},
+                        "order": 6,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "noise",
+                                "name": "noise",
+                                "type": "NOISE",
+                                "link": 40
+                            },
+                            {
+                                "localized_name": "guider",
+                                "name": "guider",
+                                "type": "GUIDER",
+                                "link": 12
+                            },
+                            {
+                                "localized_name": "sampler",
+                                "name": "sampler",
+                                "type": "SAMPLER",
+                                "link": 16
+                            },
+                            {
+                                "localized_name": "sigmas",
+                                "name": "sigmas",
+                                "type": "SIGMAS",
+                                "link": 18
+                            },
+                            {
+                                "localized_name": "latent_image",
+                                "name": "latent_image",
+                                "type": "LATENT",
+                                "link": 188
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "output",
+                                "name": "output",
+                                "type": "LATENT",
+                                "links": [
+                                    225,
+                                    226
+                                ]
+                            },
+                            {
+                                "localized_name": "denoised_output",
+                                "name": "denoised_output",
+                                "type": "LATENT",
+                                "links": null
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "SamplerCustomAdvanced"
+                        },
+                        "widgets_values": []
+                    },
+                    {
+                        "id": 16,
+                        "type": "BasicGuider",
+                        "pos": [
+                            -790,
+                            4800
+                        ],
+                        "size": [
+                            359.984375,
+                            59.984375
+                        ],
+                        "flags": {},
+                        "order": 8,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "model",
+                                "name": "model",
+                                "type": "MODEL",
+                                "link": 230
+                            },
+                            {
+                                "localized_name": "conditioning",
+                                "name": "conditioning",
+                                "type": "CONDITIONING",
+                                "link": 187
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "GUIDER",
+                                "name": "GUIDER",
+                                "type": "GUIDER",
+                                "links": [
+                                    12
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "BasicGuider"
+                        },
+                        "widgets_values": []
+                    },
+                    {
+                        "id": 6,
+                        "type": "UNETLoader",
+                        "pos": [
+                            -2020,
+                            4630
+                        ],
+                        "size": [
+                            639.984375,
+                            89.984375
+                        ],
+                        "flags": {},
+                        "order": 1,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "unet_name",
+                                "name": "unet_name",
+                                "type": "COMBO",
+                                "widget": {
+                                    "name": "unet_name"
+                                },
+                                "link": 221
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "MODEL",
+                                "name": "MODEL",
+                                "type": "MODEL",
+                                "links": [
+                                    228
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "UNETLoader",
+                            "models": [
+                                {
+                                    "name": "minimax_h3_fl2va_pruned_int8_convrot.safetensors",
+                                    "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors",
+                                    "directory": "diffusion_models"
+                                }
+                            ]
+                        },
+                        "widgets_values": [
+                            "minimax_h3_fl2va_pruned_int8_convrot.safetensors",
+                            "default"
+                        ]
+                    },
+                    {
+                        "id": 13,
+                        "type": "CLIPLoader",
+                        "pos": [
+                            -2020,
+                            4780
+                        ],
+                        "size": [
+                            639.984375,
+                            120
+                        ],
+                        "flags": {},
+                        "order": 5,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "clip_name",
+                                "name": "clip_name",
+                                "type": "COMBO",
+                                "widget": {
+                                    "name": "clip_name"
+                                },
+                                "link": 227
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "CLIP",
+                                "name": "CLIP",
+                                "type": "CLIP",
+                                "links": [
+                                    189
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "CLIPLoader",
+                            "models": [
+                                {
+                                    "name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+                                    "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+                                    "directory": "text_encoders"
+                                }
+                            ]
+                        },
+                        "widgets_values": [
+                            "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+                            "minimax",
+                            "default"
+                        ]
+                    },
+                    {
+                        "id": 15,
+                        "type": "RandomNoise",
+                        "pos": [
+                            -790,
+                            4660
+                        ],
+                        "size": [
+                            359.984375,
+                            89.984375
+                        ],
+                        "flags": {},
+                        "order": 7,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "noise_seed",
+                                "name": "noise_seed",
+                                "type": "INT",
+                                "widget": {
+                                    "name": "noise_seed"
+                                },
+                                "link": 207
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "NOISE",
+                                "name": "NOISE",
+                                "type": "NOISE",
+                                "links": [
+                                    40
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "RandomNoise"
+                        },
+                        "widgets_values": [
+                            42,
+                            "randomize"
+                        ]
+                    },
+                    {
+                        "id": 91,
+                        "type": "CreateVideo",
+                        "pos": [
+                            260,
+                            4790
+                        ],
+                        "size": [
+                            269.984375,
+                            109.984375
+                        ],
+                        "flags": {},
+                        "order": 11,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "images",
+                                "name": "images",
+                                "type": "IMAGE",
+                                "link": 167
+                            },
+                            {
+                                "localized_name": "audio",
+                                "name": "audio",
+                                "shape": 7,
+                                "type": "AUDIO",
+                                "link": 166
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "VIDEO",
+                                "name": "VIDEO",
+                                "type": "VIDEO",
+                                "links": [
+                                    168
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "CreateVideo"
+                        },
+                        "widgets_values": [
+                            24,
+                            8
+                        ]
+                    },
+                    {
+                        "id": 104,
+                        "type": "MiniMaxH3ImageToVideo",
+                        "pos": [
+                            -1290,
+                            4650
+                        ],
+                        "size": [
+                            409.984375,
+                            509.984375
+                        ],
+                        "flags": {},
+                        "order": 12,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "clip",
+                                "name": "clip",
+                                "type": "CLIP",
+                                "link": 189
+                            },
+                            {
+                                "localized_name": "vae",
+                                "name": "vae",
+                                "type": "VAE",
+                                "link": 190
+                            },
+                            {
+                                "localized_name": "first_frame",
+                                "name": "first_frame",
+                                "shape": 7,
+                                "type": "IMAGE",
+                                "link": 195
+                            },
+                            {
+                                "localized_name": "last_frame",
+                                "name": "last_frame",
+                                "shape": 7,
+                                "type": "IMAGE",
+                                "link": 196
+                            },
+                            {
+                                "localized_name": "prompt",
+                                "name": "prompt",
+                                "type": "STRING",
+                                "widget": {
+                                    "name": "prompt"
+                                },
+                                "link": 197
+                            },
+                            {
+                                "localized_name": "width",
+                                "name": "width",
+                                "type": "INT",
+                                "widget": {
+                                    "name": "width"
+                                },
+                                "link": 200
+                            },
+                            {
+                                "localized_name": "height",
+                                "name": "height",
+                                "type": "INT",
+                                "widget": {
+                                    "name": "height"
+                                },
+                                "link": 201
+                            },
+                            {
+                                "localized_name": "length",
+                                "name": "length",
+                                "type": "INT",
+                                "widget": {
+                                    "name": "length"
+                                },
+                                "link": 199
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "positive",
+                                "name": "positive",
+                                "type": "CONDITIONING",
+                                "links": [
+                                    187
+                                ]
+                            },
+                            {
+                                "localized_name": "LATENT",
+                                "name": "LATENT",
+                                "type": "LATENT",
+                                "links": [
+                                    188
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "MiniMaxH3ImageToVideo"
+                        },
+                        "widgets_values": [
+                            "",
+                            864,
+                            480,
+                            481
+                        ]
+                    },
+                    {
+                        "id": 107,
+                        "type": "ComfyMathExpression",
+                        "pos": [
+                            -1710,
+                            5300
+                        ],
+                        "size": [
+                            359.984375,
+                            159.984375
+                        ],
+                        "flags": {
+                            "collapsed": false
+                        },
+                        "order": 13,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "label": "a",
+                                "localized_name": "values.a",
+                                "name": "values.a",
+                                "type": "FLOAT,INT,BOOLEAN",
+                                "link": 205
+                            },
+                            {
+                                "label": "b",
+                                "localized_name": "values.b",
+                                "name": "values.b",
+                                "shape": 7,
+                                "type": "FLOAT,INT,BOOLEAN",
+                                "link": null
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "FLOAT",
+                                "name": "FLOAT",
+                                "type": "FLOAT",
+                                "links": null
+                            },
+                            {
+                                "localized_name": "INT",
+                                "name": "INT",
+                                "type": "INT",
+                                "links": [
+                                    199
+                                ]
+                            },
+                            {
+                                "localized_name": "BOOL",
+                                "name": "BOOL",
+                                "type": "BOOLEAN",
+                                "links": null
+                            }
+                        ],
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "ComfyMathExpression"
+                        },
+                        "widgets_values": [
+                            "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17"
+                        ]
+                    },
+                    {
+                        "id": 111,
+                        "type": "PrimitiveFloat",
+                        "pos": [
+                            -2020,
+                            5300
+                        ],
+                        "size": [
+                            269.984375,
+                            70
+                        ],
+                        "flags": {},
+                        "order": 14,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "value",
+                                "name": "value",
+                                "type": "FLOAT",
+                                "widget": {
+                                    "name": "value"
+                                },
+                                "link": 206
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "FLOAT",
+                                "name": "FLOAT",
+                                "type": "FLOAT",
+                                "links": [
+                                    205
+                                ]
+                            }
+                        ],
+                        "title": "Float (duration)",
+                        "properties": {
+                            "cnr_id": "comfy-core",
+                            "ver": "0.30.1",
+                            "Node name for S&R": "PrimitiveFloat"
+                        },
+                        "widgets_values": [
+                            2
+                        ]
+                    },
+                    {
+                        "id": 119,
+                        "type": "RHMiniMaxH3SageAttentionPatch",
+                        "pos": [
+                            -1291.1626922066582,
+                            5296.506970111788
+                        ],
+                        "size": [
+                            419.984375,
+                            99.984375
+                        ],
+                        "flags": {},
+                        "order": 15,
+                        "mode": 0,
+                        "inputs": [
+                            {
+                                "localized_name": "model",
+                                "name": "model",
+                                "type": "MODEL",
+                                "link": 228
+                            }
+                        ],
+                        "outputs": [
+                            {
+                                "localized_name": "model",
+                                "name": "model",
+                                "type": "MODEL",
+                                "links": [
+                                    229,
+                                    230
+                                ]
+                            }
+                        ],
+                        "properties": {
+                            "aux_id": "HM-RunningHub/ComfyUI_RH_MinMaxH3",
+                            "ver": "445fccb8f4e2d2a866d446a6e5589dddc5f0ec6a",
+                            "Node name for S&R": "RHMiniMaxH3SageAttentionPatch"
+                        },
+                        "widgets_values": [
+                            true
+                        ]
+                    }
+                ],
+                "groups": [
+                    {
+                        "id": 1,
+                        "title": "Models",
+                        "bounding": [
+                            -2050,
+                            4540,
+                            700,
+                            670
+                        ],
+                        "color": "#3f789e",
+                        "flags": {}
+                    },
+                    {
+                        "id": 2,
+                        "title": "Sampling",
+                        "bounding": [
+                            -810,
+                            4540,
+                            690,
+                            670
+                        ],
+                        "color": "#3f789e",
+                        "flags": {}
+                    },
+                    {
+                        "id": 3,
+                        "title": "Conditioning",
+                        "bounding": [
+                            -1320,
+                            4540,
+                            480,
+                            670
+                        ],
+                        "color": "#3f789e",
+                        "flags": {}
+                    },
+                    {
+                        "id": 4,
+                        "title": "Decoding and create video",
+                        "bounding": [
+                            -90,
+                            4540,
+                            670,
+                            670
+                        ],
+                        "color": "#3f789e",
+                        "flags": {}
+                    },
+                    {
+                        "id": 5,
+                        "title": "Bundled SageAttention backend",
+                        "bounding": [
+                            -1320,
+                            5220,
+                            480,
+                            220
+                        ],
+                        "color": "#6f4b9e",
+                        "flags": {}
+                    }
+                ],
+                "links": [
+                    {
+                        "id": 23,
+                        "origin_id": 24,
+                        "origin_slot": 0,
+                        "target_id": 23,
+                        "target_slot": 1,
+                        "type": "VAE"
+                    },
+                    {
+                        "id": 8,
+                        "origin_id": 11,
+                        "origin_slot": 0,
+                        "target_id": 10,
+                        "target_slot": 1,
+                        "type": "VAE"
+                    },
+                    {
+                        "id": 40,
+                        "origin_id": 15,
+                        "origin_slot": 0,
+                        "target_id": 14,
+                        "target_slot": 0,
+                        "type": "NOISE"
+                    },
+                    {
+                        "id": 12,
+                        "origin_id": 16,
+                        "origin_slot": 0,
+                        "target_id": 14,
+                        "target_slot": 1,
+                        "type": "GUIDER"
+                    },
+                    {
+                        "id": 16,
+                        "origin_id": 17,
+                        "origin_slot": 0,
+                        "target_id": 14,
+                        "target_slot": 2,
+                        "type": "SAMPLER"
+                    },
+                    {
+                        "id": 18,
+                        "origin_id": 9,
+                        "origin_slot": 0,
+                        "target_id": 14,
+                        "target_slot": 3,
+                        "type": "SIGMAS"
+                    },
+                    {
+                        "id": 188,
+                        "origin_id": 104,
+                        "origin_slot": 1,
+                        "target_id": 14,
+                        "target_slot": 4,
+                        "type": "LATENT"
+                    },
+                    {
+                        "id": 187,
+                        "origin_id": 104,
+                        "origin_slot": 0,
+                        "target_id": 16,
+                        "target_slot": 1,
+                        "type": "CONDITIONING"
+                    },
+                    {
+                        "id": 167,
+                        "origin_id": 10,
+                        "origin_slot": 0,
+                        "target_id": 91,
+                        "target_slot": 0,
+                        "type": "IMAGE"
+                    },
+                    {
+                        "id": 166,
+                        "origin_id": 23,
+                        "origin_slot": 0,
+                        "target_id": 91,
+                        "target_slot": 1,
+                        "type": "AUDIO"
+                    },
+                    {
+                        "id": 189,
+                        "origin_id": 13,
+                        "origin_slot": 0,
+                        "target_id": 104,
+                        "target_slot": 0,
+                        "type": "CLIP"
+                    },
+                    {
+                        "id": 190,
+                        "origin_id": 11,
+                        "origin_slot": 0,
+                        "target_id": 104,
+                        "target_slot": 1,
+                        "type": "VAE"
+                    },
+                    {
+                        "id": 168,
+                        "origin_id": 91,
+                        "origin_slot": 0,
+                        "target_id": -20,
+                        "target_slot": 0,
+                        "type": "VIDEO"
+                    },
+                    {
+                        "id": 195,
+                        "origin_id": -10,
+                        "origin_slot": 0,
+                        "target_id": 104,
+                        "target_slot": 2,
+                        "type": "IMAGE"
+                    },
+                    {
+                        "id": 196,
+                        "origin_id": -10,
+                        "origin_slot": 1,
+                        "target_id": 104,
+                        "target_slot": 3,
+                        "type": "IMAGE"
+                    },
+                    {
+                        "id": 197,
+                        "origin_id": -10,
+                        "origin_slot": 2,
+                        "target_id": 104,
+                        "target_slot": 4,
+                        "type": "STRING"
+                    },
+                    {
+                        "id": 199,
+                        "origin_id": 107,
+                        "origin_slot": 1,
+                        "target_id": 104,
+                        "target_slot": 7,
+                        "type": "INT"
+                    },
+                    {
+                        "id": 200,
+                        "origin_id": -10,
+                        "origin_slot": 3,
+                        "target_id": 104,
+                        "target_slot": 5,
+                        "type": "INT"
+                    },
+                    {
+                        "id": 201,
+                        "origin_id": -10,
+                        "origin_slot": 4,
+                        "target_id": 104,
+                        "target_slot": 6,
+                        "type": "INT"
+                    },
+                    {
+                        "id": 205,
+                        "origin_id": 111,
+                        "origin_slot": 0,
+                        "target_id": 107,
+                        "target_slot": 0,
+                        "type": "FLOAT"
+                    },
+                    {
+                        "id": 206,
+                        "origin_id": -10,
+                        "origin_slot": 5,
+                        "target_id": 111,
+                        "target_slot": 0,
+                        "type": "FLOAT"
+                    },
+                    {
+                        "id": 207,
+                        "origin_id": -10,
+                        "origin_slot": 6,
+                        "target_id": 15,
+                        "target_slot": 0,
+                        "type": "INT"
+                    },
+                    {
+                        "id": 221,
+                        "origin_id": -10,
+                        "origin_slot": 7,
+                        "target_id": 6,
+                        "target_slot": 0,
+                        "type": "COMBO"
+                    },
+                    {
+                        "id": 223,
+                        "origin_id": -10,
+                        "origin_slot": 9,
+                        "target_id": 11,
+                        "target_slot": 0,
+                        "type": "COMBO"
+                    },
+                    {
+                        "id": 224,
+                        "origin_id": -10,
+                        "origin_slot": 10,
+                        "target_id": 24,
+                        "target_slot": 0,
+                        "type": "COMBO"
+                    },
+                    {
+                        "id": 225,
+                        "origin_id": 14,
+                        "origin_slot": 0,
+                        "target_id": 10,
+                        "target_slot": 0,
+                        "type": "LATENT"
+                    },
+                    {
+                        "id": 226,
+                        "origin_id": 14,
+                        "origin_slot": 0,
+                        "target_id": 23,
+                        "target_slot": 0,
+                        "type": "LATENT"
+                    },
+                    {
+                        "id": 227,
+                        "origin_id": -10,
+                        "origin_slot": 8,
+                        "target_id": 13,
+                        "target_slot": 0,
+                        "type": "COMBO"
+                    },
+                    {
+                        "id": 228,
+                        "origin_id": 6,
+                        "origin_slot": 0,
+                        "target_id": 119,
+                        "target_slot": 0,
+                        "type": "MODEL"
+                    },
+                    {
+                        "id": 229,
+                        "origin_id": 119,
+                        "origin_slot": 0,
+                        "target_id": 9,
+                        "target_slot": 0,
+                        "type": "MODEL"
+                    },
+                    {
+                        "id": 230,
+                        "origin_id": 119,
+                        "origin_slot": 0,
+                        "target_id": 16,
+                        "target_slot": 0,
+                        "type": "MODEL"
+                    }
+                ],
+                "extra": {}
+            }
+        ]
+    },
+    "config": {},
+    "extra": {
+        "frontendVersion": "1.47.12",
+        "VHS_latentpreview": false,
+        "VHS_latentpreviewrate": 0,
+        "VHS_MetadataImage": true,
+        "VHS_KeepIntermediate": true,
+        "workflow_note": "Director-to-native-H3 example: one shared duration (20 seconds), 25 sampling steps, H3 prompt output 13.",
+        "seed_widgets": {
+            "105": 42
+        },
+        "ds": {
+            "scale": 0.55,
+            "offset": [
+                190,
+                120
+            ]
+        }
+    },
+    "version": 0.4,
+    "seed_widgets": {
+        "105": 4
+    }
+}
diff --git a/examples/10_minimax_h3_ref2va_director.json b/examples/10_minimax_h3_ref2va_director.json
new file mode 100644
index 0000000..dc8bb32
--- /dev/null
+++ b/examples/10_minimax_h3_ref2va_director.json
@@ -0,0 +1,1881 @@
+{
+  "id": "b2a0a20e-8a76-4d91-90a5-2dce9c0f2a10",
+  "revision": 0,
+  "last_node_id": 151,
+  "last_link_id": 42,
+  "nodes": [
+    {
+      "id": 116,
+      "type": "MarkdownNote",
+      "pos": [
+        0,
+        980
+      ],
+      "size": [
+        3180,
+        510
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "# DiffusionGemma Director to native MiniMax-H3 (Ref2VA)\n\n1. Replace the creative brief and both reference images. Keep their order aligned with the manifest: <Picture 1> is the protagonist identity; <Picture 2> is the environment and 2D animation style. The references guide the full clip, not first/last keyframes.\n2. Set Duration seconds. The same value controls Director planning and native H3 frame math. The sample is 20 seconds: at 24 fps it resolves to 481 valid frames on H3's 17k+5 temporal grid.\n3. Queue the workflow. DiffusionGemma unloads before H3 sampling. The generated prompt appears in the prompt preview; ComfyUI does not copy a linked runtime value into the native node's static prompt widget. A fail-closed readiness gate now prevents H3 conditioning and sampling when validation rejects or blanks the prompt.\n\nThe Director analysis batch resizes <Picture 2> to <Picture 1> only for multimodal inspection. Both original images feed the native H3 reference sockets separately and in the same order. This graph uses the Ref2VA transformer and 25 sampling steps."
+      ],
+      "color": "#222",
+      "bgcolor": "#000",
+      "title": "How this sample works"
+    },
+    {
+      "id": 138,
+      "type": "PrimitiveStringMultiline",
+      "pos": [
+        0,
+        0
+      ],
+      "size": [
+        500,
+        320
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": [1]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "PrimitiveStringMultiline"
+      },
+      "widgets_values": [
+        "Create a 20-second 2D hand-drawn animated monster-duel short in a late-1800s spaghetti-western town square. Use <Picture 1> only for the protagonist's identity, appearance, wardrobe, and accessories. Use <Picture 2> only for the town-square environment, palette, lighting, and visual style; do not transfer any pictured subject identity. Stage a coherent three-beat confrontation with stable character designs, clear action and reaction, synchronized impacts, deliberate camera movement, and a resolved final frame. No subtitles, logos, or watermarks."
+      ],
+      "title": "1. Creative brief - state 2D animation explicitly"
+    },
+    {
+      "id": 132,
+      "type": "PrimitiveFloat",
+      "pos": [
+        0,
+        370
+      ],
+      "size": [
+        330,
+        70
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "FLOAT",
+          "type": "FLOAT",
+          "links": [
+            11,
+            17
+          ]
+        }
+      ],
+      "title": "2. Duration seconds (shared)",
+      "properties": {
+        "Node name for S&R": "PrimitiveFloat"
+      },
+      "widgets_values": [
+        20
+      ],
+      "color": "#322",
+      "bgcolor": "#533"
+    },
+    {
+      "id": 137,
+      "type": "LoadImage",
+      "pos": [
+        540,
+        0
+      ],
+      "size": [
+        290,
+        330
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            2,
+            18
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "example.png",
+        "image"
+      ],
+      "title": "2a. <Picture 1> - protagonist identity"
+    },
+    {
+      "id": 148,
+      "type": "LoadImage",
+      "pos": [
+        540,
+        310
+      ],
+      "size": [
+        290,
+        330
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            3,
+            19
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "example.png",
+        "image"
+      ],
+      "title": "2b. <Picture 2> - environment + 2D style"
+    },
+    {
+      "id": 147,
+      "type": "BatchImagesNode",
+      "pos": [
+        540,
+        650
+      ],
+      "size": [
+        300,
+        130
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "image0",
+          "localized_name": "images.image0",
+          "name": "images.image0",
+          "type": "IMAGE",
+          "link": 2
+        },
+        {
+          "label": "image1",
+          "localized_name": "images.image1",
+          "name": "images.image1",
+          "type": "IMAGE",
+          "link": 3
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "IMAGE",
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [4]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "BatchImagesNode"
+      },
+      "title": "Analyze references in native order"
+    },
+    {
+      "id": 142,
+      "type": "DiffusionGemmaModelLoader",
+      "pos": [
+        1120,
+        0
+      ],
+      "size": [
+        500,
+        320
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "model_config",
+          "type": "DG_MODEL_CONFIG",
+          "links": [5]
+        },
+        {
+          "name": "runtime_status_json",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaModelLoader"
+      },
+      "widgets_values": [
+        "C:\\ComfyUI\\app\\models\\LLM\\diffusiongemma-26B-A4B-it-NVFP4",
+        "transformers_inprocess",
+        "auto",
+        "modelopt_nvfp4",
+        true,
+        "unload_after_run",
+        "20.0 GB - safe practical",
+        0.35
+      ],
+      "title": "3. Load Director (unloads before H3)"
+    },
+    {
+      "id": 143,
+      "type": "DiffusionGemmaH3ReferenceContext",
+      "pos": [
+        1660,
+        0
+      ],
+      "size": [
+        470,
+        590
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "user_prompt",
+          "type": "STRING",
+          "link": 1
+        },
+        {
+          "name": "reference_images",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": 4
+        },
+        {
+          "name": "reference_video",
+          "shape": 7,
+          "type": "VIDEO",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "gemma_context",
+          "type": "DG_CONTEXT",
+          "links": [
+            6,
+            9
+          ]
+        },
+        {
+          "name": "context_json",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "context_preview",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaH3ReferenceContext"
+      },
+      "widgets_values": [
+        "<Picture 1>: protagonist identity, face, body, silhouette, colors, wardrobe, accessories, and distinguishing traits; preserve exactly across every shot; this is a full-sequence reference, not a first or last frame.\n<Picture 2>: town-square environment, late-1800s spaghetti-western production design, palette, lighting, and 2D animation style; do not copy any pictured person or creature identity; this is a full-sequence reference, not a keyframe.",
+        1,
+        60,
+        60,
+        "trim",
+        ""
+      ],
+      "title": "4. H3 ordered reference context"
+    },
+    {
+      "id": 144,
+      "type": "DiffusionGemmaTargetProfile",
+      "pos": [
+        1120,
+        380
+      ],
+      "size": [
+        470,
+        590
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "target_duration_seconds",
+          "name": "target_duration_seconds",
+          "type": "FLOAT",
+          "widget": {
+            "name": "target_duration_seconds"
+          },
+          "link": 11
+        }
+      ],
+      "outputs": [
+        {
+          "name": "target_profile_config",
+          "type": "DG_TARGET_PROFILE_CONFIG",
+          "links": [
+            7,
+            10
+          ]
+        },
+        {
+          "name": "target_profile_json",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaTargetProfile"
+      },
+      "widgets_values": [
+        "minimax_h3",
+        "auto_scene_audio",
+        "",
+        20,
+        "",
+        "16:9",
+        "",
+        "",
+        false,
+        "auto",
+        "",
+        "ref2va"
+      ],
+      "title": "5. MiniMax-H3 Ref2VA target profile"
+    },
+    {
+      "id": 145,
+      "type": "DiffusionGemmaCoTGenerator",
+      "pos": [
+        2200,
+        0
+      ],
+      "size": [
+        430,
+        304
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model_config",
+          "type": "DG_MODEL_CONFIG",
+          "link": 5
+        },
+        {
+          "name": "gemma_context",
+          "type": "DG_CONTEXT",
+          "link": 6
+        },
+        {
+          "name": "target_profile_config",
+          "type": "DG_TARGET_PROFILE_CONFIG",
+          "link": 7
+        }
+      ],
+      "outputs": [
+        {
+          "name": "final_json",
+          "type": "STRING",
+          "links": [8]
+        },
+        {
+          "name": "reasoning_text",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "raw_response",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "metadata_json",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaCoTGenerator"
+      },
+      "widgets_values": [
+        0.45,
+        "cinematic",
+        0.7,
+        "on",
+        2048
+      ],
+      "title": "6. Write strict H3 Ref2VA prompt"
+    },
+    {
+      "id": 146,
+      "type": "DiffusionGemmaJSONSplitter",
+      "pos": [
+        2200,
+        350
+      ],
+      "size": [
+        430,
+        510
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "final_json",
+          "type": "STRING",
+          "link": 8
+        },
+        {
+          "name": "gemma_context",
+          "shape": 7,
+          "type": "DG_CONTEXT",
+          "link": 9
+        },
+        {
+          "name": "target_profile_config",
+          "shape": 7,
+          "type": "DG_TARGET_PROFILE_CONFIG",
+          "link": 10
+        }
+      ],
+      "outputs": [
+        {
+          "name": "ltx_prompt",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "ideogram_prompt",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "negative_prompt",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "aspect_ratio",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "metadata_json",
+          "type": "STRING",
+          "links": [
+            13,
+            41
+          ]
+        },
+        {
+          "name": "scene_segments_json",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "local_prompts",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "segment_lengths",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "is_valid",
+          "type": "BOOLEAN",
+          "links": null
+        },
+        {
+          "name": "resolution_selector_preset",
+          "type": "STRING",
+          "links": null
+        },
+        {
+          "name": "resolution_width",
+          "type": "INT",
+          "links": [15]
+        },
+        {
+          "name": "resolution_height",
+          "type": "INT",
+          "links": [16]
+        },
+        {
+          "name": "ready_for_generation",
+          "type": "BOOLEAN",
+          "links": [40]
+        },
+        {
+          "name": "minimax_h3_prompt",
+          "type": "STRING",
+          "links": [
+            12,
+            14
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaJSONSplitter"
+      },
+      "widgets_values": [
+        "",
+        0.4,
+        32
+      ],
+      "title": "7. Route Ref2VA prompt + resolution"
+    },
+    {
+      "id": 149,
+      "type": "PreviewAny",
+      "pos": [
+        2690,
+        0
+      ],
+      "size": [
+        550,
+        340
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 14
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "title": "Generated MiniMax-H3 Ref2VA prompt",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.15.1",
+        "Node name for S&R": "PreviewAny",
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65
+      },
+      "widgets_values": [
+        null,
+        null,
+        null
+      ],
+      "color": "#2a363b",
+      "bgcolor": "#3f5159"
+    },
+    {
+      "id": 150,
+      "type": "PreviewAny",
+      "pos": [
+        2690,
+        400
+      ],
+      "size": [
+        550,
+        300
+      ],
+      "flags": {},
+      "order": 12,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "source",
+          "type": "*",
+          "link": 13
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "title": "H3 validation / blocked reasons",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.15.1",
+        "Node name for S&R": "PreviewAny",
+        "enableTabs": false,
+        "tabWidth": 65,
+        "tabXOffset": 10,
+        "hasSecondTab": false,
+        "secondTabText": "Send Back",
+        "secondTabOffset": 80,
+        "secondTabWidth": 65
+      },
+      "widgets_values": [
+        null,
+        null,
+        null
+      ],
+      "color": "#2a363b",
+      "bgcolor": "#3f5159"
+    },
+    {
+      "id": 151,
+      "type": "DiffusionGemmaGenerationGate",
+      "pos": [
+        2690,
+        740
+      ],
+      "size": [
+        550,
+        120
+      ],
+      "flags": {},
+      "order": 29,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "prompt",
+          "type": "STRING",
+          "link": 12
+        },
+        {
+          "name": "ready_for_generation",
+          "type": "BOOLEAN",
+          "link": 40
+        },
+        {
+          "name": "metadata_json",
+          "shape": 7,
+          "type": "STRING",
+          "link": 41
+        }
+      ],
+      "outputs": [
+        {
+          "name": "prompt",
+          "type": "STRING",
+          "links": [42]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "DiffusionGemmaGenerationGate"
+      },
+      "title": "8. Fail closed before expensive H3 render"
+    },
+    {
+      "id": 127,
+      "type": "UNETLoader",
+      "pos": [
+        3280,
+        180
+      ],
+      "size": [
+        640,
+        90
+      ],
+      "flags": {},
+      "order": 13,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [20]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "UNETLoader",
+        "models": [
+          {
+            "name": "minimax_h3_ref2va_int8_convrot.safetensors",
+            "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors",
+            "directory": "diffusion_models"
+          }
+        ]
+      },
+      "widgets_values": [
+        "minimax_h3_ref2va_int8_convrot.safetensors",
+        "default"
+      ]
+    },
+    {
+      "id": 128,
+      "type": "CLIPLoader",
+      "pos": [
+        3280,
+        330
+      ],
+      "size": [
+        640,
+        120
+      ],
+      "flags": {},
+      "order": 14,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [23]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CLIPLoader",
+        "models": [
+          {
+            "name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+            "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+            "directory": "text_encoders"
+          }
+        ]
+      },
+      "widgets_values": [
+        "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
+        "minimax",
+        "default"
+      ]
+    },
+    {
+      "id": 131,
+      "type": "ComfyMathExpression",
+      "pos": [
+        4000,
+        650
+      ],
+      "size": [
+        225,
+        0
+      ],
+      "flags": {
+        "collapsed": true
+      },
+      "order": 21,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "a",
+          "name": "values.a",
+          "type": "FLOAT,INT,BOOLEAN",
+          "link": 17
+        },
+        {
+          "label": "b",
+          "name": "values.b",
+          "shape": 7,
+          "type": "FLOAT,INT,BOOLEAN",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "FLOAT",
+          "type": "FLOAT",
+          "links": null
+        },
+        {
+          "name": "INT",
+          "type": "INT",
+          "links": [26]
+        },
+        {
+          "name": "BOOL",
+          "type": "BOOLEAN",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "ComfyMathExpression"
+      },
+      "widgets_values": [
+        "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17"
+      ]
+    },
+    {
+      "id": 119,
+      "type": "VAELoader",
+      "pos": [
+        3280,
+        500
+      ],
+      "size": [
+        640,
+        70
+      ],
+      "flags": {},
+      "order": 15,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            24,
+            35
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VAELoader",
+        "models": [
+          {
+            "name": "minimax_h3_video_vae_fp16.safetensors",
+            "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors",
+            "directory": "vae"
+          }
+        ]
+      },
+      "widgets_values": [
+        "minimax_h3_video_vae_fp16.safetensors"
+      ]
+    },
+    {
+      "id": 120,
+      "type": "VAELoader",
+      "pos": [
+        3280,
+        630
+      ],
+      "size": [
+        650,
+        70
+      ],
+      "flags": {},
+      "order": 16,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            25,
+            36
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VAELoader",
+        "models": [
+          {
+            "name": "minimax_h3_audio_vae_fp32.safetensors",
+            "url": "https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors",
+            "directory": "vae"
+          }
+        ]
+      },
+      "widgets_values": [
+        "minimax_h3_audio_vae_fp32.safetensors"
+      ]
+    },
+    {
+      "id": 141,
+      "type": "RHMiniMaxH3SageAttentionPatch",
+      "pos": [
+        3280,
+        0
+      ],
+      "size": [
+        640,
+        100
+      ],
+      "flags": {},
+      "order": 17,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 20
+        }
+      ],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            21,
+            22
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "RHMiniMaxH3SageAttentionPatch"
+      },
+      "widgets_values": [
+        true
+      ],
+      "title": "Bundled SageAttention Patch"
+    },
+    {
+      "id": 129,
+      "type": "RandomNoise",
+      "pos": [
+        4520,
+        80
+      ],
+      "size": [
+        360,
+        90
+      ],
+      "flags": {},
+      "order": 18,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [29]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        42,
+        "fixed"
+      ]
+    },
+    {
+      "id": 123,
+      "type": "KSamplerSelect",
+      "pos": [
+        4520,
+        230
+      ],
+      "size": [
+        370,
+        70
+      ],
+      "flags": {},
+      "order": 19,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [31]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "res_multistep"
+      ]
+    },
+    {
+      "id": 124,
+      "type": "BasicScheduler",
+      "pos": [
+        4520,
+        350
+      ],
+      "size": [
+        370,
+        130
+      ],
+      "flags": {},
+      "order": 20,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 21
+        }
+      ],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [32]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "BasicScheduler"
+      },
+      "widgets_values": [
+        "simple",
+        25,
+        1
+      ]
+    },
+    {
+      "id": 136,
+      "type": "MiniMaxH3ReferenceToVideo",
+      "pos": [
+        4000,
+        80
+      ],
+      "size": [
+        470,
+        570
+      ],
+      "flags": {},
+      "order": 22,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 23
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 24
+        },
+        {
+          "name": "audio_vae",
+          "type": "VAE",
+          "link": 25
+        },
+        {
+          "label": "ref_image_0",
+          "name": "ref_images.ref_image_0",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": 18
+        },
+        {
+          "label": "ref_image_1",
+          "name": "ref_images.ref_image_1",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": 19
+        },
+        {
+          "label": "ref_image_2",
+          "name": "ref_images.ref_image_2",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "label": "ref_video_0",
+          "name": "ref_videos.ref_video_0",
+          "shape": 7,
+          "type": "IMAGE",
+          "link": null
+        },
+        {
+          "label": "ref_video_audio_0",
+          "name": "ref_video_audios.ref_video_audio_0",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": null
+        },
+        {
+          "label": "ref_audio_0",
+          "name": "ref_audios.ref_audio_0",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": null
+        },
+        {
+          "name": "prompt",
+          "type": "STRING",
+          "widget": {
+            "name": "prompt"
+          },
+          "link": 42
+        },
+        {
+          "name": "width",
+          "type": "INT",
+          "widget": {
+            "name": "width"
+          },
+          "link": 15
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "widget": {
+            "name": "height"
+          },
+          "link": 16
+        },
+        {
+          "name": "length",
+          "type": "INT",
+          "widget": {
+            "name": "length"
+          },
+          "link": 26
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [27]
+        },
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [28]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "MiniMaxH3ReferenceToVideo"
+      },
+      "widgets_values": [
+        "",
+        864,
+        480,
+        481,
+        "match"
+      ],
+      "title": "8. Native MiniMax-H3 Ref2VA - ordered refs, 25 steps"
+    },
+    {
+      "id": 126,
+      "type": "BasicGuider",
+      "pos": [
+        4520,
+        510
+      ],
+      "size": [
+        360,
+        60
+      ],
+      "flags": {},
+      "order": 23,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 22
+        },
+        {
+          "name": "conditioning",
+          "type": "CONDITIONING",
+          "link": 27
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [30]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "BasicGuider"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 125,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        4910,
+        260
+      ],
+      "size": [
+        230,
+        140
+      ],
+      "flags": {},
+      "order": 24,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 29
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 30
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 31
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 32
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 28
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": [
+            33,
+            34
+          ]
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 122,
+      "type": "VAEDecode",
+      "pos": [
+        5260,
+        250
+      ],
+      "size": [
+        230,
+        60
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 25,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 33
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 35
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [37]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VAEDecode"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 121,
+      "type": "VAEDecodeAudio",
+      "pos": [
+        5260,
+        390
+      ],
+      "size": [
+        230,
+        60
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 26,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 34
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 36
+        }
+      ],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [38]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "VAEDecodeAudio"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 130,
+      "type": "CreateVideo",
+      "pos": [
+        5570,
+        320
+      ],
+      "size": [
+        270,
+        110
+      ],
+      "flags": {},
+      "order": 27,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 37
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 38
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [39]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "CreateVideo"
+      },
+      "widgets_values": [
+        24,
+        8
+      ]
+    },
+    {
+      "id": 92,
+      "type": "SaveVideo",
+      "pos": [
+        5900,
+        300
+      ],
+      "size": [
+        650,
+        126
+      ],
+      "flags": {},
+      "order": 28,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 39
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "links": null
+        }
+      ],
+      "properties": {},
+      "widgets_values": [
+        "video/minimax_h3/diffusiongemma_director_ref2va",
+        "mp4",
+        "h264"
+      ],
+      "title": "9. Save generated video"
+    }
+  ],
+  "links": [
+    [
+      1,
+      138,
+      0,
+      143,
+      0,
+      "STRING"
+    ],
+    [
+      2,
+      137,
+      0,
+      147,
+      0,
+      "IMAGE"
+    ],
+    [
+      3,
+      148,
+      0,
+      147,
+      1,
+      "IMAGE"
+    ],
+    [
+      4,
+      147,
+      0,
+      143,
+      1,
+      "IMAGE"
+    ],
+    [
+      5,
+      142,
+      0,
+      145,
+      0,
+      "DG_MODEL_CONFIG"
+    ],
+    [
+      6,
+      143,
+      0,
+      145,
+      1,
+      "DG_CONTEXT"
+    ],
+    [
+      7,
+      144,
+      0,
+      145,
+      2,
+      "DG_TARGET_PROFILE_CONFIG"
+    ],
+    [
+      8,
+      145,
+      0,
+      146,
+      0,
+      "STRING"
+    ],
+    [
+      9,
+      143,
+      0,
+      146,
+      1,
+      "DG_CONTEXT"
+    ],
+    [
+      10,
+      144,
+      0,
+      146,
+      2,
+      "DG_TARGET_PROFILE_CONFIG"
+    ],
+    [
+      11,
+      132,
+      0,
+      144,
+      0,
+      "FLOAT"
+    ],
+    [
+      12,
+      146,
+      13,
+      151,
+      0,
+      "STRING"
+    ],
+    [
+      13,
+      146,
+      4,
+      150,
+      0,
+      "STRING"
+    ],
+    [
+      14,
+      146,
+      13,
+      149,
+      0,
+      "STRING"
+    ],
+    [
+      15,
+      146,
+      10,
+      136,
+      10,
+      "INT"
+    ],
+    [
+      16,
+      146,
+      11,
+      136,
+      11,
+      "INT"
+    ],
+    [
+      17,
+      132,
+      0,
+      131,
+      0,
+      "FLOAT"
+    ],
+    [
+      18,
+      137,
+      0,
+      136,
+      3,
+      "IMAGE"
+    ],
+    [
+      19,
+      148,
+      0,
+      136,
+      4,
+      "IMAGE"
+    ],
+    [
+      20,
+      127,
+      0,
+      141,
+      0,
+      "MODEL"
+    ],
+    [
+      21,
+      141,
+      0,
+      124,
+      0,
+      "MODEL"
+    ],
+    [
+      22,
+      141,
+      0,
+      126,
+      0,
+      "MODEL"
+    ],
+    [
+      23,
+      128,
+      0,
+      136,
+      0,
+      "CLIP"
+    ],
+    [
+      24,
+      119,
+      0,
+      136,
+      1,
+      "VAE"
+    ],
+    [
+      25,
+      120,
+      0,
+      136,
+      2,
+      "VAE"
+    ],
+    [
+      26,
+      131,
+      1,
+      136,
+      12,
+      "INT"
+    ],
+    [
+      27,
+      136,
+      0,
+      126,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      28,
+      136,
+      1,
+      125,
+      4,
+      "LATENT"
+    ],
+    [
+      29,
+      129,
+      0,
+      125,
+      0,
+      "NOISE"
+    ],
+    [
+      30,
+      126,
+      0,
+      125,
+      1,
+      "GUIDER"
+    ],
+    [
+      31,
+      123,
+      0,
+      125,
+      2,
+      "SAMPLER"
+    ],
+    [
+      32,
+      124,
+      0,
+      125,
+      3,
+      "SIGMAS"
+    ],
+    [
+      33,
+      125,
+      0,
+      122,
+      0,
+      "LATENT"
+    ],
+    [
+      34,
+      125,
+      0,
+      121,
+      0,
+      "LATENT"
+    ],
+    [
+      35,
+      119,
+      0,
+      122,
+      1,
+      "VAE"
+    ],
+    [
+      36,
+      120,
+      0,
+      121,
+      1,
+      "VAE"
+    ],
+    [
+      37,
+      122,
+      0,
+      130,
+      0,
+      "IMAGE"
+    ],
+    [
+      38,
+      121,
+      0,
+      130,
+      1,
+      "AUDIO"
+    ],
+    [
+      39,
+      130,
+      0,
+      92,
+      0,
+      "VIDEO"
+    ],
+    [
+      40,
+      146,
+      12,
+      151,
+      1,
+      "BOOLEAN"
+    ],
+    [
+      41,
+      146,
+      4,
+      151,
+      2,
+      "STRING"
+    ],
+    [
+      42,
+      151,
+      0,
+      136,
+      9,
+      "STRING"
+    ]
+  ],
+  "groups": [
+    {
+      "id": 1,
+      "title": "INPUTS - brief, duration, ordered references",
+      "bounding": [
+        -40,
+        -40,
+        1100,
+        900
+      ],
+      "color": "#3f789e",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 2,
+      "title": "DIFFUSIONGEMMA DIRECTOR - Ref2VA prompt",
+      "bounding": [
+        1080,
+        -40,
+        2150,
+        900
+      ],
+      "color": "#8a6f3d",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 3,
+      "title": "MINIMAX-H3 REF2VA MODELS",
+      "bounding": [
+        3240,
+        -40,
+        720,
+        850
+      ],
+      "color": "#6a4c93",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 4,
+      "title": "NATIVE REF2VA CONDITIONING",
+      "bounding": [
+        3960,
+        -40,
+        500,
+        900
+      ],
+      "color": "#3f789e",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 5,
+      "title": "25-STEP SAMPLING",
+      "bounding": [
+        4480,
+        -40,
+        750,
+        900
+      ],
+      "color": "#355c4a",
+      "font_size": 24,
+      "flags": {}
+    },
+    {
+      "id": 6,
+      "title": "DECODE + SAVE",
+      "bounding": [
+        5220,
+        -40,
+        1400,
+        900
+      ],
+      "color": "#75524a",
+      "font_size": 24,
+      "flags": {}
+    }
+  ],
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.75,
+      "offset": [
+        -20,
+        80
+      ]
+    },
+    "frontendVersion": "1.47.11",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true,
+    "workflow_note": "DiffusionGemma Director Ref2VA sample: two ordered picture references, native MiniMaxH3ReferenceToVideo, shared 20-second duration (481 frames at 24 fps), generated prompt preview plus fail-closed readiness gate, 25 sampling steps."
+  },
+  "version": 0.4
+}
diff --git a/examples/README.md b/examples/README.md
index 23cb403..3d978fe 100644
--- a/examples/README.md
+++ b/examples/README.md
@@ -1,6 +1,6 @@
-# DiffusionGemma Example Workflow
+# DiffusionGemma Example Workflows
 
-This package ships one release workflow:
+The established character-motion release workflow is:
 
 ```text
 07_ltx23_character_motion_transfer.json
@@ -47,3 +47,56 @@ examples/assets/character_motion_transfer/diffusiongemma_ltx_character_demo.mp4
 ```
 
 Wire downstream generation nodes from the JSON Splitter outputs only when `ready_for_generation=true`.
+
+## MiniMax H3 Director T2VA
+
+The native text-to-video/audio example is:
+
+```text
+09_minimax_h3_t2va_director.json
+```
+
+It converts a normal creative brief into Director's dedicated `minimax_h3_prompt`, wires the splitter's width and height outputs into native MiniMax-H3, and shares one editable duration between the Director profile and H3 frame calculation. Parallel previews show the generated H3 prompt and the splitter metadata/blocked reasons without interrupting the direct prompt connection. The included preset is 20 seconds, 481 frames at 24 fps, 0.4 MP at multiple-of-32 resolution, and 25 sampling steps.
+
+The workflow uses the native `MiniMaxH3ImageToVideo` path plus `RHMiniMaxH3SageAttentionPatch`; install `ComfyUI_RH_MinMaxH3` and place the four model files selected in the embedded H3 group in their normal ComfyUI model folders. DiffusionGemma is set to unload after prompt generation so H3 can reclaim VRAM before sampling.
+
+## MiniMax H3 Director Ref2VA
+
+The ordered-reference example is:
+
+```text
+10_minimax_h3_ref2va_director.json
+```
+
+It uses two images in explicit `<Picture 1>` / `<Picture 2>` order, analyzes that same ordered batch with `DiffusionGemmaH3ReferenceContext`, and sends the dedicated Ref2VA prompt directly to native `MiniMaxH3ReferenceToVideo`. The sample creative brief explicitly requests 2D hand-drawn animation and assigns protagonist identity to Picture 1 and environment/style to Picture 2, preventing vague “inspired” wording from drifting into generic CGI subjects.
+
+One editable duration is shared by Director planning and native frame calculation; the included setting is 20 seconds, 481 frames at 24 fps, with 25 sampling steps. The Director uses its full 2048-token output budget because the official Ref2VA representation includes subject definitions, retention analysis, and a detailed timeline in addition to audio sections. Replace both `example.png` placeholders before running and keep their socket order aligned with the reference manifest. The Director loader unloads before H3 sampling, and both original images feed the native Ref2VA node separately.
+
+## LTX 2.3 Reframe with selectable audio
+
+The provisional, unexported repository prototype is:
+
+```text
+gemmaREFRAME_Experimental_audio_on-off.json
+```
+
+The current package entry point does not register the `LTXReframe*` nodes, and the retired layout frontend is not shipped. Keep these reframe artifacts separate from the stable Director examples when preparing a Registry release.
+
+Its **Use Custom Vocal Audio?** switch defaults to off:
+
+- **Off:** the source video's audio is used for both LTX audio conditioning and the final video. Videos without an audio track use duration-matched silence.
+- **On:** upload a custom audio track, choose its start time in **Custom Track Window**, and use the MelBandRoFormer vocal stem. Long tracks are trimmed and short tracks are padded with silence to match the effective video range.
+
+Required custom-node dependencies:
+
+- `ComfyUI-MelBandRoFormer`
+- `comfyui-kjnodes`
+- `comfy-mtb`
+
+Required model:
+
+```text
+ComfyUI/models/diffusion_models/MelBandRoformer_fp16.safetensors
+```
+
+No custom music is bundled. Source-audio mode can remain selected without choosing a custom file; custom-audio mode reports a clear error until an audio file is uploaded or selected.
diff --git a/examples/gemmaREFRAME_Experimental_audio_on-off.json b/examples/gemmaREFRAME_Experimental_audio_on-off.json
new file mode 100644
index 0000000..a2c91c8
--- /dev/null
+++ b/examples/gemmaREFRAME_Experimental_audio_on-off.json
@@ -0,0 +1,7677 @@
+{
+  "id": "28666934-dc99-4884-932b-b72715bf49ce",
+  "revision": 1,
+  "last_node_id": 9113,
+  "last_link_id": 14999,
+  "nodes": [
+    {
+      "id": 5527,
+      "type": "CLIPTextEncode",
+      "pos": [
+        2575.390383856118,
+        3744.343183320789
+      ],
+      "size": [
+        637.234375,
+        229.125
+      ],
+      "flags": {},
+      "order": 13,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 14650
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            14808
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Positive Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "Re-create video with new dimensions; maintain spatial and temporal consistency."
+      ],
+      "color": "#232",
+      "bgcolor": "#353"
+    },
+    {
+      "id": 5528,
+      "type": "CLIPTextEncode",
+      "pos": [
+        2577.6981904891463,
+        4050.883080687562
+      ],
+      "size": [
+        627.171875,
+        204.515625
+      ],
+      "flags": {},
+      "order": 14,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 14651
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            14809
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Negative Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "pc game, console game, video game, cartoon, childish, ugly"
+      ],
+      "color": "#322",
+      "bgcolor": "#533"
+    },
+    {
+      "id": 5533,
+      "type": "ManualSigmas",
+      "pos": [
+        3633.623873133871,
+        4496.696824944616
+      ],
+      "size": [
+        275,
+        107.5
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            14816
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "ManualSigmas"
+      },
+      "widgets_values": [
+        "1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0"
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5534,
+      "type": "CFGGuider",
+      "pos": [
+        3636.5298689167216,
+        4304.7245415307825
+      ],
+      "size": [
+        268.75,
+        127.515625
+      ],
+      "flags": {},
+      "order": 23,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14655
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14656
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14657
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            14814
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5537,
+      "type": "LTXVLaplacianPyramidBlend",
+      "pos": [
+        3678.3045102368214,
+        3272.4291189665087
+      ],
+      "size": [
+        287.5,
+        179.484375
+      ],
+      "flags": {},
+      "order": 24,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_a",
+          "type": "IMAGE",
+          "link": 14899
+        },
+        {
+          "name": "image_b",
+          "type": "IMAGE",
+          "link": 14897
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14894
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14948
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVLaplacianPyramidBlend"
+      },
+      "widgets_values": [
+        true,
+        15
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5538,
+      "type": "KSamplerSelect",
+      "pos": [
+        3634.913561245836,
+        4156.955968329219
+      ],
+      "size": [
+        343.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            14815
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler_ancestral_cfg_pp"
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5539,
+      "type": "CreateVideo",
+      "pos": [
+        2606.9830617789557,
+        4523.002287812876
+      ],
+      "size": [
+        231.25,
+        103.515625
+      ],
+      "flags": {},
+      "order": 25,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14963
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "widget": {
+            "name": "fps"
+          },
+          "link": 14999
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 14666
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [
+            14711
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CreateVideo"
+      },
+      "widgets_values": [
+        25
+      ],
+      "title": "Create Video · Selected Audio + Source FPS"
+    },
+    {
+      "id": 5541,
+      "type": "LTXVImgToVideoConditionOnly",
+      "pos": [
+        3304.983004089232,
+        3735.7554941731955
+      ],
+      "size": [
+        312.5,
+        179.484375
+      ],
+      "flags": {},
+      "order": 26,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14668
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14669
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14896
+        },
+        {
+          "name": "bypass",
+          "shape": 7,
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "bypass"
+          },
+          "link": 14900
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14681
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVImgToVideoConditionOnly"
+      },
+      "widgets_values": [
+        0.7,
+        false
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5542,
+      "type": "LTXVPreprocess",
+      "pos": [
+        3306.869214304367,
+        3601.4181348043135
+      ],
+      "size": [
+        268.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 27,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14901
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output_image",
+          "type": "IMAGE",
+          "links": [
+            14669
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LTXVPreprocess"
+      },
+      "widgets_values": [
+        8
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5544,
+      "type": "LTXAddVideoICLoRAGuideAdvanced",
+      "pos": [
+        3634.9941529503017,
+        3615.716215524626
+      ],
+      "size": [
+        387.5,
+        443.359375
+      ],
+      "flags": {},
+      "order": 28,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14820
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14821
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14680
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14681
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14898
+        },
+        {
+          "name": "attention_mask",
+          "shape": 7,
+          "type": "MASK",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            14656,
+            14811
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            14657,
+            14812
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14817
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXAddVideoICLoRAGuideAdvanced"
+      },
+      "widgets_values": [
+        0,
+        1,
+        1,
+        "disabled",
+        false,
+        256,
+        64,
+        1
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5552,
+      "type": "SaveVideo",
+      "pos": [
+        4958.6991337986365,
+        2510.3077251206746
+      ],
+      "size": [
+        1413.171875,
+        2646.296875
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 29,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 14778
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "video/LTX_Reframe_Outpaint",
+        "mp4",
+        "auto"
+      ]
+    },
+    {
+      "id": 5553,
+      "type": "LTXReframeLayout",
+      "pos": [
+        216.0319099395074,
+        2816.9066157891793
+      ],
+      "size": [
+        1589.03125,
+        1531
+      ],
+      "flags": {},
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "padded_frames",
+          "type": "IMAGE",
+          "links": [
+            14980
+          ]
+        },
+        {
+          "name": "outpaint_mask",
+          "type": "MASK",
+          "links": [
+            14981
+          ]
+        },
+        {
+          "name": "source_mask",
+          "type": "MASK",
+          "links": [
+            14777,
+            14982
+          ]
+        },
+        {
+          "name": "source_frames",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            14990,
+            14991
+          ]
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "links": [
+            14810,
+            14999
+          ]
+        },
+        {
+          "name": "source_video",
+          "type": "VIDEO",
+          "links": [
+            14902
+          ]
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "links": [
+            14950
+          ]
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "links": [
+            14951
+          ]
+        },
+        {
+          "name": "placement_json",
+          "type": "STRING",
+          "links": []
+        }
+      ],
+      "title": "1. LTX Reframe Layout · Source, Framing & Frame Range",
+      "properties": {
+        "Node name for S&R": "LTXReframeLayout",
+        "ltx_reframe_range_v2": {
+          "limit_frames": false,
+          "first_frame": 0,
+          "last_frame": -1
+        }
+      },
+      "widgets_values": [
+        "shuffle.mp4",
+        "9:16",
+        1088,
+        1920,
+        0.335,
+        0.707,
+        1.549,
+        0,
+        false,
+        0,
+        -1,
+        "image",
+        ""
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 5557,
+      "type": "GrowMaskWithBlur",
+      "pos": [
+        4469.038454259336,
+        3254.2301742540003
+      ],
+      "size": [
+        362.5,
+        339.390625
+      ],
+      "flags": {},
+      "order": 30,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14988
+        }
+      ],
+      "outputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": [
+            14985
+          ]
+        },
+        {
+          "name": "mask_inverted",
+          "type": "MASK",
+          "links": []
+        }
+      ],
+      "title": "Source Seam · Inward 12px / Blur 8px",
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "GrowMaskWithBlur"
+      },
+      "widgets_values": [
+        -64,
+        0,
+        true,
+        false,
+        0,
+        1,
+        1,
+        false
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5559,
+      "type": "c1aa79cc-b9f8-43f1-8d88-3d7406881f62",
+      "pos": [
+        2898.1433506846097,
+        4431.1426195048725
+      ],
+      "size": [
+        287.5,
+        159.5
+      ],
+      "flags": {},
+      "order": 31,
+      "mode": 4,
+      "inputs": [
+        {
+          "label": "Before Video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": 14903
+        },
+        {
+          "label": "After Video",
+          "name": "video_1",
+          "type": "VIDEO",
+          "link": 14902
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [
+            14778
+          ]
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [
+          [
+            "79",
+            "direction"
+          ],
+          [
+            "79",
+            "match_image_size"
+          ],
+          [
+            "79",
+            "spacing_width"
+          ],
+          [
+            "79",
+            "spacing_color"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.13.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5562,
+      "type": "easy cleanGpuUsed",
+      "pos": [
+        2606.835589078068,
+        4393.748696788327
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {},
+      "order": 32,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "anything",
+          "type": "*",
+          "link": 14711
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "*",
+          "links": [
+            14903
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-easy-use",
+        "ver": "1.3.6",
+        "Node name for S&R": "easy cleanGpuUsed"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5563,
+      "type": "DualCLIPLoader",
+      "pos": [
+        1937.7445025108225,
+        4331.113964716672
+      ],
+      "size": [
+        518.75,
+        203.4375
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "showAdvanced": true,
+      "inputs": [
+        {
+          "label": "clip_name1",
+          "name": "clip_name1",
+          "type": "COMBO",
+          "widget": {
+            "name": "clip_name1"
+          },
+          "link": null
+        },
+        {
+          "label": "clip_name2",
+          "name": "clip_name2",
+          "type": "COMBO",
+          "widget": {
+            "name": "clip_name2"
+          },
+          "link": null
+        },
+        {
+          "label": "type",
+          "name": "type",
+          "type": "COMBO",
+          "widget": {
+            "name": "type"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "shape": 7,
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "CLIP",
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            14650,
+            14651
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "DualCLIPLoader",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "clip_name1": true,
+          "clip_name2": true,
+          "type": true,
+          "device": true
+        }
+      },
+      "widgets_values": [
+        "gemma_3_12B_it_fp4_mixed.safetensors",
+        "ltx-2.3_text_projection_bf16.safetensors",
+        "ltxv",
+        "default"
+      ]
+    },
+    {
+      "id": 5564,
+      "type": "VAELoaderKJ",
+      "pos": [
+        1940.4294038177754,
+        4009.9917036487186
+      ],
+      "size": [
+        518.75,
+        167.453125
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "vae_name",
+          "name": "vae_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "vae_name"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "VAE",
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            14819
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "VAELoaderKJ",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "device": true,
+          "vae_name": true
+        }
+      },
+      "widgets_values": [
+        "LTX23_audio_vae_bf16.safetensors",
+        "main_device",
+        "bf16"
+      ]
+    },
+    {
+      "id": 5565,
+      "type": "VAELoaderKJ",
+      "pos": [
+        1940.427248667633,
+        3751.492211578288
+      ],
+      "size": [
+        518.75,
+        167.453125
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "vae_name",
+          "name": "vae_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "vae_name"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "VAE",
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            14668,
+            14680,
+            14893,
+            14949
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "VAELoaderKJ",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "device": true,
+          "vae_name": true
+        }
+      },
+      "widgets_values": [
+        "LTX23_video_vae_bf16.safetensors",
+        "main_device",
+        "bf16"
+      ]
+    },
+    {
+      "id": 5566,
+      "type": "UNETLoader",
+      "pos": [
+        1952.67417621041,
+        3240.7162320271464
+      ],
+      "size": [
+        521.875,
+        107.5
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "unet_name",
+          "name": "unet_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "unet_name"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "MODEL",
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            14771,
+            14779,
+            14781
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "UNETLoader",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "unet_name": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors",
+        "default"
+      ]
+    },
+    {
+      "id": 5567,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        1944.1849874543964,
+        3448.505050167105
+      ],
+      "size": [
+        521.875,
+        160.625
+      ],
+      "flags": {},
+      "order": 17,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "link": 14781
+        },
+        {
+          "label": "lora_name",
+          "name": "lora_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "lora_name"
+          },
+          "link": null
+        },
+        {
+          "label": "strength_model",
+          "name": "strength_model",
+          "type": "FLOAT",
+          "widget": {
+            "name": "strength_model"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            14655
+          ]
+        },
+        {
+          "label": "latent_downscale_factor",
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "531512f7286963dc7aff1fd8bf5556e95eae03af",
+        "Node name for S&R": "LTXICLoRALoaderModelOnly",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "lora_name": true,
+          "strength_model": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors",
+        0.75
+      ]
+    },
+    {
+      "id": 5568,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        2636.546876800274,
+        3441.737089318999
+      ],
+      "size": [
+        518.75,
+        159.484375
+      ],
+      "flags": {},
+      "order": 19,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "link": 14780
+        },
+        {
+          "label": "lora_name",
+          "name": "lora_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "lora_name"
+          },
+          "link": null
+        },
+        {
+          "label": "strength_model",
+          "name": "strength_model",
+          "type": "FLOAT",
+          "widget": {
+            "name": "strength_model"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            14728
+          ]
+        },
+        {
+          "label": "latent_downscale_factor",
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "531512f7286963dc7aff1fd8bf5556e95eae03af",
+        "Node name for S&R": "LTXICLoRALoaderModelOnly",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "lora_name": true,
+          "strength_model": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-decompression-0.9.safetensors",
+        0.5
+      ]
+    },
+    {
+      "id": 5574,
+      "type": "RandomNoise",
+      "pos": [
+        3306.439297944243,
+        3982.189277301762
+      ],
+      "size": [
+        278.125,
+        81.984375
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [
+            14813,
+            14955
+          ]
+        }
+      ],
+      "title": "Stage 2 Noise · Shared Seed",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        42,
+        "fixed"
+      ],
+      "color": "#222",
+      "bgcolor": "#000"
+    },
+    {
+      "id": 5575,
+      "type": "CFGGuider",
+      "pos": [
+        4543.70977613057,
+        3982.15170888756
+      ],
+      "size": [
+        309.375,
+        127.515625
+      ],
+      "flags": {},
+      "order": 33,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14728
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14822
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14823
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            14956
+          ]
+        }
+      ],
+      "title": "Stage 2 Guider · Distilled Model Only",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 5576,
+      "type": "KSamplerSelect",
+      "pos": [
+        4107.2231052936,
+        3939.836864075704
+      ],
+      "size": [
+        343.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            14957
+          ]
+        }
+      ],
+      "title": "Stage 2 Sampler · Euler CFG++",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler_ancestral_cfg_pp"
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9042,
+      "type": "LatentUpscaleModelLoader",
+      "pos": [
+        4100.7247467984125,
+        3778.494191029667
+      ],
+      "size": [
+        365.921875,
+        79.515625
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "LATENT_UPSCALE_MODEL",
+          "type": "LATENT_UPSCALE_MODEL",
+          "links": [
+            14960
+          ]
+        }
+      ],
+      "title": "Native LTX Upscale Model · 1.5× / 2×",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LatentUpscaleModelLoader"
+      },
+      "widgets_values": [
+        "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9045,
+      "type": "BasicScheduler",
+      "pos": [
+        4538.282636818714,
+        3763.490495949545
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {},
+      "order": 15,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14771
+        }
+      ],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            14958
+          ]
+        }
+      ],
+      "title": "Stage 2 BasicScheduler",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "BasicScheduler"
+      },
+      "widgets_values": [
+        "simple",
+        3,
+        0.2
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9047,
+      "type": "LTXReframeRectilinearCorrection",
+      "pos": [
+        4476.417958946322,
+        4263.513485857038
+      ],
+      "size": [
+        331.25,
+        347.359375
+      ],
+      "flags": {},
+      "order": 34,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14962
+        },
+        {
+          "name": "source_mask",
+          "shape": 7,
+          "type": "MASK",
+          "link": 14777
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            14953
+          ]
+        },
+        {
+          "name": "correction_mask",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXReframeRectilinearCorrection"
+      },
+      "widgets_values": [
+        true,
+        0.01,
+        0.5,
+        0.5,
+        1,
+        192,
+        "bicubic",
+        8
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9051,
+      "type": "Note",
+      "pos": [
+        1844.1145463193513,
+        2850.3147849607235
+      ],
+      "size": [
+        343.84375,
+        165.015625
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "Iimit_frames toggle",
+      "properties": {},
+      "widgets_values": [
+        "- Turn ON when you clip/set custom frame limits.\n\nTurn OFF to use the full source video."
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9052,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        2636.2309483206473,
+        3209.809099597265
+      ],
+      "size": [
+        518.75,
+        159.484375
+      ],
+      "flags": {},
+      "order": 16,
+      "mode": 4,
+      "inputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "link": 14779
+        },
+        {
+          "label": "lora_name",
+          "name": "lora_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "lora_name"
+          },
+          "link": null
+        },
+        {
+          "label": "strength_model",
+          "name": "strength_model",
+          "type": "FLOAT",
+          "widget": {
+            "name": "strength_model"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            14780
+          ]
+        },
+        {
+          "label": "latent_downscale_factor",
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "531512f7286963dc7aff1fd8bf5556e95eae03af",
+        "Node name for S&R": "LTXICLoRALoaderModelOnly",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "lora_name": true,
+          "strength_model": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-deblur-0.9.safetensors",
+        0.5
+      ]
+    },
+    {
+      "id": 9055,
+      "type": "defb337f-bdbe-4ed9-8b41-8e9269dbd16f",
+      "pos": [
+        3347.431385778906,
+        4191.958828880091
+      ],
+      "size": [
+        225,
+        347.5
+      ],
+      "flags": {},
+      "order": 35,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14808
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14809
+        },
+        {
+          "name": "frame_rate",
+          "type": "FLOAT",
+          "widget": {
+            "name": "frame_rate"
+          },
+          "link": 14810
+        },
+        {
+          "name": "positive_1",
+          "type": "CONDITIONING",
+          "link": 14811
+        },
+        {
+          "name": "negative_1",
+          "type": "CONDITIONING",
+          "link": 14812
+        },
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 14813
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 14814
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 14815
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 14816
+        },
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "link": 14817
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14818
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14819
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            14820
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            14821
+          ]
+        },
+        {
+          "name": "positive_1",
+          "type": "CONDITIONING",
+          "links": [
+            14822
+          ]
+        },
+        {
+          "name": "negative_1",
+          "type": "CONDITIONING",
+          "links": [
+            14823
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14892,
+            14959
+          ]
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "links": [
+            14954
+          ]
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [
+          [
+            "5521",
+            "frame_rate"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 9075,
+      "type": "2eabb368-3326-4160-9984-dd85c7e7e498",
+      "pos": [
+        3339.2961997270263,
+        3243.4742597750196
+      ],
+      "size": [
+        225,
+        255.5
+      ],
+      "flags": {},
+      "order": 36,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14986
+        },
+        {
+          "name": "input_1",
+          "type": "IMAGE,MASK",
+          "link": 14987
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14892
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14893
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14894
+          ]
+        },
+        {
+          "name": "batch_size",
+          "type": "INT",
+          "links": [
+            14983
+          ]
+        },
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14896
+          ]
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14897,
+            14898
+          ]
+        },
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14899
+          ]
+        },
+        {
+          "name": "BOOLEAN",
+          "type": "BOOLEAN",
+          "links": [
+            14900
+          ]
+        },
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14901
+          ]
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [
+          [
+            "9072",
+            "image"
+          ],
+          [
+            "9072",
+            "upload"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 9077,
+      "type": "MelBandRoFormerSampler",
+      "pos": [
+        1780,
+        4950
+      ],
+      "size": [
+        246.78125,
+        95.515625
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 20,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MELROFORMERMODEL",
+          "link": 14905
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14915
+        }
+      ],
+      "outputs": [
+        {
+          "name": "vocals",
+          "type": "AUDIO",
+          "links": [
+            14917
+          ]
+        },
+        {
+          "name": "instruments",
+          "type": "AUDIO",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-melbandroformer",
+        "ver": "1.0.1",
+        "Node name for S&R": "MelBandRoFormerSampler"
+      },
+      "widgets_values": [],
+      "title": "Extract Custom Vocal Stem"
+    },
+    {
+      "id": 9079,
+      "type": "Audio Duration (mtb)",
+      "pos": [
+        700,
+        5210
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 21,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14991
+        }
+      ],
+      "outputs": [
+        {
+          "name": "duration_ms",
+          "type": "INT",
+          "links": [
+            14992
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-mtb",
+        "ver": "0.5.4",
+        "Node name for S&R": "Audio Duration (mtb)"
+      },
+      "widgets_values": [],
+      "title": "Selected Video Audio Duration"
+    },
+    {
+      "id": 9083,
+      "type": "MelBandRoFormerModelLoader",
+      "pos": [
+        1450,
+        4950
+      ],
+      "size": [
+        307.421875,
+        103.5
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "model",
+          "type": "MELROFORMERMODEL",
+          "links": [
+            14905
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-melbandroformer",
+        "ver": "1.0.1",
+        "Node name for S&R": "MelBandRoFormerModelLoader"
+      },
+      "widgets_values": [
+        "MelBandRoformer_fp16.safetensors"
+      ],
+      "color": "#323",
+      "bgcolor": "#535",
+      "title": "MelBandRoFormer Model · Custom Audio Only"
+    },
+    {
+      "id": 9084,
+      "type": "TrimAudioDuration",
+      "pos": [
+        1080,
+        4950
+      ],
+      "size": [
+        306,
+        111.5
+      ],
+      "flags": {},
+      "order": 18,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14912
+        },
+        {
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {
+            "name": "duration"
+          },
+          "link": 14993
+        }
+      ],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            14915
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.9.1",
+        "Node name for S&R": "TrimAudioDuration"
+      },
+      "widgets_values": [
+        0,
+        10
+      ],
+      "color": "#323",
+      "bgcolor": "#535",
+      "title": "Custom Track Window · Start Offset / Video Duration"
+    },
+    {
+      "id": 9086,
+      "type": "LTXReframeCustomAudioLoader",
+      "pos": [
+        700,
+        4950
+      ],
+      "size": [
+        330,
+        150
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 12,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            14912
+          ]
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXReframeCustomAudioLoader"
+      },
+      "widgets_values": [
+        ""
+      ],
+      "color": "#323",
+      "bgcolor": "#535",
+      "title": "Custom Audio Upload · Lazy-Safe"
+    },
+    {
+      "id": 9088,
+      "type": "PreviewAudio",
+      "pos": [
+        3150,
+        4950
+      ],
+      "size": [
+        300,
+        70
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 22,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14918
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.9.1",
+        "Node name for S&R": "PreviewAudio"
+      },
+      "widgets_values": [],
+      "color": "#323",
+      "bgcolor": "#535",
+      "title": "Preview Selected Audio"
+    },
+    {
+      "id": 9100,
+      "type": "8c3490c3-2129-435f-b9ef-9c340f8a2a25",
+      "pos": [
+        4124.96850299175,
+        4250.602168693877
+      ],
+      "size": [
+        225,
+        363.515625
+      ],
+      "flags": {},
+      "order": 37,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14948
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14949
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "link": 14950
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "link": 14951
+        },
+        {
+          "name": "on_false",
+          "type": "*",
+          "link": 14989
+        },
+        {
+          "name": "on_true",
+          "type": "*",
+          "link": 14953
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14954
+        },
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 14955
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 14956
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 14957
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 14958
+        },
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14959
+        },
+        {
+          "name": "upscale_model",
+          "type": "LATENT_UPSCALE_MODEL",
+          "link": 14960
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14984
+          ]
+        },
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            14962
+          ]
+        },
+        {
+          "name": "*",
+          "type": "*",
+          "links": [
+            14963
+          ]
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [],
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 9107,
+      "type": "d100f318-044f-47e1-aae2-b2ea5e0e88df",
+      "pos": [
+        4152.048606361803,
+        3324.986559802312
+      ],
+      "size": [
+        225,
+        203.5
+      ],
+      "flags": {},
+      "order": 38,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14980
+        },
+        {
+          "name": "input_1",
+          "type": "IMAGE,MASK",
+          "link": 14981
+        },
+        {
+          "name": "input_2",
+          "type": "IMAGE,MASK",
+          "link": 14982
+        },
+        {
+          "name": "size",
+          "type": "INT",
+          "widget": {
+            "name": "size"
+          },
+          "link": 14983
+        },
+        {
+          "name": "image_from",
+          "type": "IMAGE",
+          "link": 14984
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14985
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14986
+          ]
+        },
+        {
+          "name": "resized_1",
+          "type": "MASK",
+          "links": [
+            14987
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14988
+          ]
+        },
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14989
+          ]
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [
+          [
+            "9105",
+            "size"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 9108,
+      "type": "ComfyMathExpression",
+      "pos": [
+        1040,
+        5210
+      ],
+      "size": [
+        340,
+        150
+      ],
+      "flags": {},
+      "order": 39,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "a",
+          "localized_name": "values.a",
+          "name": "values.a",
+          "type": "FLOAT,INT,BOOLEAN",
+          "link": 14992
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "FLOAT",
+          "name": "FLOAT",
+          "type": "FLOAT",
+          "links": [
+            14993,
+            14994,
+            14997
+          ]
+        },
+        {
+          "localized_name": "INT",
+          "name": "INT",
+          "type": "INT",
+          "links": null
+        },
+        {
+          "localized_name": "BOOL",
+          "name": "BOOL",
+          "type": "BOOLEAN",
+          "links": null
+        }
+      ],
+      "title": "Video Duration · Milliseconds to Seconds",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "ComfyMathExpression"
+      },
+      "widgets_values": [
+        "a / 1000.0"
+      ],
+      "color": "#323",
+      "bgcolor": "#535"
+    },
+    {
+      "id": 9109,
+      "type": "EmptyAudio",
+      "pos": [
+        1450,
+        5200
+      ],
+      "size": [
+        300,
+        135
+      ],
+      "flags": {},
+      "order": 40,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {
+            "name": "duration"
+          },
+          "link": 14994
+        }
+      ],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            14995
+          ]
+        }
+      ],
+      "title": "Silence Padding · 44.1 kHz Stereo",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "EmptyAudio"
+      },
+      "widgets_values": [
+        10,
+        44100,
+        2
+      ],
+      "color": "#323",
+      "bgcolor": "#535"
+    },
+    {
+      "id": 9110,
+      "type": "AudioConcat",
+      "pos": [
+        2100,
+        4950
+      ],
+      "size": [
+        280,
+        120
+      ],
+      "flags": {},
+      "order": 41,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio1",
+          "type": "AUDIO",
+          "link": 14917
+        },
+        {
+          "name": "audio2",
+          "type": "AUDIO",
+          "link": 14995
+        }
+      ],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            14996
+          ]
+        }
+      ],
+      "title": "Vocals + Silence Padding",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "AudioConcat"
+      },
+      "widgets_values": [
+        "after"
+      ],
+      "color": "#323",
+      "bgcolor": "#535"
+    },
+    {
+      "id": 9111,
+      "type": "TrimAudioDuration",
+      "pos": [
+        2410,
+        4950
+      ],
+      "size": [
+        300,
+        120
+      ],
+      "flags": {},
+      "order": 42,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14996
+        },
+        {
+          "name": "duration",
+          "type": "FLOAT",
+          "widget": {
+            "name": "duration"
+          },
+          "link": 14997
+        }
+      ],
+      "outputs": [
+        {
+          "name": "AUDIO",
+          "type": "AUDIO",
+          "links": [
+            14998
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.9.1",
+        "Node name for S&R": "TrimAudioDuration"
+      },
+      "widgets_values": [
+        0,
+        10
+      ],
+      "color": "#323",
+      "bgcolor": "#535",
+      "title": "Exact Video Duration · Trim Padded Vocals"
+    },
+    {
+      "id": 9112,
+      "type": "LazySwitchKJ",
+      "pos": [
+        2760,
+        4950
+      ],
+      "size": [
+        340,
+        145
+      ],
+      "flags": {},
+      "order": 43,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "on_false",
+          "name": "on_false",
+          "type": "*",
+          "link": 14990
+        },
+        {
+          "localized_name": "on_true",
+          "name": "on_true",
+          "type": "*",
+          "link": 14998
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "*",
+          "name": "*",
+          "type": "*",
+          "links": [
+            14666,
+            14818,
+            14918
+          ]
+        }
+      ],
+      "title": "Use Custom Vocal Audio? · Off = Source Video",
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "LazySwitchKJ"
+      },
+      "widgets_values": [
+        false
+      ],
+      "color": "#263423",
+      "bgcolor": "#395034"
+    },
+    {
+      "id": 9113,
+      "type": "Note",
+      "pos": [
+        2100,
+        5200
+      ],
+      "size": [
+        1350,
+        170
+      ],
+      "flags": {},
+      "order": 44,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "Custom audio controls",
+      "properties": {},
+      "widgets_values": [
+        "OFF uses the source video's audio (or duration-matched silence).\n\nON loads the uploaded track, starts at Custom Track Window's start_index, isolates vocals with MelBandRoFormer, and trims or silence-pads them to the effective video range. The same selected audio drives LTX and the final export."
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    }
+  ],
+  "links": [
+    [
+      14650,
+      5563,
+      0,
+      5527,
+      0,
+      "CLIP"
+    ],
+    [
+      14651,
+      5563,
+      0,
+      5528,
+      0,
+      "CLIP"
+    ],
+    [
+      14655,
+      5567,
+      0,
+      5534,
+      0,
+      "MODEL"
+    ],
+    [
+      14656,
+      5544,
+      0,
+      5534,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14657,
+      5544,
+      1,
+      5534,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      14666,
+      9112,
+      0,
+      5539,
+      2,
+      "AUDIO"
+    ],
+    [
+      14668,
+      5565,
+      0,
+      5541,
+      0,
+      "VAE"
+    ],
+    [
+      14669,
+      5542,
+      0,
+      5541,
+      1,
+      "IMAGE"
+    ],
+    [
+      14680,
+      5565,
+      0,
+      5544,
+      2,
+      "VAE"
+    ],
+    [
+      14681,
+      5541,
+      0,
+      5544,
+      3,
+      "LATENT"
+    ],
+    [
+      14711,
+      5539,
+      0,
+      5562,
+      0,
+      "VIDEO"
+    ],
+    [
+      14728,
+      5568,
+      0,
+      5575,
+      0,
+      "MODEL"
+    ],
+    [
+      14771,
+      5566,
+      0,
+      9045,
+      0,
+      "MODEL"
+    ],
+    [
+      14777,
+      5553,
+      2,
+      9047,
+      1,
+      "MASK"
+    ],
+    [
+      14778,
+      5559,
+      0,
+      5552,
+      0,
+      "VIDEO"
+    ],
+    [
+      14779,
+      5566,
+      0,
+      9052,
+      0,
+      "MODEL"
+    ],
+    [
+      14780,
+      9052,
+      0,
+      5568,
+      0,
+      "MODEL"
+    ],
+    [
+      14781,
+      5566,
+      0,
+      5567,
+      0,
+      "MODEL"
+    ],
+    [
+      14808,
+      5527,
+      0,
+      9055,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14809,
+      5528,
+      0,
+      9055,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14810,
+      5553,
+      5,
+      9055,
+      2,
+      "FLOAT"
+    ],
+    [
+      14811,
+      5544,
+      0,
+      9055,
+      3,
+      "CONDITIONING"
+    ],
+    [
+      14812,
+      5544,
+      1,
+      9055,
+      4,
+      "CONDITIONING"
+    ],
+    [
+      14813,
+      5574,
+      0,
+      9055,
+      5,
+      "NOISE"
+    ],
+    [
+      14814,
+      5534,
+      0,
+      9055,
+      6,
+      "GUIDER"
+    ],
+    [
+      14815,
+      5538,
+      0,
+      9055,
+      7,
+      "SAMPLER"
+    ],
+    [
+      14816,
+      5533,
+      0,
+      9055,
+      8,
+      "SIGMAS"
+    ],
+    [
+      14817,
+      5544,
+      2,
+      9055,
+      9,
+      "LATENT"
+    ],
+    [
+      14818,
+      9112,
+      0,
+      9055,
+      10,
+      "AUDIO"
+    ],
+    [
+      14819,
+      5564,
+      0,
+      9055,
+      11,
+      "VAE"
+    ],
+    [
+      14820,
+      9055,
+      0,
+      5544,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14821,
+      9055,
+      1,
+      5544,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14822,
+      9055,
+      2,
+      5575,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14823,
+      9055,
+      3,
+      5575,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      14892,
+      9055,
+      4,
+      9075,
+      2,
+      "LATENT"
+    ],
+    [
+      14893,
+      5565,
+      0,
+      9075,
+      3,
+      "VAE"
+    ],
+    [
+      14894,
+      9075,
+      0,
+      5537,
+      2,
+      "MASK"
+    ],
+    [
+      14896,
+      9075,
+      2,
+      5541,
+      2,
+      "LATENT"
+    ],
+    [
+      14897,
+      9075,
+      3,
+      5537,
+      1,
+      "IMAGE"
+    ],
+    [
+      14898,
+      9075,
+      3,
+      5544,
+      4,
+      "IMAGE"
+    ],
+    [
+      14899,
+      9075,
+      4,
+      5537,
+      0,
+      "IMAGE"
+    ],
+    [
+      14900,
+      9075,
+      5,
+      5541,
+      3,
+      "BOOLEAN"
+    ],
+    [
+      14901,
+      9075,
+      6,
+      5542,
+      0,
+      "IMAGE"
+    ],
+    [
+      14902,
+      5553,
+      6,
+      5559,
+      1,
+      "VIDEO"
+    ],
+    [
+      14903,
+      5562,
+      0,
+      5559,
+      0,
+      "VIDEO"
+    ],
+    [
+      14905,
+      9083,
+      0,
+      9077,
+      0,
+      "MELROFORMERMODEL"
+    ],
+    [
+      14912,
+      9086,
+      0,
+      9084,
+      0,
+      "AUDIO"
+    ],
+    [
+      14915,
+      9084,
+      0,
+      9077,
+      1,
+      "AUDIO"
+    ],
+    [
+      14917,
+      9077,
+      0,
+      9110,
+      0,
+      "AUDIO"
+    ],
+    [
+      14918,
+      9112,
+      0,
+      9088,
+      0,
+      "AUDIO"
+    ],
+    [
+      14948,
+      5537,
+      0,
+      9100,
+      0,
+      "IMAGE"
+    ],
+    [
+      14949,
+      5565,
+      0,
+      9100,
+      1,
+      "VAE"
+    ],
+    [
+      14950,
+      5553,
+      7,
+      9100,
+      2,
+      "INT"
+    ],
+    [
+      14951,
+      5553,
+      8,
+      9100,
+      3,
+      "INT"
+    ],
+    [
+      14953,
+      9047,
+      0,
+      9100,
+      5,
+      "IMAGE"
+    ],
+    [
+      14954,
+      9055,
+      5,
+      9100,
+      6,
+      "LATENT"
+    ],
+    [
+      14955,
+      5574,
+      0,
+      9100,
+      7,
+      "NOISE"
+    ],
+    [
+      14956,
+      5575,
+      0,
+      9100,
+      8,
+      "GUIDER"
+    ],
+    [
+      14957,
+      5576,
+      0,
+      9100,
+      9,
+      "SAMPLER"
+    ],
+    [
+      14958,
+      9045,
+      0,
+      9100,
+      10,
+      "SIGMAS"
+    ],
+    [
+      14959,
+      9055,
+      4,
+      9100,
+      11,
+      "LATENT"
+    ],
+    [
+      14960,
+      9042,
+      0,
+      9100,
+      12,
+      "LATENT_UPSCALE_MODEL"
+    ],
+    [
+      14962,
+      9100,
+      1,
+      9047,
+      0,
+      "IMAGE"
+    ],
+    [
+      14963,
+      9100,
+      2,
+      5539,
+      0,
+      "IMAGE"
+    ],
+    [
+      14980,
+      5553,
+      0,
+      9107,
+      0,
+      "IMAGE"
+    ],
+    [
+      14981,
+      5553,
+      1,
+      9107,
+      1,
+      "MASK"
+    ],
+    [
+      14982,
+      5553,
+      2,
+      9107,
+      2,
+      "MASK"
+    ],
+    [
+      14983,
+      9075,
+      1,
+      9107,
+      3,
+      "INT"
+    ],
+    [
+      14984,
+      9100,
+      0,
+      9107,
+      4,
+      "IMAGE"
+    ],
+    [
+      14985,
+      5557,
+      0,
+      9107,
+      5,
+      "MASK"
+    ],
+    [
+      14986,
+      9107,
+      0,
+      9075,
+      0,
+      "IMAGE"
+    ],
+    [
+      14987,
+      9107,
+      1,
+      9075,
+      1,
+      "MASK"
+    ],
+    [
+      14988,
+      9107,
+      2,
+      5557,
+      0,
+      "MASK"
+    ],
+    [
+      14989,
+      9107,
+      3,
+      9100,
+      4,
+      "IMAGE"
+    ],
+    [
+      14990,
+      5553,
+      4,
+      9112,
+      0,
+      "AUDIO"
+    ],
+    [
+      14991,
+      5553,
+      4,
+      9079,
+      0,
+      "AUDIO"
+    ],
+    [
+      14992,
+      9079,
+      0,
+      9108,
+      0,
+      "INT"
+    ],
+    [
+      14993,
+      9108,
+      0,
+      9084,
+      1,
+      "FLOAT"
+    ],
+    [
+      14994,
+      9108,
+      0,
+      9109,
+      0,
+      "FLOAT"
+    ],
+    [
+      14995,
+      9109,
+      0,
+      9110,
+      1,
+      "AUDIO"
+    ],
+    [
+      14996,
+      9110,
+      0,
+      9111,
+      0,
+      "AUDIO"
+    ],
+    [
+      14997,
+      9108,
+      0,
+      9111,
+      1,
+      "FLOAT"
+    ],
+    [
+      14998,
+      9111,
+      0,
+      9112,
+      1,
+      "AUDIO"
+    ],
+    [
+      14999,
+      5553,
+      5,
+      5539,
+      1,
+      "FLOAT"
+    ]
+  ],
+  "groups": [
+    {
+      "id": 201,
+      "title": "Group",
+      "bounding": [
+        1842.031559935187,
+        3057.8037686804973,
+        3074.581739716744,
+        1646.4847221201553
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 202,
+      "title": "Load Models",
+      "bounding": [
+        1881.3987882980887,
+        3125.8632949328417,
+        649.9095453339949,
+        1527.8976283992993
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 206,
+      "title": "1. Prepare Two-Stage Outpaint Conditions",
+      "bounding": [
+        3261.108318887105,
+        3129.14706001421,
+        787.1075732081713,
+        1528.0708281365696
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 203,
+      "title": "These settings effect strength of source video",
+      "bounding": [
+        3279.525230864959,
+        3525.90397829568,
+        758.0756507166602,
+        1118.6070301690256
+      ],
+      "color": "#8A8",
+      "flags": {}
+    },
+    {
+      "id": 208,
+      "title": "1. Reframe Source Video",
+      "bounding": [
+        2557.777115692966,
+        3660.7219572530757,
+        675.5252377975253,
+        613.2277736458445
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 205,
+      "title": "Generate Low Resolution Stage 1 video",
+      "bounding": [
+        3313.621655886097,
+        4083.8181989970576,
+        688.9389712203624,
+        535.6137160379835
+      ],
+      "color": "#8AA",
+      "flags": {}
+    },
+    {
+      "id": 207,
+      "title": "Decode and Blend (STAGE 1)",
+      "bounding": [
+        3641.5550484144087,
+        3187.034241743885,
+        358.54492028553705,
+        300.380738936552
+      ],
+      "color": "#8A8",
+      "flags": {}
+    },
+    {
+      "id": 211,
+      "title": "2. Stage 1 Direct Finish + Source Restoration",
+      "bounding": [
+        4070.2049616395702,
+        3131.9797746856343,
+        818.1489809712725,
+        499.78751423245876
+      ],
+      "color": "#4f8a5b",
+      "flags": {}
+    },
+    {
+      "id": 209,
+      "title": "Stage 2 Full-Resolution Mask",
+      "bounding": [
+        4097.197022022294,
+        3240.9841101536413,
+        309.93384815295394,
+        307.8821232758537
+      ],
+      "color": "#b06634",
+      "flags": {}
+    },
+    {
+      "id": 212,
+      "title": "3. Stable Stage 2 · Optional Native 1.5× / 2× Upscale + 4-Step Full-Clip Synthesis",
+      "bounding": [
+        4074.253339906803,
+        3652.357833798646,
+        815.7816776935447,
+        1006.4457576785549
+      ],
+      "color": "#b06634",
+      "flags": {}
+    },
+    {
+      "id": 213,
+      "title": "REPAIR FISH-EYE LENS/MOTION DISTORTION",
+      "bounding": [
+        4403.079071236254,
+        4167.731050677765,
+        465.9290251293796,
+        468.3302283839621
+      ],
+      "color": "#8A8",
+      "flags": {}
+    },
+    {
+      "id": 214,
+      "title": "Final Output Settings",
+      "bounding": [
+        2562.436348054896,
+        4297.78389609365,
+        670.2073778343365,
+        358.211417797731
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 215,
+      "title": "USE THIS NODE TO BLEND SEAM",
+      "bounding": [
+        4453.592477418698,
+        3179.982201162255,
+        396.81195871477485,
+        434.5649344230974
+      ],
+      "color": "#8A8",
+      "flags": {}
+    },
+    {
+      "id": 216,
+      "title": "Additional LoRAs",
+      "bounding": [
+        2557.5990441033214,
+        3123.9490169315186,
+        672.0062505482947,
+        509.69769307314164
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 210,
+      "title": "Stage 1 Half-Resolution Mask",
+      "bounding": [
+        3299.8017937223126,
+        3172.71488442258,
+        303.1237577437173,
+        328.73693124208467
+      ],
+      "color": "#8AA",
+      "flags": {}
+    },
+    {
+      "id": 217,
+      "title": "BYPASS COMPARISON",
+      "bounding": [
+        2888.6039612477803,
+        4365.97243246716,
+        310.61709551998683,
+        243.50798583668256
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 218,
+      "title": "Optional Custom Audio · OFF = Source Video",
+      "bounding": [
+        650,
+        4830,
+        2860,
+        620
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    }
+  ],
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "c1aa79cc-b9f8-43f1-8d88-3d7406881f62",
+        "version": 1,
+        "state": {
+          "lastGroupId": 217,
+          "lastNodeId": 9107,
+          "lastLinkId": 14989,
+          "lastRerouteId": 228
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Video Stitch",
+        "description": "Stitches multiple video clips into a single sequential video file.",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            -6810,
+            2580,
+            143.55859375,
+            160
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            -4770,
+            2600,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "85555afe-c7a1-4f6e-b073-7c37f7bace7f",
+            "name": "video",
+            "type": "VIDEO",
+            "linkIds": [
+              253
+            ],
+            "localized_name": "video",
+            "label": "Before Video",
+            "pos": [
+              -6686.44140625,
+              2600
+            ]
+          },
+          {
+            "id": "022773ee-6b4f-4e3d-bead-68b3e75e2d20",
+            "name": "video_1",
+            "type": "VIDEO",
+            "linkIds": [
+              254
+            ],
+            "localized_name": "video_1",
+            "label": "After Video",
+            "pos": [
+              -6686.44140625,
+              2620
+            ]
+          },
+          {
+            "id": "7bcd7cbc-e918-472a-a0cf-2e0900545372",
+            "name": "direction",
+            "type": "COMBO",
+            "linkIds": [
+              259
+            ],
+            "pos": [
+              -6686.44140625,
+              2640
+            ]
+          },
+          {
+            "id": "9a00389d-c1c8-40d5-87fe-f41019b61fbc",
+            "name": "match_image_size",
+            "type": "BOOLEAN",
+            "linkIds": [
+              260
+            ],
+            "pos": [
+              -6686.44140625,
+              2660
+            ]
+          },
+          {
+            "id": "b95e0440-3ea8-4ae0-887e-12e75701042a",
+            "name": "spacing_width",
+            "type": "INT",
+            "linkIds": [
+              261
+            ],
+            "pos": [
+              -6686.44140625,
+              2680
+            ]
+          },
+          {
+            "id": "83ab9382-0a70-4169-b26a-66ab026b43c4",
+            "name": "spacing_color",
+            "type": "COMBO",
+            "linkIds": [
+              262
+            ],
+            "pos": [
+              -6686.44140625,
+              2700
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "09707f43-7552-4a6e-bd23-d962d31801c2",
+            "name": "VIDEO",
+            "type": "VIDEO",
+            "linkIds": [
+              255
+            ],
+            "localized_name": "VIDEO",
+            "pos": [
+              -4750,
+              2620
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 78,
+            "type": "GetVideoComponents",
+            "pos": [
+              -6390,
+              2600
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 254
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  249
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": null
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "GetVideoComponents"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 77,
+            "type": "GetVideoComponents",
+            "pos": [
+              -6390,
+              2420
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 253
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  248
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": [
+                  251
+                ]
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": [
+                  252
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "GetVideoComponents"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 90,
+            "type": "GetImageSize",
+            "pos": [
+              -6390,
+              3030
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "link": 266
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "width",
+                "name": "width",
+                "type": "INT",
+                "links": [
+                  274
+                ]
+              },
+              {
+                "localized_name": "height",
+                "name": "height",
+                "type": "INT",
+                "links": [
+                  276
+                ]
+              },
+              {
+                "localized_name": "batch_size",
+                "name": "batch_size",
+                "type": "INT",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "GetImageSize"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 80,
+            "type": "CreateVideo",
+            "pos": [
+              -5190,
+              2420
+            ],
+            "size": [
+              270,
+              130
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 282
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": 251
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "fps"
+                },
+                "link": 252
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "VIDEO",
+                "name": "VIDEO",
+                "type": "VIDEO",
+                "links": [
+                  255
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "CreateVideo"
+            },
+            "widgets_values": [
+              30
+            ]
+          },
+          {
+            "id": 95,
+            "type": "ComfyMathExpression",
+            "pos": [
+              -6040,
+              3020
+            ],
+            "size": [
+              400,
+              200
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "a",
+                "localized_name": "values.a",
+                "name": "values.a",
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": 274
+              },
+              {
+                "label": "b",
+                "localized_name": "values.b",
+                "name": "values.b",
+                "shape": 7,
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": null
+              },
+              {
+                "localized_name": "INT",
+                "name": "INT",
+                "type": "INT",
+                "links": [
+                  279
+                ]
+              },
+              {
+                "localized_name": "BOOL",
+                "name": "BOOL",
+                "type": "BOOLEAN",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ComfyMathExpression"
+            },
+            "widgets_values": [
+              "a & ~1"
+            ]
+          },
+          {
+            "id": 96,
+            "type": "ComfyMathExpression",
+            "pos": [
+              -6040,
+              3290
+            ],
+            "size": [
+              400,
+              200
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "a",
+                "localized_name": "values.a",
+                "name": "values.a",
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": 276
+              },
+              {
+                "label": "b",
+                "localized_name": "values.b",
+                "name": "values.b",
+                "shape": 7,
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": null
+              },
+              {
+                "localized_name": "INT",
+                "name": "INT",
+                "type": "INT",
+                "links": [
+                  280
+                ]
+              },
+              {
+                "localized_name": "BOOL",
+                "name": "BOOL",
+                "type": "BOOLEAN",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ComfyMathExpression"
+            },
+            "widgets_values": [
+              "a & ~1"
+            ]
+          },
+          {
+            "id": 79,
+            "type": "ImageStitch",
+            "pos": [
+              -6390,
+              2780
+            ],
+            "size": [
+              270,
+              160
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image1",
+                "name": "image1",
+                "type": "IMAGE",
+                "link": 248
+              },
+              {
+                "localized_name": "image2",
+                "name": "image2",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": 249
+              },
+              {
+                "localized_name": "direction",
+                "name": "direction",
+                "type": "COMBO",
+                "widget": {
+                  "name": "direction"
+                },
+                "link": 259
+              },
+              {
+                "localized_name": "match_image_size",
+                "name": "match_image_size",
+                "type": "BOOLEAN",
+                "widget": {
+                  "name": "match_image_size"
+                },
+                "link": 260
+              },
+              {
+                "localized_name": "spacing_width",
+                "name": "spacing_width",
+                "type": "INT",
+                "widget": {
+                  "name": "spacing_width"
+                },
+                "link": 261
+              },
+              {
+                "localized_name": "spacing_color",
+                "name": "spacing_color",
+                "type": "COMBO",
+                "widget": {
+                  "name": "spacing_color"
+                },
+                "link": 262
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  266,
+                  281
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "ImageStitch"
+            },
+            "widgets_values": [
+              "right",
+              false,
+              0,
+              "white"
+            ]
+          },
+          {
+            "id": 97,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              -5560,
+              2790
+            ],
+            "size": [
+              270,
+              154
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 281
+              },
+              {
+                "localized_name": "width",
+                "name": "resize_type.width",
+                "type": "INT",
+                "widget": {
+                  "name": "resize_type.width"
+                },
+                "link": 279
+              },
+              {
+                "localized_name": "height",
+                "name": "resize_type.height",
+                "type": "INT",
+                "widget": {
+                  "name": "resize_type.height"
+                },
+                "link": 280
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE",
+                "links": [
+                  282
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale dimensions",
+              512,
+              512,
+              "center",
+              "area"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 248,
+            "origin_id": 77,
+            "origin_slot": 0,
+            "target_id": 79,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 249,
+            "origin_id": 78,
+            "origin_slot": 0,
+            "target_id": 79,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 251,
+            "origin_id": 77,
+            "origin_slot": 1,
+            "target_id": 80,
+            "target_slot": 1,
+            "type": "AUDIO"
+          },
+          {
+            "id": 252,
+            "origin_id": 77,
+            "origin_slot": 2,
+            "target_id": 80,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 253,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 77,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 254,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 78,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 255,
+            "origin_id": 80,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 259,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 79,
+            "target_slot": 2,
+            "type": "COMBO"
+          },
+          {
+            "id": 260,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 79,
+            "target_slot": 3,
+            "type": "BOOLEAN"
+          },
+          {
+            "id": 261,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 79,
+            "target_slot": 4,
+            "type": "INT"
+          },
+          {
+            "id": 262,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 79,
+            "target_slot": 5,
+            "type": "COMBO"
+          },
+          {
+            "id": 266,
+            "origin_id": 79,
+            "origin_slot": 0,
+            "target_id": 90,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 274,
+            "origin_id": 90,
+            "origin_slot": 0,
+            "target_id": 95,
+            "target_slot": 0,
+            "type": "INT"
+          },
+          {
+            "id": 276,
+            "origin_id": 90,
+            "origin_slot": 1,
+            "target_id": 96,
+            "target_slot": 0,
+            "type": "INT"
+          },
+          {
+            "id": 279,
+            "origin_id": 95,
+            "origin_slot": 1,
+            "target_id": 97,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 280,
+            "origin_id": 96,
+            "origin_slot": 1,
+            "target_id": 97,
+            "target_slot": 2,
+            "type": "INT"
+          },
+          {
+            "id": 281,
+            "origin_id": 79,
+            "origin_slot": 0,
+            "target_id": 97,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 282,
+            "origin_id": 97,
+            "origin_slot": 0,
+            "target_id": 80,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        }
+      },
+      {
+        "id": "defb337f-bdbe-4ed9-8b41-8e9269dbd16f",
+        "version": 1,
+        "state": {
+          "lastGroupId": 217,
+          "lastNodeId": 9107,
+          "lastLinkId": 14989,
+          "lastRerouteId": 228
+        },
+        "revision": 1,
+        "config": {},
+        "name": "New Subgraph",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            2717.2287792596712,
+            4431.633638619667,
+            128,
+            288
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            3552.177733217622,
+            4491.633638619667,
+            128,
+            168
+          ]
+        },
+        "inputs": [
+          {
+            "id": "003322be-3eca-4d07-b6af-3ccc257185a9",
+            "name": "positive",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14640
+            ],
+            "localized_name": "positive",
+            "pos": [
+              51,
+              24
+            ]
+          },
+          {
+            "id": "5bd0ef5b-cbce-4257-aed9-b005118687c5",
+            "name": "negative",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14641
+            ],
+            "localized_name": "negative",
+            "pos": [
+              51,
+              44
+            ]
+          },
+          {
+            "id": "c4f852bc-5899-48a4-8177-df7700c67e93",
+            "name": "frame_rate",
+            "type": "FLOAT",
+            "linkIds": [
+              14642
+            ],
+            "localized_name": "frame_rate",
+            "pos": [
+              51,
+              64
+            ]
+          },
+          {
+            "id": "f2aaa1b6-72f5-45fb-914e-f335af1903ed",
+            "name": "positive_1",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14659,
+              14693
+            ],
+            "localized_name": "positive_1",
+            "pos": [
+              51,
+              84
+            ]
+          },
+          {
+            "id": "2a4074ea-ae4f-4c64-984a-0c49a59dba66",
+            "name": "negative_1",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14660,
+              14694
+            ],
+            "localized_name": "negative_1",
+            "pos": [
+              51,
+              104
+            ]
+          },
+          {
+            "id": "563764b6-6913-4060-a0c9-0b82ae46c2f7",
+            "name": "noise",
+            "type": "NOISE",
+            "linkIds": [
+              14757
+            ],
+            "localized_name": "noise",
+            "pos": [
+              51,
+              124
+            ]
+          },
+          {
+            "id": "c897500a-7bf9-4ce0-83eb-65aaecc1a710",
+            "name": "guider",
+            "type": "GUIDER",
+            "linkIds": [
+              14674
+            ],
+            "localized_name": "guider",
+            "pos": [
+              51,
+              144
+            ]
+          },
+          {
+            "id": "092e3781-123d-46a3-b207-781374a16889",
+            "name": "sampler",
+            "type": "SAMPLER",
+            "linkIds": [
+              14675
+            ],
+            "localized_name": "sampler",
+            "pos": [
+              51,
+              164
+            ]
+          },
+          {
+            "id": "25ab9970-283b-452b-b1e7-8f33c5f711db",
+            "name": "sigmas",
+            "type": "SIGMAS",
+            "linkIds": [
+              14773
+            ],
+            "localized_name": "sigmas",
+            "pos": [
+              51,
+              184
+            ]
+          },
+          {
+            "id": "f1527a08-94d2-4895-8d37-38adbc241999",
+            "name": "video_latent",
+            "type": "LATENT",
+            "linkIds": [
+              14683
+            ],
+            "localized_name": "video_latent",
+            "pos": [
+              51,
+              204
+            ]
+          },
+          {
+            "id": "db94507f-0727-48fc-96ae-8279fb9f730b",
+            "name": "audio",
+            "type": "AUDIO",
+            "linkIds": [
+              14691
+            ],
+            "localized_name": "audio",
+            "pos": [
+              51,
+              224
+            ]
+          },
+          {
+            "id": "71a173c2-f028-4d3d-ad69-315d1d743764",
+            "name": "vae",
+            "type": "VAE",
+            "linkIds": [
+              14692
+            ],
+            "localized_name": "vae",
+            "pos": [
+              51,
+              244
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "583ec700-b4ae-45bf-a00a-ec5d85526b79",
+            "name": "positive",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14678
+            ],
+            "localized_name": "positive",
+            "pos": [
+              24,
+              24
+            ]
+          },
+          {
+            "id": "5157d8e6-03b9-4a60-9a1b-78ac6d2b7fa7",
+            "name": "negative",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14679
+            ],
+            "localized_name": "negative",
+            "pos": [
+              24,
+              44
+            ]
+          },
+          {
+            "id": "73d80047-f223-42ce-b907-55dc52525239",
+            "name": "positive_1",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14729
+            ],
+            "localized_name": "positive_1",
+            "pos": [
+              24,
+              64
+            ]
+          },
+          {
+            "id": "9b6bf980-25b9-4def-bcdc-581a2e1c84a1",
+            "name": "negative_1",
+            "type": "CONDITIONING",
+            "linkIds": [
+              14730
+            ],
+            "localized_name": "negative_1",
+            "pos": [
+              24,
+              84
+            ]
+          },
+          {
+            "id": "4a0c2686-2636-417c-bd23-090011707577",
+            "name": "latent",
+            "type": "LATENT",
+            "linkIds": [
+              14742,
+              14742
+            ],
+            "localized_name": "latent",
+            "pos": [
+              24,
+              104
+            ]
+          },
+          {
+            "id": "68fc5ab2-4916-48e2-803e-6f7137df46b0",
+            "name": "audio_latent",
+            "type": "LATENT",
+            "linkIds": [
+              14798
+            ],
+            "localized_name": "audio_latent",
+            "pos": [
+              24,
+              124
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 5521,
+            "type": "LTXVConditioning",
+            "pos": [
+              2910.5157413139486,
+              4528.584822489217
+            ],
+            "size": [
+              225,
+              103.515625
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "link": 14640
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "link": 14641
+              },
+              {
+                "localized_name": "frame_rate",
+                "name": "frame_rate",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "frame_rate"
+                },
+                "link": 14642
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "slot_index": 0,
+                "links": [
+                  14678
+                ]
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "slot_index": 1,
+                "links": [
+                  14679
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.3.28",
+              "Node name for S&R": "LTXVConditioning"
+            },
+            "widgets_values": [
+              24
+            ]
+          },
+          {
+            "id": 5536,
+            "type": "LTXVCropGuides",
+            "pos": [
+              2905.8603821211464,
+              4329.044405548306
+            ],
+            "size": [
+              225,
+              98.125
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "link": 14659
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "link": 14660
+              },
+              {
+                "localized_name": "latent",
+                "name": "latent",
+                "type": "LATENT",
+                "link": 14661
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "links": [
+                  14729
+                ]
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "links": [
+                  14730
+                ]
+              },
+              {
+                "localized_name": "latent",
+                "name": "latent",
+                "type": "LATENT",
+                "links": [
+                  14742
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.3.75",
+              "Node name for S&R": "LTXVCropGuides"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 5543,
+            "type": "SamplerCustomAdvanced",
+            "pos": [
+              3267.177733217622,
+              4320.921381132472
+            ],
+            "size": [
+              225,
+              143.53125
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "noise",
+                "name": "noise",
+                "type": "NOISE",
+                "link": 14757
+              },
+              {
+                "localized_name": "guider",
+                "name": "guider",
+                "type": "GUIDER",
+                "link": 14674
+              },
+              {
+                "localized_name": "sampler",
+                "name": "sampler",
+                "type": "SAMPLER",
+                "link": 14675
+              },
+              {
+                "localized_name": "sigmas",
+                "name": "sigmas",
+                "type": "SIGMAS",
+                "link": 14773
+              },
+              {
+                "localized_name": "latent_image",
+                "name": "latent_image",
+                "type": "LATENT",
+                "link": 14677
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "output",
+                "name": "output",
+                "type": "LATENT",
+                "links": []
+              },
+              {
+                "localized_name": "denoised_output",
+                "name": "denoised_output",
+                "type": "LATENT",
+                "links": [
+                  14688
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.14.1",
+              "Node name for S&R": "SamplerCustomAdvanced"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 5545,
+            "type": "LTXVConcatAVLatent",
+            "pos": [
+              3262.9521642726554,
+              4788.814646106863
+            ],
+            "size": [
+              225,
+              71.53125
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video_latent",
+                "name": "video_latent",
+                "type": "LATENT",
+                "link": 14683
+              },
+              {
+                "localized_name": "audio_latent",
+                "name": "audio_latent",
+                "type": "LATENT",
+                "link": 14684
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "latent",
+                "name": "latent",
+                "type": "LATENT",
+                "links": [
+                  14677
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.5.1",
+              "Node name for S&R": "LTXVConcatAVLatent"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 5547,
+            "type": "LTXVSeparateAVLatent",
+            "pos": [
+              3263.8457261466633,
+              4515.248628747466
+            ],
+            "size": [
+              225,
+              71.53125
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "av_latent",
+                "name": "av_latent",
+                "type": "LATENT",
+                "link": 14688
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "video_latent",
+                "name": "video_latent",
+                "type": "LATENT",
+                "links": [
+                  14661
+                ]
+              },
+              {
+                "localized_name": "audio_latent",
+                "name": "audio_latent",
+                "type": "LATENT",
+                "links": [
+                  14798
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.5.1",
+              "Node name for S&R": "LTXVSeparateAVLatent"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 5549,
+            "type": "VAEEncodeAudio",
+            "pos": [
+              3262.150694045136,
+              4636.482667961747
+            ],
+            "size": [
+              225,
+              98.125
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "link": 14691
+              },
+              {
+                "localized_name": "vae",
+                "name": "vae",
+                "type": "VAE",
+                "link": 14692
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "LATENT",
+                "name": "LATENT",
+                "type": "LATENT",
+                "links": [
+                  14695
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "VAEEncodeAudio"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 5550,
+            "type": "LTXVSetAudioRefTokens",
+            "pos": [
+              2905.2287792596712,
+              4735.695105609564
+            ],
+            "size": [
+              268.75,
+              119.515625
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "link": 14693
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "link": 14694
+              },
+              {
+                "localized_name": "audio_latent",
+                "name": "audio_latent",
+                "type": "LATENT",
+                "link": 14695
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "positive",
+                "name": "positive",
+                "type": "CONDITIONING",
+                "links": []
+              },
+              {
+                "localized_name": "negative",
+                "name": "negative",
+                "type": "CONDITIONING",
+                "links": []
+              },
+              {
+                "localized_name": "frozen_audio",
+                "name": "frozen_audio",
+                "type": "LATENT",
+                "links": [
+                  14684
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "ComfyUI-LTXVideo",
+              "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+              "Node name for S&R": "LTXVSetAudioRefTokens"
+            },
+            "widgets_values": []
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 14661,
+            "origin_id": 5547,
+            "origin_slot": 0,
+            "target_id": 5536,
+            "target_slot": 2,
+            "type": "LATENT"
+          },
+          {
+            "id": 14677,
+            "origin_id": 5545,
+            "origin_slot": 0,
+            "target_id": 5543,
+            "target_slot": 4,
+            "type": "LATENT"
+          },
+          {
+            "id": 14684,
+            "origin_id": 5550,
+            "origin_slot": 2,
+            "target_id": 5545,
+            "target_slot": 1,
+            "type": "LATENT"
+          },
+          {
+            "id": 14688,
+            "origin_id": 5543,
+            "origin_slot": 1,
+            "target_id": 5547,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14695,
+            "origin_id": 5549,
+            "origin_slot": 0,
+            "target_id": 5550,
+            "target_slot": 2,
+            "type": "LATENT"
+          },
+          {
+            "id": 14640,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 5521,
+            "target_slot": 0,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14641,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 5521,
+            "target_slot": 1,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14642,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 5521,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 14659,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 5536,
+            "target_slot": 0,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14693,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 5550,
+            "target_slot": 0,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14660,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 5536,
+            "target_slot": 1,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14694,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 5550,
+            "target_slot": 1,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14757,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 5543,
+            "target_slot": 0,
+            "type": "NOISE"
+          },
+          {
+            "id": 14674,
+            "origin_id": -10,
+            "origin_slot": 6,
+            "target_id": 5543,
+            "target_slot": 1,
+            "type": "GUIDER"
+          },
+          {
+            "id": 14675,
+            "origin_id": -10,
+            "origin_slot": 7,
+            "target_id": 5543,
+            "target_slot": 2,
+            "type": "SAMPLER"
+          },
+          {
+            "id": 14773,
+            "origin_id": -10,
+            "origin_slot": 8,
+            "target_id": 5543,
+            "target_slot": 3,
+            "type": "SIGMAS"
+          },
+          {
+            "id": 14683,
+            "origin_id": -10,
+            "origin_slot": 9,
+            "target_id": 5545,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14691,
+            "origin_id": -10,
+            "origin_slot": 10,
+            "target_id": 5549,
+            "target_slot": 0,
+            "type": "AUDIO"
+          },
+          {
+            "id": 14692,
+            "origin_id": -10,
+            "origin_slot": 11,
+            "target_id": 5549,
+            "target_slot": 1,
+            "type": "VAE"
+          },
+          {
+            "id": 14678,
+            "origin_id": 5521,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14679,
+            "origin_id": 5521,
+            "origin_slot": 1,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14729,
+            "origin_id": 5536,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14730,
+            "origin_id": 5536,
+            "origin_slot": 1,
+            "target_id": -20,
+            "target_slot": 3,
+            "type": "CONDITIONING"
+          },
+          {
+            "id": 14742,
+            "origin_id": 5536,
+            "origin_slot": 2,
+            "target_id": -20,
+            "target_slot": 4,
+            "type": "LATENT"
+          },
+          {
+            "id": 14798,
+            "origin_id": 5547,
+            "origin_slot": 1,
+            "target_id": -20,
+            "target_slot": 5,
+            "type": "LATENT"
+          }
+        ],
+        "extra": {}
+      },
+      {
+        "id": "2eabb368-3326-4160-9984-dd85c7e7e498",
+        "version": 1,
+        "state": {
+          "lastGroupId": 217,
+          "lastNodeId": 9107,
+          "lastLinkId": 14989,
+          "lastRerouteId": 228
+        },
+        "revision": 1,
+        "config": {},
+        "name": "New Subgraph",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            -507.91119182386876,
+            3189.278650785782,
+            128,
+            128
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            735.9619384083298,
+            3159.278650785782,
+            128,
+            188
+          ]
+        },
+        "inputs": [
+          {
+            "id": "8c477ae2-714b-41d1-9529-446f76030021",
+            "name": "input",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14877
+            ],
+            "localized_name": "input",
+            "pos": [
+              51,
+              24
+            ]
+          },
+          {
+            "id": "6bc1504b-1095-43cc-8bac-c19d86abfe05",
+            "name": "input_1",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14878
+            ],
+            "localized_name": "input_1",
+            "pos": [
+              51,
+              44
+            ]
+          },
+          {
+            "id": "83123c82-dd05-448a-83bf-a0d369917548",
+            "name": "samples",
+            "type": "LATENT",
+            "linkIds": [
+              14879
+            ],
+            "localized_name": "samples",
+            "pos": [
+              51,
+              64
+            ]
+          },
+          {
+            "id": "0fadec64-17e8-4c71-9b03-e3cccfe12cfc",
+            "name": "vae",
+            "type": "VAE",
+            "linkIds": [
+              14880
+            ],
+            "localized_name": "vae",
+            "pos": [
+              51,
+              84
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "e780cf22-88cd-43df-9cad-6a48d355f633",
+            "name": "MASK",
+            "type": "MASK",
+            "linkIds": [
+              14881
+            ],
+            "localized_name": "MASK",
+            "pos": [
+              24,
+              24
+            ]
+          },
+          {
+            "id": "ef1aa961-8a67-44f6-9e88-27cfe6975113",
+            "name": "batch_size",
+            "type": "INT",
+            "linkIds": [
+              14882
+            ],
+            "localized_name": "batch_size",
+            "pos": [
+              24,
+              44
+            ]
+          },
+          {
+            "id": "f8c10d1e-5869-48b1-867c-bfec08416ebf",
+            "name": "LATENT",
+            "type": "LATENT",
+            "linkIds": [
+              14883
+            ],
+            "localized_name": "LATENT",
+            "pos": [
+              24,
+              64
+            ]
+          },
+          {
+            "id": "96f65e9f-4efd-4610-b97f-8d749ddf8381",
+            "name": "image",
+            "type": "IMAGE",
+            "linkIds": [
+              14884,
+              14884
+            ],
+            "localized_name": "image",
+            "pos": [
+              24,
+              84
+            ]
+          },
+          {
+            "id": "3f7e4e8f-bf73-4ec7-bf54-e295ea7d9195",
+            "name": "IMAGE",
+            "type": "IMAGE",
+            "linkIds": [
+              14886
+            ],
+            "localized_name": "IMAGE",
+            "pos": [
+              24,
+              104
+            ]
+          },
+          {
+            "id": "cef8c0a4-3f41-4eac-b75d-7eaec3d6b50f",
+            "name": "BOOLEAN",
+            "type": "BOOLEAN",
+            "linkIds": [
+              14888
+            ],
+            "localized_name": "BOOLEAN",
+            "pos": [
+              24,
+              124
+            ]
+          },
+          {
+            "id": "17ae8008-98c9-4db5-b1e6-95a8cb6f5d47",
+            "name": "resized",
+            "type": "IMAGE",
+            "linkIds": [
+              14889
+            ],
+            "localized_name": "resized",
+            "pos": [
+              24,
+              144
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 9065,
+            "type": "MaskExpandBatch+",
+            "pos": [
+              355.05568840832984,
+              3237.0894585488136
+            ],
+            "size": [
+              320.9458984375,
+              135.484375
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "mask",
+                "name": "mask",
+                "type": "MASK",
+                "link": 14868
+              },
+              {
+                "localized_name": "size",
+                "name": "size",
+                "type": "INT",
+                "widget": {
+                  "name": "size"
+                },
+                "link": 14869
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "MASK",
+                "name": "MASK",
+                "type": "MASK",
+                "links": [
+                  14881
+                ]
+              }
+            ],
+            "title": "Half-Res Mask · Match Video Frame Count",
+            "properties": {
+              "cnr_id": "comfyui_essentials",
+              "ver": "1.1.0",
+              "Node name for S&R": "MaskExpandBatch+"
+            },
+            "widgets_values": [
+              121,
+              "expand"
+            ],
+            "color": "#5b4636",
+            "bgcolor": "#3c3027"
+          },
+          {
+            "id": 9066,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              -264.92222244890036,
+              3266.433797066296
+            ],
+            "size": [
+              268.75,
+              126
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14877
+              },
+              {
+                "label": "match",
+                "localized_name": "resize_type.match",
+                "name": "resize_type.match",
+                "type": "IMAGE,MASK",
+                "link": 14870
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14875
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "match size",
+              "center",
+              "area"
+            ]
+          },
+          {
+            "id": 9067,
+            "type": "GetImageSize",
+            "pos": [
+              55.365551246529776,
+              3249.9189952660117
+            ],
+            "size": [
+              225,
+              159.046875
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "link": 14871
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "width",
+                "name": "width",
+                "type": "INT",
+                "links": [
+                  14872
+                ]
+              },
+              {
+                "localized_name": "height",
+                "name": "height",
+                "type": "INT",
+                "links": [
+                  14873
+                ]
+              },
+              {
+                "localized_name": "batch_size",
+                "name": "batch_size",
+                "type": "INT",
+                "links": [
+                  14869,
+                  14874,
+                  14882
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.16.0",
+              "Node name for S&R": "GetImageSize"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9068,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              -266.351633177399,
+              3487.753574978044
+            ],
+            "size": [
+              268.75,
+              106
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14878
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14868,
+                  14870,
+                  14876
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale by multiplier",
+              0.5,
+              "area"
+            ]
+          },
+          {
+            "id": 9069,
+            "type": "EmptyLTXVLatentVideo",
+            "pos": [
+              53.04870558034099,
+              3461.538985708232
+            ],
+            "size": [
+              225,
+              175.453125
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "width",
+                "name": "width",
+                "type": "INT",
+                "widget": {
+                  "name": "width"
+                },
+                "link": 14872
+              },
+              {
+                "localized_name": "height",
+                "name": "height",
+                "type": "INT",
+                "widget": {
+                  "name": "height"
+                },
+                "link": 14873
+              },
+              {
+                "localized_name": "length",
+                "name": "length",
+                "type": "INT",
+                "widget": {
+                  "name": "length"
+                },
+                "link": 14874
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "LATENT",
+                "name": "LATENT",
+                "type": "LATENT",
+                "links": [
+                  14883
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.3.43",
+              "Node name for S&R": "EmptyLTXVLatentVideo"
+            },
+            "widgets_values": [
+              960,
+              544,
+              121,
+              1
+            ]
+          },
+          {
+            "id": 9070,
+            "type": "LTXVInpaintPreprocess",
+            "pos": [
+              -254.6084782673488,
+              3674.161553001181
+            ],
+            "size": [
+              246.875,
+              95.515625
+            ],
+            "flags": {},
+            "order": 8,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 14875
+              },
+              {
+                "localized_name": "mask",
+                "name": "mask",
+                "type": "MASK",
+                "link": 14876
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "links": [
+                  14871,
+                  14884
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "ComfyUI-LTXVideo",
+              "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+              "Node name for S&R": "LTXVInpaintPreprocess"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9071,
+            "type": "VAEDecode",
+            "pos": [
+              50.34490056620734,
+              3692.0782378135063
+            ],
+            "size": [
+              225,
+              71.53125
+            ],
+            "flags": {},
+            "order": 9,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "samples",
+                "name": "samples",
+                "type": "LATENT",
+                "link": 14879
+              },
+              {
+                "localized_name": "vae",
+                "name": "vae",
+                "type": "VAE",
+                "link": 14880
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  14886
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "VAEDecode"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9072,
+            "type": "LoadImage",
+            "pos": [
+              -319.91119182386876,
+              2766.8801235703822
+            ],
+            "size": [
+              319.421875,
+              428.328125
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  14887
+                ]
+              },
+              {
+                "localized_name": "MASK",
+                "name": "MASK",
+                "type": "MASK",
+                "links": []
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.3.56",
+              "Node name for S&R": "LoadImage"
+            },
+            "widgets_values": [
+              "example.png",
+              "image"
+            ]
+          },
+          {
+            "id": 9073,
+            "type": "PrimitiveBoolean",
+            "pos": [
+              48.35072156245121,
+              2814.811600796802
+            ],
+            "size": [
+              268.75,
+              75.53125
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [],
+            "outputs": [
+              {
+                "localized_name": "BOOLEAN",
+                "name": "BOOLEAN",
+                "type": "BOOLEAN",
+                "links": [
+                  14888
+                ]
+              }
+            ],
+            "title": "bypass_i2v",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "PrimitiveBoolean"
+            },
+            "widgets_values": [
+              true
+            ]
+          },
+          {
+            "id": 9074,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              49.86215630340473,
+              2956.4228254616874
+            ],
+            "size": [
+              268.75,
+              106
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14887
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE",
+                "links": [
+                  14889
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.16.0",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale longer dimension",
+              1024,
+              "lanczos"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 14868,
+            "origin_id": 9068,
+            "origin_slot": 0,
+            "target_id": 9065,
+            "target_slot": 0,
+            "type": "MASK"
+          },
+          {
+            "id": 14869,
+            "origin_id": 9067,
+            "origin_slot": 2,
+            "target_id": 9065,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14870,
+            "origin_id": 9068,
+            "origin_slot": 0,
+            "target_id": 9066,
+            "target_slot": 1,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14871,
+            "origin_id": 9070,
+            "origin_slot": 0,
+            "target_id": 9067,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14872,
+            "origin_id": 9067,
+            "origin_slot": 0,
+            "target_id": 9069,
+            "target_slot": 0,
+            "type": "INT"
+          },
+          {
+            "id": 14873,
+            "origin_id": 9067,
+            "origin_slot": 1,
+            "target_id": 9069,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14874,
+            "origin_id": 9067,
+            "origin_slot": 2,
+            "target_id": 9069,
+            "target_slot": 2,
+            "type": "INT"
+          },
+          {
+            "id": 14875,
+            "origin_id": 9066,
+            "origin_slot": 0,
+            "target_id": 9070,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14876,
+            "origin_id": 9068,
+            "origin_slot": 0,
+            "target_id": 9070,
+            "target_slot": 1,
+            "type": "MASK"
+          },
+          {
+            "id": 14887,
+            "origin_id": 9072,
+            "origin_slot": 0,
+            "target_id": 9074,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14877,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 9066,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14878,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 9068,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14879,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 9071,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14880,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 9071,
+            "target_slot": 1,
+            "type": "VAE"
+          },
+          {
+            "id": 14881,
+            "origin_id": 9065,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "MASK"
+          },
+          {
+            "id": 14882,
+            "origin_id": 9067,
+            "origin_slot": 2,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14883,
+            "origin_id": 9069,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "LATENT"
+          },
+          {
+            "id": 14884,
+            "origin_id": 9070,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 3,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14886,
+            "origin_id": 9071,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 4,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14888,
+            "origin_id": 9073,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 5,
+            "type": "BOOLEAN"
+          },
+          {
+            "id": 14889,
+            "origin_id": 9074,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 6,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {}
+      },
+      {
+        "id": "8c3490c3-2129-435f-b9ef-9c340f8a2a25",
+        "version": 1,
+        "state": {
+          "lastGroupId": 217,
+          "lastNodeId": 9107,
+          "lastLinkId": 14989,
+          "lastRerouteId": 228
+        },
+        "revision": 1,
+        "config": {},
+        "name": "New Subgraph",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            2876.252974517697,
+            5179.740098342359,
+            133.224609375,
+            308
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            4551.839338978017,
+            5279.740098342359,
+            128,
+            108
+          ]
+        },
+        "inputs": [
+          {
+            "id": "32ddc223-fc82-43de-a627-5b67c8ff6154",
+            "name": "input",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14927,
+              14928
+            ],
+            "localized_name": "input",
+            "pos": [
+              51,
+              24
+            ]
+          },
+          {
+            "id": "c173bc59-989e-4dd3-bcd0-8455212e1278",
+            "name": "vae",
+            "type": "VAE",
+            "linkIds": [
+              14929,
+              14930,
+              14931
+            ],
+            "localized_name": "vae",
+            "pos": [
+              51,
+              44
+            ]
+          },
+          {
+            "id": "6f4889b1-8aa9-484f-9dd0-86ac2a60e7f2",
+            "name": "target_width",
+            "type": "INT",
+            "linkIds": [
+              14932,
+              14933
+            ],
+            "localized_name": "target_width",
+            "pos": [
+              51,
+              64
+            ]
+          },
+          {
+            "id": "893bdb13-57f5-4d1c-a18e-b1b462d6a439",
+            "name": "target_height",
+            "type": "INT",
+            "linkIds": [
+              14934,
+              14935
+            ],
+            "localized_name": "target_height",
+            "pos": [
+              51,
+              84
+            ]
+          },
+          {
+            "id": "2e17f767-0dbd-4aa1-bdbc-28794e02769e",
+            "name": "on_false",
+            "type": "*",
+            "linkIds": [
+              14936
+            ],
+            "localized_name": "on_false",
+            "pos": [
+              51,
+              104
+            ]
+          },
+          {
+            "id": "cf478415-d1f2-4227-9dd9-69250cade078",
+            "name": "on_true",
+            "type": "*",
+            "linkIds": [
+              14937
+            ],
+            "localized_name": "on_true",
+            "pos": [
+              51,
+              124
+            ]
+          },
+          {
+            "id": "4d4b7177-d98b-4cee-b847-46e217dadd02",
+            "name": "audio_latent",
+            "type": "LATENT",
+            "linkIds": [
+              14938
+            ],
+            "localized_name": "audio_latent",
+            "pos": [
+              51,
+              144
+            ]
+          },
+          {
+            "id": "c4342e2e-f8db-420e-9f5a-7c531ca40616",
+            "name": "noise",
+            "type": "NOISE",
+            "linkIds": [
+              14939
+            ],
+            "localized_name": "noise",
+            "pos": [
+              51,
+              164
+            ]
+          },
+          {
+            "id": "9238b0fe-ef06-4d1a-9d47-0eb678642b76",
+            "name": "guider",
+            "type": "GUIDER",
+            "linkIds": [
+              14940
+            ],
+            "localized_name": "guider",
+            "pos": [
+              51,
+              184
+            ]
+          },
+          {
+            "id": "bff114ea-76df-4323-a8e6-680c5005027c",
+            "name": "sampler",
+            "type": "SAMPLER",
+            "linkIds": [
+              14941
+            ],
+            "localized_name": "sampler",
+            "pos": [
+              51,
+              204
+            ]
+          },
+          {
+            "id": "9f6469d2-3a4d-4e3b-9b56-44aebc22dbb6",
+            "name": "sigmas",
+            "type": "SIGMAS",
+            "linkIds": [
+              14942
+            ],
+            "localized_name": "sigmas",
+            "pos": [
+              51,
+              224
+            ]
+          },
+          {
+            "id": "ee71316a-20ab-4bc3-9a0f-c9c53425d450",
+            "name": "samples",
+            "type": "LATENT",
+            "linkIds": [
+              14943
+            ],
+            "localized_name": "samples",
+            "pos": [
+              51,
+              244
+            ]
+          },
+          {
+            "id": "c3b534cf-f83c-4304-b5ee-8431f9539872",
+            "name": "upscale_model",
+            "type": "LATENT_UPSCALE_MODEL",
+            "linkIds": [
+              14944
+            ],
+            "localized_name": "upscale_model",
+            "pos": [
+              51,
+              264
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "f50a1349-d3f5-44ee-953a-31bc5b260a86",
+            "name": "resized",
+            "type": "IMAGE",
+            "linkIds": [
+              14945
+            ],
+            "localized_name": "resized",
+            "pos": [
+              24,
+              24
+            ]
+          },
+          {
+            "id": "8ae6e313-34ae-4698-96c9-d859bdfd00eb",
+            "name": "images",
+            "type": "IMAGE",
+            "linkIds": [
+              14946
+            ],
+            "localized_name": "images",
+            "pos": [
+              24,
+              44
+            ]
+          },
+          {
+            "id": "5d857f6c-cb9d-4860-b786-f4635ec277a5",
+            "name": "*",
+            "type": "*",
+            "linkIds": [
+              14947
+            ],
+            "localized_name": "*",
+            "pos": [
+              24,
+              64
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 9089,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              3532.142806297189,
+              5478.400147117672
+            ],
+            "size": [
+              268.75,
+              106
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14927
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14945
+                ]
+              }
+            ],
+            "title": "Baseline · Stage 1 + 2x Lanczos",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale by multiplier",
+              1,
+              "lanczos"
+            ]
+          },
+          {
+            "id": 9090,
+            "type": "VAEDecodeTiled",
+            "pos": [
+              4144.095824117898,
+              5071.794354871104
+            ],
+            "size": [
+              337.5,
+              149.984375
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "samples",
+                "name": "samples",
+                "type": "LATENT",
+                "link": 14919
+              },
+              {
+                "localized_name": "vae",
+                "name": "vae",
+                "type": "VAE",
+                "link": 14929
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  14920
+                ]
+              }
+            ],
+            "title": "Stage 2 Stable Refine · Tiled Decode",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "VAEDecodeTiled"
+            },
+            "widgets_values": [
+              512,
+              128,
+              64,
+              16
+            ],
+            "color": "#223",
+            "bgcolor": "#335"
+          },
+          {
+            "id": 9091,
+            "type": "LTXReframeTargetFit",
+            "pos": [
+              4136.298932784887,
+              5284.281995829486
+            ],
+            "size": [
+              340.625,
+              119.515625
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 14920
+              },
+              {
+                "localized_name": "target_width",
+                "name": "target_width",
+                "type": "INT",
+                "link": 14932
+              },
+              {
+                "localized_name": "target_height",
+                "name": "target_height",
+                "type": "INT",
+                "link": 14934
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  14946
+                ]
+              }
+            ],
+            "title": "Stage 2 Stable Refine · Exact Output Fit",
+            "properties": {
+              "Node name for S&R": "LTXReframeTargetFit"
+            },
+            "widgets_values": [],
+            "color": "#223",
+            "bgcolor": "#335"
+          },
+          {
+            "id": 9092,
+            "type": "LazySwitchKJ",
+            "pos": [
+              4028.52431605131,
+              5473.770594170146
+            ],
+            "size": [
+              462.5,
+              135.625
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "on_false",
+                "name": "on_false",
+                "type": "*",
+                "link": 14936
+              },
+              {
+                "localized_name": "on_true",
+                "name": "on_true",
+                "type": "*",
+                "link": 14937
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "*",
+                "name": "*",
+                "type": "*",
+                "links": [
+                  14947
+                ]
+              }
+            ],
+            "title": "Stable Stage 2? · Off = Known-Good Stage 1 Baseline",
+            "properties": {
+              "cnr_id": "comfyui-kjnodes",
+              "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+              "Node name for S&R": "LazySwitchKJ"
+            },
+            "widgets_values": [
+              true
+            ],
+            "color": "#263423",
+            "bgcolor": "#395034"
+          },
+          {
+            "id": 9093,
+            "type": "LTXVConcatAVLatent",
+            "pos": [
+              3531.373582987637,
+              5327.976567857297
+            ],
+            "size": [
+              418.75,
+              71.53125
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video_latent",
+                "name": "video_latent",
+                "type": "LATENT",
+                "link": 14921
+              },
+              {
+                "localized_name": "audio_latent",
+                "name": "audio_latent",
+                "type": "LATENT",
+                "link": 14938
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "latent",
+                "name": "latent",
+                "type": "LATENT",
+                "links": [
+                  14922
+                ]
+              }
+            ],
+            "title": "Stage 2 Stable Refine · Recombine Video + Audio",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.5.1",
+              "Node name for S&R": "LTXVConcatAVLatent"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9094,
+            "type": "SamplerCustomAdvanced",
+            "pos": [
+              3513.227135957148,
+              4868.363938270976
+            ],
+            "size": [
+              371.875,
+              143.53125
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "noise",
+                "name": "noise",
+                "type": "NOISE",
+                "link": 14939
+              },
+              {
+                "localized_name": "guider",
+                "name": "guider",
+                "type": "GUIDER",
+                "link": 14940
+              },
+              {
+                "localized_name": "sampler",
+                "name": "sampler",
+                "type": "SAMPLER",
+                "link": 14941
+              },
+              {
+                "localized_name": "sigmas",
+                "name": "sigmas",
+                "type": "SIGMAS",
+                "link": 14942
+              },
+              {
+                "localized_name": "latent_image",
+                "name": "latent_image",
+                "type": "LATENT",
+                "link": 14922
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "output",
+                "name": "output",
+                "type": "LATENT",
+                "links": [
+                  14923
+                ]
+              },
+              {
+                "localized_name": "denoised_output",
+                "name": "denoised_output",
+                "type": "LATENT",
+                "links": []
+              }
+            ],
+            "title": "Stage 2 Stable Refine · Full-Clip Synthesis",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.14.1",
+              "Node name for S&R": "SamplerCustomAdvanced"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9095,
+            "type": "LTXVSeparateAVLatent",
+            "pos": [
+              3071.512248549222,
+              5080.490158717988
+            ],
+            "size": [
+              334.375,
+              71.53125
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "av_latent",
+                "name": "av_latent",
+                "type": "LATENT",
+                "link": 14923
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "video_latent",
+                "name": "video_latent",
+                "type": "LATENT",
+                "links": [
+                  14919
+                ]
+              },
+              {
+                "localized_name": "audio_latent",
+                "name": "audio_latent",
+                "type": "LATENT",
+                "links": []
+              }
+            ],
+            "title": "Stage 2 Stable Refine · Separate Video",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.5.1",
+              "Node name for S&R": "LTXVSeparateAVLatent"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9096,
+            "type": "LTXReframeTargetFit",
+            "pos": [
+              3069.477583892697,
+              5230.030485744866
+            ],
+            "size": [
+              378.125,
+              119.515625
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 14928
+              },
+              {
+                "localized_name": "target_width",
+                "name": "target_width",
+                "type": "INT",
+                "link": 14933
+              },
+              {
+                "localized_name": "target_height",
+                "name": "target_height",
+                "type": "INT",
+                "link": 14935
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  14924
+                ]
+              }
+            ],
+            "title": "Stage 2 Stable Initializer · Exact 2x Pixel Fit",
+            "properties": {
+              "Node name for S&R": "LTXReframeTargetFit"
+            },
+            "widgets_values": [],
+            "color": "#223",
+            "bgcolor": "#335"
+          },
+          {
+            "id": 9097,
+            "type": "VAEEncodeTiled",
+            "pos": [
+              3527.0974854421174,
+              5104.003549100957
+            ],
+            "size": [
+              521.875,
+              149.984375
+            ],
+            "flags": {},
+            "order": 8,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "pixels",
+                "name": "pixels",
+                "type": "IMAGE",
+                "link": 14924
+              },
+              {
+                "localized_name": "vae",
+                "name": "vae",
+                "type": "VAE",
+                "link": 14930
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "LATENT",
+                "name": "LATENT",
+                "type": "LATENT",
+                "links": [
+                  14925
+                ]
+              }
+            ],
+            "title": "Stage 2 Stable Initializer · Spatial Tiles / Whole-Clip Temporal Encode",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "VAEEncodeTiled"
+            },
+            "widgets_values": [
+              512,
+              128,
+              4096,
+              4
+            ]
+          },
+          {
+            "id": 9098,
+            "type": "LTXVLatentUpsampler",
+            "pos": [
+              3535.4434072102304,
+              5709.600633413741
+            ],
+            "size": [
+              345.984375,
+              119.515625
+            ],
+            "flags": {},
+            "order": 9,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "samples",
+                "name": "samples",
+                "type": "LATENT",
+                "link": 14943
+              },
+              {
+                "localized_name": "upscale_model",
+                "name": "upscale_model",
+                "type": "LATENT_UPSCALE_MODEL",
+                "link": 14944
+              },
+              {
+                "localized_name": "vae",
+                "name": "vae",
+                "type": "VAE",
+                "link": 14931
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "LATENT",
+                "name": "LATENT",
+                "type": "LATENT",
+                "links": [
+                  14926
+                ]
+              }
+            ],
+            "title": "Stage 2 Initializer · Native Latent Upscale",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "LTXVLatentUpsampler"
+            },
+            "widgets_values": [],
+            "color": "#223",
+            "bgcolor": "#335"
+          },
+          {
+            "id": 9099,
+            "type": "LazySwitchKJ",
+            "pos": [
+              4051.2924639780167,
+              4871.598382584078
+            ],
+            "size": [
+              440.546875,
+              134.71875
+            ],
+            "flags": {},
+            "order": 10,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "on_false",
+                "name": "on_false",
+                "type": "*",
+                "link": 14925
+              },
+              {
+                "localized_name": "on_true",
+                "name": "on_true",
+                "type": "*",
+                "link": 14926
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "*",
+                "name": "*",
+                "type": "*",
+                "links": [
+                  14921
+                ]
+              }
+            ],
+            "title": "Native Latent Upscale? · Off = Stable Initializer",
+            "properties": {
+              "cnr_id": "comfyui-kjnodes",
+              "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+              "Node name for S&R": "LazySwitchKJ"
+            },
+            "widgets_values": [
+              true
+            ],
+            "color": "#263423",
+            "bgcolor": "#395034"
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 14919,
+            "origin_id": 9095,
+            "origin_slot": 0,
+            "target_id": 9090,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14920,
+            "origin_id": 9090,
+            "origin_slot": 0,
+            "target_id": 9091,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14921,
+            "origin_id": 9099,
+            "origin_slot": 0,
+            "target_id": 9093,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14922,
+            "origin_id": 9093,
+            "origin_slot": 0,
+            "target_id": 9094,
+            "target_slot": 4,
+            "type": "LATENT"
+          },
+          {
+            "id": 14923,
+            "origin_id": 9094,
+            "origin_slot": 0,
+            "target_id": 9095,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14924,
+            "origin_id": 9096,
+            "origin_slot": 0,
+            "target_id": 9097,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14925,
+            "origin_id": 9097,
+            "origin_slot": 0,
+            "target_id": 9099,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14926,
+            "origin_id": 9098,
+            "origin_slot": 0,
+            "target_id": 9099,
+            "target_slot": 1,
+            "type": "LATENT"
+          },
+          {
+            "id": 14927,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 9089,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14928,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 9096,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14929,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 9090,
+            "target_slot": 1,
+            "type": "VAE"
+          },
+          {
+            "id": 14930,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 9097,
+            "target_slot": 1,
+            "type": "VAE"
+          },
+          {
+            "id": 14931,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 9098,
+            "target_slot": 2,
+            "type": "VAE"
+          },
+          {
+            "id": 14932,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 9091,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14933,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 9096,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14934,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 9091,
+            "target_slot": 2,
+            "type": "INT"
+          },
+          {
+            "id": 14935,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 9096,
+            "target_slot": 2,
+            "type": "INT"
+          },
+          {
+            "id": 14936,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 9092,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14937,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 9092,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14938,
+            "origin_id": -10,
+            "origin_slot": 6,
+            "target_id": 9093,
+            "target_slot": 1,
+            "type": "LATENT"
+          },
+          {
+            "id": 14939,
+            "origin_id": -10,
+            "origin_slot": 7,
+            "target_id": 9094,
+            "target_slot": 0,
+            "type": "NOISE"
+          },
+          {
+            "id": 14940,
+            "origin_id": -10,
+            "origin_slot": 8,
+            "target_id": 9094,
+            "target_slot": 1,
+            "type": "GUIDER"
+          },
+          {
+            "id": 14941,
+            "origin_id": -10,
+            "origin_slot": 9,
+            "target_id": 9094,
+            "target_slot": 2,
+            "type": "SAMPLER"
+          },
+          {
+            "id": 14942,
+            "origin_id": -10,
+            "origin_slot": 10,
+            "target_id": 9094,
+            "target_slot": 3,
+            "type": "SIGMAS"
+          },
+          {
+            "id": 14943,
+            "origin_id": -10,
+            "origin_slot": 11,
+            "target_id": 9098,
+            "target_slot": 0,
+            "type": "LATENT"
+          },
+          {
+            "id": 14944,
+            "origin_id": -10,
+            "origin_slot": 12,
+            "target_id": 9098,
+            "target_slot": 1,
+            "type": "LATENT_UPSCALE_MODEL"
+          },
+          {
+            "id": 14945,
+            "origin_id": 9089,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14946,
+            "origin_id": 9091,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14947,
+            "origin_id": 9092,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {}
+      },
+      {
+        "id": "d100f318-044f-47e1-aae2-b2ea5e0e88df",
+        "version": 1,
+        "state": {
+          "lastGroupId": 217,
+          "lastNodeId": 9107,
+          "lastLinkId": 14989,
+          "lastRerouteId": 228
+        },
+        "revision": 1,
+        "config": {},
+        "name": "New Subgraph",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            3064.7737797934724,
+            5247.072702467725,
+            128,
+            168
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            4192.891925264606,
+            5267.072702467725,
+            128,
+            128
+          ]
+        },
+        "inputs": [
+          {
+            "id": "30fd4f9c-2b36-4705-9578-ca8825c4454d",
+            "name": "input",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14970
+            ],
+            "localized_name": "input",
+            "pos": [
+              51,
+              24
+            ]
+          },
+          {
+            "id": "38d6aa01-9ec0-4a00-a740-311e60a4f0aa",
+            "name": "input_1",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14971
+            ],
+            "localized_name": "input_1",
+            "pos": [
+              51,
+              44
+            ]
+          },
+          {
+            "id": "2c4c7407-9155-44e7-920c-b708db84b2ef",
+            "name": "input_2",
+            "type": "IMAGE,MASK",
+            "linkIds": [
+              14972
+            ],
+            "localized_name": "input_2",
+            "pos": [
+              51,
+              64
+            ]
+          },
+          {
+            "id": "24e2bf35-a476-4269-9dca-53ff533621d1",
+            "name": "size",
+            "type": "INT",
+            "linkIds": [
+              14973
+            ],
+            "localized_name": "size",
+            "pos": [
+              51,
+              84
+            ]
+          },
+          {
+            "id": "57e3a0da-31fb-439c-8e01-303823397c59",
+            "name": "image_from",
+            "type": "IMAGE",
+            "linkIds": [
+              14974
+            ],
+            "localized_name": "image_from",
+            "pos": [
+              51,
+              104
+            ]
+          },
+          {
+            "id": "f3af26ee-b7aa-4be1-a29d-c3d41a9df8a2",
+            "name": "mask",
+            "type": "MASK",
+            "linkIds": [
+              14975
+            ],
+            "localized_name": "mask",
+            "pos": [
+              51,
+              124
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "61de8ae3-4bf1-4ad1-93be-4082e1c727bc",
+            "name": "resized",
+            "type": "IMAGE",
+            "linkIds": [
+              14976
+            ],
+            "localized_name": "resized",
+            "pos": [
+              24,
+              24
+            ]
+          },
+          {
+            "id": "fd393fb0-4ba4-4171-a5cd-69412d288760",
+            "name": "resized_1",
+            "type": "MASK",
+            "linkIds": [
+              14977
+            ],
+            "localized_name": "resized_1",
+            "pos": [
+              24,
+              44
+            ]
+          },
+          {
+            "id": "3e8176ab-eeeb-4e0f-928f-7ae3c567deaa",
+            "name": "MASK",
+            "type": "MASK",
+            "linkIds": [
+              14978
+            ],
+            "localized_name": "MASK",
+            "pos": [
+              24,
+              64
+            ]
+          },
+          {
+            "id": "7754612d-6731-41b7-acf2-a0e62b52862a",
+            "name": "IMAGE",
+            "type": "IMAGE",
+            "linkIds": [
+              14979
+            ],
+            "localized_name": "IMAGE",
+            "pos": [
+              24,
+              84
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 9101,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              3298.117916081525,
+              5098.587118788428
+            ],
+            "size": [
+              268.75,
+              106
+            ],
+            "flags": {
+              "collapsed": false
+            },
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14970
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14964,
+                  14965,
+                  14967,
+                  14969,
+                  14976
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale to multiple",
+              64,
+              "lanczos"
+            ]
+          },
+          {
+            "id": 9102,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              3298.171986539672,
+              5294.379178271545
+            ],
+            "size": [
+              268.75,
+              126
+            ],
+            "flags": {
+              "collapsed": false
+            },
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14971
+              },
+              {
+                "label": "match",
+                "localized_name": "resize_type.match",
+                "name": "resize_type.match",
+                "type": "IMAGE,MASK",
+                "link": 14964
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14966,
+                  14977
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "match size",
+              "center",
+              "area"
+            ]
+          },
+          {
+            "id": 9103,
+            "type": "LTXVInpaintPreprocess",
+            "pos": [
+              3252.7737797934724,
+              5531.481544491496
+            ],
+            "size": [
+              317.234375,
+              95.515625
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 14965
+              },
+              {
+                "localized_name": "mask",
+                "name": "mask",
+                "type": "MASK",
+                "link": 14966
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "links": []
+              }
+            ],
+            "properties": {
+              "cnr_id": "ComfyUI-LTXVideo",
+              "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+              "Node name for S&R": "LTXVInpaintPreprocess"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 9104,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              3692.2669252646056,
+              5060.157910063769
+            ],
+            "size": [
+              440.625,
+              126
+            ],
+            "flags": {
+              "collapsed": false
+            },
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 14972
+              },
+              {
+                "label": "match",
+                "localized_name": "resize_type.match",
+                "name": "resize_type.match",
+                "type": "IMAGE,MASK",
+                "link": 14967
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE,MASK",
+                "links": [
+                  14968
+                ]
+              }
+            ],
+            "title": "Binary Source Mask · Match Final Resolution",
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.20.1",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "match size",
+              "center",
+              "area"
+            ],
+            "color": "#5b4636",
+            "bgcolor": "#3c3027"
+          },
+          {
+            "id": 9105,
+            "type": "MaskExpandBatch+",
+            "pos": [
+              3695.345868141436,
+              5286.776458801101
+            ],
+            "size": [
+              434.375,
+              135.484375
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "mask",
+                "name": "mask",
+                "type": "MASK",
+                "link": 14968
+              },
+              {
+                "localized_name": "size",
+                "name": "size",
+                "type": "INT",
+                "widget": {
+                  "name": "size"
+                },
+                "link": 14973
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "MASK",
+                "name": "MASK",
+                "type": "MASK",
+                "links": [
+                  14978
+                ]
+              }
+            ],
+            "title": "Source Mask · Match Video Frame Count",
+            "properties": {
+              "cnr_id": "comfyui_essentials",
+              "ver": "1.1.0",
+              "Node name for S&R": "MaskExpandBatch+"
+            },
+            "widgets_values": [
+              121,
+              "expand"
+            ],
+            "color": "#5b4636",
+            "bgcolor": "#3c3027"
+          },
+          {
+            "id": 9106,
+            "type": "ImageCompositeFromMaskBatch+",
+            "pos": [
+              3700.9772644276436,
+              5489.096869871682
+            ],
+            "size": [
+              427.84375,
+              142.890625
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image_from",
+                "name": "image_from",
+                "type": "IMAGE",
+                "link": 14974
+              },
+              {
+                "localized_name": "image_to",
+                "name": "image_to",
+                "type": "IMAGE",
+                "link": 14969
+              },
+              {
+                "localized_name": "mask",
+                "name": "mask",
+                "type": "MASK",
+                "link": 14975
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  14979
+                ]
+              }
+            ],
+            "title": "Baseline Only · Stage 1 Source Restoration Composite",
+            "properties": {
+              "cnr_id": "comfyui_essentials",
+              "ver": "1.1.0",
+              "Node name for S&R": "ImageCompositeFromMaskBatch+"
+            },
+            "widgets_values": [],
+            "color": "#35533f",
+            "bgcolor": "#243b2c"
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 14964,
+            "origin_id": 9101,
+            "origin_slot": 0,
+            "target_id": 9102,
+            "target_slot": 1,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14965,
+            "origin_id": 9101,
+            "origin_slot": 0,
+            "target_id": 9103,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14966,
+            "origin_id": 9102,
+            "origin_slot": 0,
+            "target_id": 9103,
+            "target_slot": 1,
+            "type": "MASK"
+          },
+          {
+            "id": 14967,
+            "origin_id": 9101,
+            "origin_slot": 0,
+            "target_id": 9104,
+            "target_slot": 1,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14968,
+            "origin_id": 9104,
+            "origin_slot": 0,
+            "target_id": 9105,
+            "target_slot": 0,
+            "type": "MASK"
+          },
+          {
+            "id": 14969,
+            "origin_id": 9101,
+            "origin_slot": 0,
+            "target_id": 9106,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14970,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 9101,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14971,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 9102,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14972,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 9104,
+            "target_slot": 0,
+            "type": "IMAGE,MASK"
+          },
+          {
+            "id": 14973,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 9105,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 14974,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 9106,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14975,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 9106,
+            "target_slot": 2,
+            "type": "MASK"
+          },
+          {
+            "id": 14976,
+            "origin_id": 9101,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 14977,
+            "origin_id": 9102,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "MASK"
+          },
+          {
+            "id": 14978,
+            "origin_id": 9105,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "MASK"
+          },
+          {
+            "id": 14979,
+            "origin_id": 9106,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 3,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {}
+      }
+    ]
+  },
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.21435888100001196,
+      "offset": [
+        804.2820762650078,
+        -1657.451274140016
+      ]
+    },
+    "reroutes": [
+      {
+        "id": 227,
+        "pos": [
+          2588.7836225212227,
+          3561.962434201263
+        ],
+        "linkIds": [
+          14655
+        ]
+      },
+      {
+        "id": 228,
+        "parentId": 227,
+        "pos": [
+          2880.3306495998672,
+          4096.8065108382225
+        ],
+        "linkIds": [
+          14655
+        ]
+      }
+    ],
+    "frontendVersion": "1.45.15",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true,
+    "linkExtensions": [
+      {
+        "id": 14655,
+        "parentId": 228
+      }
+    ]
+  },
+  "version": 0.4
+}
diff --git a/examples/gemmaREFRAME_v3.4_RENTICULAR_FIXED.json b/examples/gemmaREFRAME_v3.4_RENTICULAR_FIXED.json
new file mode 100644
index 0000000..dba04a1
--- /dev/null
+++ b/examples/gemmaREFRAME_v3.4_RENTICULAR_FIXED.json
@@ -0,0 +1,5017 @@
+{
+  "id": "28666934-dc99-4884-932b-b72715bf49ce",
+  "revision": 1,
+  "last_node_id": 9050,
+  "last_link_id": 14777,
+  "nodes": [
+    {
+      "id": 5521,
+      "type": "LTXVConditioning",
+      "pos": [
+        2223.8711691822973,
+        4600.794391168895
+      ],
+      "size": [
+        225,
+        103.515625
+      ],
+      "flags": {},
+      "order": 24,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14640
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14641
+        },
+        {
+          "name": "frame_rate",
+          "type": "FLOAT",
+          "widget": {
+            "name": "frame_rate"
+          },
+          "link": 14642
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            14678
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "slot_index": 1,
+          "links": [
+            14679
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "LTXVConditioning"
+      },
+      "widgets_values": [
+        24
+      ]
+    },
+    {
+      "id": 5522,
+      "type": "LoadImage",
+      "pos": [
+        1560.6047157269752,
+        4615.746561432165
+      ],
+      "size": [
+        225,
+        339.5
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14654
+          ]
+        },
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.56",
+        "Node name for S&R": "LoadImage"
+      },
+      "widgets_values": [
+        "example.png",
+        "image"
+      ]
+    },
+    {
+      "id": 5523,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        2874.8454718354014,
+        3033.8930990981676
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 15,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14643
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14645,
+            14646,
+            14648,
+            14700,
+            14705
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale to multiple",
+        64,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5524,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        2874.8995422935486,
+        3229.685158581285
+      ],
+      "size": [
+        268.75,
+        167.46875
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 22,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14644
+        },
+        {
+          "label": "match",
+          "name": "resize_type.match",
+          "type": "IMAGE,MASK",
+          "link": 14645
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "MASK",
+          "links": [
+            14647,
+            14667
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "match size",
+        "center",
+        "area"
+      ]
+    },
+    {
+      "id": 5525,
+      "type": "LTXVInpaintPreprocess",
+      "pos": [
+        2884.2411782217823,
+        3458.967745659206
+      ],
+      "size": [
+        246.875,
+        95.515625
+      ],
+      "flags": {},
+      "order": 26,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14646
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14647
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVInpaintPreprocess"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5526,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        2203.2341935182085,
+        3041.545024843583
+      ],
+      "size": [
+        268.75,
+        167.46875
+      ],
+      "flags": {},
+      "order": 28,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14648
+        },
+        {
+          "label": "match",
+          "name": "resize_type.match",
+          "type": "IMAGE,MASK",
+          "link": 14649
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14689
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "match size",
+        "center",
+        "area"
+      ]
+    },
+    {
+      "id": 5527,
+      "type": "CLIPTextEncode",
+      "pos": [
+        4773.29474494944,
+        4340.993216432035
+      ],
+      "size": [
+        534.375,
+        310.625
+      ],
+      "flags": {},
+      "order": 16,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 14650
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            14640
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Positive Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "A woman wearing high-heels walks through a secret garage hideout with henchmen surrounding her."
+      ],
+      "color": "#232",
+      "bgcolor": "#353"
+    },
+    {
+      "id": 5528,
+      "type": "CLIPTextEncode",
+      "pos": [
+        4770.581475366014,
+        4721.216704968021
+      ],
+      "size": [
+        539.53125,
+        306.78125
+      ],
+      "flags": {},
+      "order": 17,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "clip",
+          "type": "CLIP",
+          "link": 14651
+        }
+      ],
+      "outputs": [
+        {
+          "name": "CONDITIONING",
+          "type": "CONDITIONING",
+          "slot_index": 0,
+          "links": [
+            14641
+          ]
+        }
+      ],
+      "title": "CLIP Text Encode (Negative Prompt)",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.28",
+        "Node name for S&R": "CLIPTextEncode"
+      },
+      "widgets_values": [
+        "pc game, console game, video game, cartoon, childish, ugly"
+      ],
+      "color": "#322",
+      "bgcolor": "#533"
+    },
+    {
+      "id": 5531,
+      "type": "PrimitiveBoolean",
+      "pos": [
+        1815.7844184289695,
+        4642.803517310191
+      ],
+      "size": [
+        268.75,
+        75.53125
+      ],
+      "flags": {},
+      "order": 1,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "BOOLEAN",
+          "type": "BOOLEAN",
+          "links": [
+            14671
+          ]
+        }
+      ],
+      "title": "bypass_i2v",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "PrimitiveBoolean"
+      },
+      "widgets_values": [
+        true
+      ]
+    },
+    {
+      "id": 5532,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        1812.2571401136952,
+        4799.531659263295
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {},
+      "order": 14,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14654
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14672
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.16.0",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale longer dimension",
+        1024,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5533,
+      "type": "ManualSigmas",
+      "pos": [
+        2225.0068603491377,
+        4437.974677023865
+      ],
+      "size": [
+        225,
+        107.5
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 2,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            14773
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "ManualSigmas"
+      },
+      "widgets_values": [
+        "1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0"
+      ]
+    },
+    {
+      "id": 5534,
+      "type": "CFGGuider",
+      "pos": [
+        2875.9672150865217,
+        4428.056710839279
+      ],
+      "size": [
+        268.75,
+        127.515625
+      ],
+      "flags": {},
+      "order": 37,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14655
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14656
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14657
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            14674
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ]
+    },
+    {
+      "id": 5535,
+      "type": "GetImageSize",
+      "pos": [
+        2194.1562508223938,
+        3708.0952392566637
+      ],
+      "size": [
+        225,
+        159.515625
+      ],
+      "flags": {},
+      "order": 30,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14658
+        }
+      ],
+      "outputs": [
+        {
+          "name": "width",
+          "type": "INT",
+          "links": [
+            14685
+          ]
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "links": [
+            14686
+          ]
+        },
+        {
+          "name": "batch_size",
+          "type": "INT",
+          "links": [
+            14687,
+            14698,
+            14702
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.16.0",
+        "Node name for S&R": "GetImageSize"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5536,
+      "type": "LTXVCropGuides",
+      "pos": [
+        2230.054083653975,
+        4280.138604836632
+      ],
+      "size": [
+        225,
+        98.125
+      ],
+      "flags": {},
+      "order": 42,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14659
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14660
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14661
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            14729
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            14730
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14742,
+            14765
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.75",
+        "Node name for S&R": "LTXVCropGuides"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5537,
+      "type": "LTXVLaplacianPyramidBlend",
+      "pos": [
+        2527.0562443070453,
+        3356.2487831475346
+      ],
+      "size": [
+        287.5,
+        179.484375
+      ],
+      "flags": {},
+      "order": 46,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_a",
+          "type": "IMAGE",
+          "link": 14744
+        },
+        {
+          "name": "image_b",
+          "type": "IMAGE",
+          "link": 14663
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14664
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14710,
+            14758
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVLaplacianPyramidBlend"
+      },
+      "widgets_values": [
+        true,
+        15
+      ],
+      "color": "#233",
+      "bgcolor": "#355"
+    },
+    {
+      "id": 5538,
+      "type": "KSamplerSelect",
+      "pos": [
+        2804.0989066216716,
+        4268.956494300035
+      ],
+      "size": [
+        343.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            14675
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "res_2s"
+      ]
+    },
+    {
+      "id": 5539,
+      "type": "CreateVideo",
+      "pos": [
+        3463.1642402172997,
+        4905.949851297977
+      ],
+      "size": [
+        231.25,
+        103.515625
+      ],
+      "flags": {},
+      "order": 59,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14741
+        },
+        {
+          "name": "audio",
+          "shape": 7,
+          "type": "AUDIO",
+          "link": 14666
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": [
+            14711
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CreateVideo"
+      },
+      "widgets_values": [
+        25
+      ]
+    },
+    {
+      "id": 5540,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        2201.80478278971,
+        3262.8648027553313
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {},
+      "order": 27,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14667
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "MASK",
+          "links": [
+            14649,
+            14690,
+            14697
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale by multiplier",
+        0.5,
+        "area"
+      ]
+    },
+    {
+      "id": 5541,
+      "type": "LTXVImgToVideoConditionOnly",
+      "pos": [
+        2444.6376386337356,
+        3900.537760212123
+      ],
+      "size": [
+        312.5,
+        179.484375
+      ],
+      "flags": {},
+      "order": 34,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14668
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14669
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14670
+        },
+        {
+          "name": "bypass",
+          "shape": 7,
+          "type": "BOOLEAN",
+          "widget": {
+            "name": "bypass"
+          },
+          "link": 14671
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14681
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVImgToVideoConditionOnly"
+      },
+      "widgets_values": [
+        0.7,
+        false
+      ]
+    },
+    {
+      "id": 5542,
+      "type": "LTXVPreprocess",
+      "pos": [
+        2467.0603913816626,
+        3736.9471384851304
+      ],
+      "size": [
+        268.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 21,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14672
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output_image",
+          "type": "IMAGE",
+          "links": [
+            14669
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LTXVPreprocess"
+      },
+      "widgets_values": [
+        8
+      ]
+    },
+    {
+      "id": 5543,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        2544.666292850023,
+        4272.015580420798
+      ],
+      "size": [
+        225,
+        143.53125
+      ],
+      "flags": {},
+      "order": 40,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 14757
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 14674
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 14675
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 14773
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 14677
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": []
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": [
+            14688
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5544,
+      "type": "LTXAddVideoICLoRAGuideAdvanced",
+      "pos": [
+        2778.0923389512786,
+        3704.2349239675345
+      ],
+      "size": [
+        387.5,
+        443.359375
+      ],
+      "flags": {},
+      "order": 36,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14678
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14679
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14680
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "link": 14681
+        },
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "link": 14682
+        },
+        {
+          "name": "attention_mask",
+          "shape": 7,
+          "type": "MASK",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": [
+            14656,
+            14659,
+            14693
+          ]
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": [
+            14657,
+            14660,
+            14694
+          ]
+        },
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14683
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXAddVideoICLoRAGuideAdvanced"
+      },
+      "widgets_values": [
+        0,
+        1,
+        1,
+        "disabled",
+        false,
+        256,
+        64,
+        1
+      ]
+    },
+    {
+      "id": 5545,
+      "type": "LTXVConcatAVLatent",
+      "pos": [
+        2540.4407239050565,
+        4739.908845395189
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {},
+      "order": 39,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "link": 14683
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14684
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14677
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVConcatAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5546,
+      "type": "EmptyLTXVLatentVideo",
+      "pos": [
+        2195.6741417028356,
+        3948.2975731976926
+      ],
+      "size": [
+        225,
+        175.453125
+      ],
+      "flags": {},
+      "order": 31,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "width",
+          "type": "INT",
+          "widget": {
+            "name": "width"
+          },
+          "link": 14685
+        },
+        {
+          "name": "height",
+          "type": "INT",
+          "widget": {
+            "name": "height"
+          },
+          "link": 14686
+        },
+        {
+          "name": "length",
+          "type": "INT",
+          "widget": {
+            "name": "length"
+          },
+          "link": 14687
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14670
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.3.43",
+        "Node name for S&R": "EmptyLTXVLatentVideo"
+      },
+      "widgets_values": [
+        960,
+        544,
+        121,
+        1
+      ]
+    },
+    {
+      "id": 5547,
+      "type": "LTXVSeparateAVLatent",
+      "pos": [
+        2541.3342857790644,
+        4466.342828035791
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {},
+      "order": 41,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "av_latent",
+          "type": "LATENT",
+          "link": 14688
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "links": [
+            14661
+          ]
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "links": [
+            14727
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVSeparateAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5548,
+      "type": "LTXVInpaintPreprocess",
+      "pos": [
+        2213.54793769976,
+        3449.272780778468
+      ],
+      "size": [
+        246.875,
+        95.515625
+      ],
+      "flags": {},
+      "order": 29,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14689
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14690
+        }
+      ],
+      "outputs": [
+        {
+          "name": "image",
+          "type": "IMAGE",
+          "links": [
+            14658,
+            14663,
+            14682
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVInpaintPreprocess"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5549,
+      "type": "VAEEncodeAudio",
+      "pos": [
+        2539.639253677537,
+        4587.5768672500735
+      ],
+      "size": [
+        225,
+        98.125
+      ],
+      "flags": {},
+      "order": 18,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "link": 14691
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14692
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14695
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "VAEEncodeAudio"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5550,
+      "type": "LTXVSetAudioRefTokens",
+      "pos": [
+        2873.504166994278,
+        4639.587589082661
+      ],
+      "size": [
+        268.75,
+        119.515625
+      ],
+      "flags": {},
+      "order": 38,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14693
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14694
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14695
+        }
+      ],
+      "outputs": [
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "links": []
+        },
+        {
+          "name": "frozen_audio",
+          "type": "LATENT",
+          "links": [
+            14684
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "4f45fd6c222eb06eb3e46605da62e7c889e4be5c",
+        "Node name for S&R": "LTXVSetAudioRefTokens"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5551,
+      "type": "Note",
+      "pos": [
+        2199.3127086287946,
+        4762.265935032064
+      ],
+      "size": [
+        287.609375,
+        238.75
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "Frozen audio helps guiding outpainting to be consistent with the sounds in the video.\nIf you don't want this guidance (e.g. you use this flow to replace a section of the frame or you prefer the off-frame sound to stay off frame) remove audio stream from the generation altogether and/or adjust positive/negative conditioning."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 5552,
+      "type": "SaveVideo",
+      "pos": [
+        1524.259914788955,
+        5137.345213744979
+      ],
+      "size": [
+        3836.1875,
+        1819.984375
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 61,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video",
+          "type": "VIDEO",
+          "link": 14696
+        }
+      ],
+      "outputs": [],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SaveVideo"
+      },
+      "widgets_values": [
+        "video/LTX_Reframe_Outpaint",
+        "mp4",
+        "auto"
+      ]
+    },
+    {
+      "id": 5553,
+      "type": "LTXReframeLayout",
+      "pos": [
+        1529.1783026096316,
+        3004.450389086466
+      ],
+      "size": [
+        560,
+        790
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "padded_frames",
+          "type": "IMAGE",
+          "links": [
+            14643
+          ]
+        },
+        {
+          "name": "outpaint_mask",
+          "type": "MASK",
+          "links": [
+            14644
+          ]
+        },
+        {
+          "name": "source_mask",
+          "type": "MASK",
+          "links": [
+            14699,
+            14777
+          ]
+        },
+        {
+          "name": "source_frames",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "name": "audio",
+          "type": "AUDIO",
+          "links": [
+            14666,
+            14691
+          ]
+        },
+        {
+          "name": "fps",
+          "type": "FLOAT",
+          "links": [
+            14642
+          ]
+        },
+        {
+          "name": "source_video",
+          "type": "VIDEO",
+          "links": [
+            14714
+          ]
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "links": [
+            14721,
+            14759
+          ]
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "links": [
+            14722,
+            14760
+          ]
+        },
+        {
+          "name": "placement_json",
+          "type": "STRING",
+          "links": []
+        }
+      ],
+      "title": "1. LTX Reframe Layout · Source, Framing & Frame Range",
+      "properties": {
+        "Node name for S&R": "LTXReframeLayout",
+        "ltx_reframe_range_v2": {
+          "limit_frames": false,
+          "first_frame": 0,
+          "last_frame": -1
+        }
+      },
+      "widgets_values": [
+        "LTX-2_00175_.mp4",
+        "21:9",
+        1792,
+        768,
+        0.493,
+        0.461,
+        0.591,
+        0,
+        false,
+        0,
+        -1,
+        "image",
+        ""
+      ],
+      "color": "#1f4d52",
+      "bgcolor": "#16383c"
+    },
+    {
+      "id": 5554,
+      "type": "MaskExpandBatch+",
+      "pos": [
+        2839.306340874133,
+        4871.888559979711
+      ],
+      "size": [
+        320.921875,
+        135.484375
+      ],
+      "flags": {},
+      "order": 32,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14697
+        },
+        {
+          "name": "size",
+          "type": "INT",
+          "widget": {
+            "name": "size"
+          },
+          "link": 14698
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14664
+          ]
+        }
+      ],
+      "title": "Half-Res Mask · Match Video Frame Count",
+      "properties": {
+        "cnr_id": "comfyui_essentials",
+        "ver": "1.1.0",
+        "Node name for S&R": "MaskExpandBatch+"
+      },
+      "widgets_values": [
+        121,
+        "expand"
+      ],
+      "color": "#5b4636",
+      "bgcolor": "#3c3027"
+    },
+    {
+      "id": 5555,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        3242.883103203625,
+        2986.790893103916
+      ],
+      "size": [
+        440.625,
+        167.46875
+      ],
+      "flags": {
+        "collapsed": false
+      },
+      "order": 23,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14699
+        },
+        {
+          "label": "match",
+          "name": "resize_type.match",
+          "type": "IMAGE,MASK",
+          "link": 14700
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "MASK",
+          "links": [
+            14701
+          ]
+        }
+      ],
+      "title": "Binary Source Mask · Match Final Resolution",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.20.1",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "match size",
+        "center",
+        "area"
+      ],
+      "color": "#5b4636",
+      "bgcolor": "#3c3027"
+    },
+    {
+      "id": 5556,
+      "type": "MaskExpandBatch+",
+      "pos": [
+        3245.2666774281115,
+        3206.2789221268936
+      ],
+      "size": [
+        434.375,
+        135.484375
+      ],
+      "flags": {},
+      "order": 33,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14701
+        },
+        {
+          "name": "size",
+          "type": "INT",
+          "widget": {
+            "name": "size"
+          },
+          "link": 14702
+        }
+      ],
+      "outputs": [
+        {
+          "name": "MASK",
+          "type": "MASK",
+          "links": [
+            14703
+          ]
+        }
+      ],
+      "title": "Source Mask · Match Video Frame Count",
+      "properties": {
+        "cnr_id": "comfyui_essentials",
+        "ver": "1.1.0",
+        "Node name for S&R": "MaskExpandBatch+"
+      },
+      "widgets_values": [
+        121,
+        "expand"
+      ],
+      "color": "#5b4636",
+      "bgcolor": "#3c3027"
+    },
+    {
+      "id": 5557,
+      "type": "GrowMaskWithBlur",
+      "pos": [
+        4292.295265283913,
+        3004.990272923734
+      ],
+      "size": [
+        315.625,
+        339.390625
+      ],
+      "flags": {},
+      "order": 35,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14703
+        }
+      ],
+      "outputs": [
+        {
+          "name": "mask",
+          "type": "MASK",
+          "links": [
+            14706
+          ]
+        },
+        {
+          "name": "mask_inverted",
+          "type": "MASK",
+          "links": []
+        }
+      ],
+      "title": "Source Seam · Inward 12px / Blur 8px",
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "GrowMaskWithBlur"
+      },
+      "widgets_values": [
+        -128,
+        0,
+        true,
+        false,
+        0,
+        1,
+        1,
+        false
+      ],
+      "color": "#35533f",
+      "bgcolor": "#243b2c"
+    },
+    {
+      "id": 5558,
+      "type": "ImageCompositeFromMaskBatch+",
+      "pos": [
+        3790.7069577475095,
+        3084.3559637211615
+      ],
+      "size": [
+        393.75,
+        145
+      ],
+      "flags": {},
+      "order": 49,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "image_from",
+          "type": "IMAGE",
+          "link": 14738
+        },
+        {
+          "name": "image_to",
+          "type": "IMAGE",
+          "link": 14705
+        },
+        {
+          "name": "mask",
+          "type": "MASK",
+          "link": 14706
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14739
+          ]
+        }
+      ],
+      "title": "Baseline Only · Stage 1 Source Restoration Composite",
+      "properties": {
+        "cnr_id": "comfyui_essentials",
+        "ver": "1.1.0",
+        "Node name for S&R": "ImageCompositeFromMaskBatch+"
+      },
+      "widgets_values": [],
+      "color": "#35533f",
+      "bgcolor": "#243b2c"
+    },
+    {
+      "id": 5559,
+      "type": "c1aa79cc-b9f8-43f1-8d88-3d7406881f62",
+      "pos": [
+        4087.7924094614086,
+        4874.636205745332
+      ],
+      "size": [
+        287.5,
+        159.5
+      ],
+      "flags": {},
+      "order": 62,
+      "mode": 4,
+      "inputs": [
+        {
+          "label": "Before Video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": 14714
+        },
+        {
+          "label": "After Video",
+          "name": "video_1",
+          "type": "VIDEO",
+          "link": 14708
+        }
+      ],
+      "outputs": [
+        {
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": []
+        }
+      ],
+      "properties": {
+        "proxyWidgets": [
+          [
+            "79",
+            "direction"
+          ],
+          [
+            "79",
+            "match_image_size"
+          ],
+          [
+            "79",
+            "spacing_width"
+          ],
+          [
+            "79",
+            "spacing_color"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.13.0"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5561,
+      "type": "ResizeImageMaskNode",
+      "pos": [
+        3770.7859207492493,
+        4514.16258336215
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {},
+      "order": 47,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "input",
+          "type": "IMAGE,MASK",
+          "link": 14710
+        }
+      ],
+      "outputs": [
+        {
+          "name": "resized",
+          "type": "IMAGE",
+          "links": [
+            14738
+          ]
+        }
+      ],
+      "title": "Baseline · Stage 1 + 2x Lanczos",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "ResizeImageMaskNode"
+      },
+      "widgets_values": [
+        "scale by multiplier",
+        1,
+        "lanczos"
+      ]
+    },
+    {
+      "id": 5562,
+      "type": "easy cleanGpuUsed",
+      "pos": [
+        3768.3403021886097,
+        4908.246701809883
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {},
+      "order": 60,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "anything",
+          "type": "*",
+          "link": 14711
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "*",
+          "links": [
+            14696,
+            14708
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-easy-use",
+        "ver": "1.3.6",
+        "Node name for S&R": "easy cleanGpuUsed"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5563,
+      "type": "DualCLIPLoader",
+      "pos": [
+        4778.511613714791,
+        4085.2601847280166
+      ],
+      "size": [
+        518.75,
+        201.25
+      ],
+      "flags": {},
+      "order": 6,
+      "mode": 0,
+      "showAdvanced": false,
+      "inputs": [
+        {
+          "label": "clip_name1",
+          "name": "clip_name1",
+          "type": "COMBO",
+          "widget": {
+            "name": "clip_name1"
+          },
+          "link": null
+        },
+        {
+          "label": "clip_name2",
+          "name": "clip_name2",
+          "type": "COMBO",
+          "widget": {
+            "name": "clip_name2"
+          },
+          "link": null
+        },
+        {
+          "label": "type",
+          "name": "type",
+          "type": "COMBO",
+          "widget": {
+            "name": "type"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "shape": 7,
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "CLIP",
+          "name": "CLIP",
+          "type": "CLIP",
+          "links": [
+            14650,
+            14651
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "DualCLIPLoader",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "clip_name1": true,
+          "clip_name2": true,
+          "type": true,
+          "device": true
+        }
+      },
+      "widgets_values": [
+        "gemma_3_12B_it_fp4_mixed.safetensors",
+        "ltx-2.3_text_projection_bf16.safetensors",
+        "ltxv",
+        "default"
+      ]
+    },
+    {
+      "id": 5564,
+      "type": "VAELoaderKJ",
+      "pos": [
+        4778.426192757166,
+        3856.565521438628
+      ],
+      "size": [
+        518.75,
+        167.453125
+      ],
+      "flags": {},
+      "order": 7,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "vae_name",
+          "name": "vae_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "vae_name"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "VAE",
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            14692
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "VAELoaderKJ",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "device": true,
+          "vae_name": true
+        }
+      },
+      "widgets_values": [
+        "LTX23_audio_vae_bf16.safetensors",
+        "main_device",
+        "bf16"
+      ]
+    },
+    {
+      "id": 5565,
+      "type": "VAELoaderKJ",
+      "pos": [
+        4781.1945737961005,
+        3634.0852460334377
+      ],
+      "size": [
+        518.75,
+        167.453125
+      ],
+      "flags": {},
+      "order": 8,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "vae_name",
+          "name": "vae_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "vae_name"
+          },
+          "link": null
+        },
+        {
+          "label": "device",
+          "name": "device",
+          "type": "COMBO",
+          "widget": {
+            "name": "device"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "VAE",
+          "name": "VAE",
+          "type": "VAE",
+          "links": [
+            14668,
+            14680,
+            14719,
+            14743,
+            14762,
+            14767
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "VAELoaderKJ",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "device": true,
+          "vae_name": true
+        }
+      },
+      "widgets_values": [
+        "LTX23_video_vae_bf16.safetensors",
+        "main_device",
+        "bf16"
+      ]
+    },
+    {
+      "id": 5566,
+      "type": "UNETLoader",
+      "pos": [
+        4778.26174461125,
+        3030.8403719514695
+      ],
+      "size": [
+        521.875,
+        107.5
+      ],
+      "flags": {},
+      "order": 9,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "unet_name",
+          "name": "unet_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "unet_name"
+          },
+          "link": null
+        },
+        {
+          "label": "weight_dtype",
+          "name": "weight_dtype",
+          "type": "COMBO",
+          "widget": {
+            "name": "weight_dtype"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "MODEL",
+          "name": "MODEL",
+          "type": "MODEL",
+          "links": [
+            14713,
+            14771
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.18.1",
+        "Node name for S&R": "UNETLoader",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "weight_dtype": true,
+          "unet_name": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors",
+        "default"
+      ]
+    },
+    {
+      "id": 5567,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        4774.991902743204,
+        3415.4732191311464
+      ],
+      "size": [
+        521.875,
+        160.625
+      ],
+      "flags": {},
+      "order": 25,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "link": 14712
+        },
+        {
+          "label": "lora_name",
+          "name": "lora_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "lora_name"
+          },
+          "link": null
+        },
+        {
+          "label": "strength_model",
+          "name": "strength_model",
+          "type": "FLOAT",
+          "widget": {
+            "name": "strength_model"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            14655
+          ]
+        },
+        {
+          "label": "latent_downscale_factor",
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "531512f7286963dc7aff1fd8bf5556e95eae03af",
+        "Node name for S&R": "LTXICLoRALoaderModelOnly",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "lora_name": true,
+          "strength_model": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors",
+        0.75
+      ]
+    },
+    {
+      "id": 5568,
+      "type": "LTXICLoRALoaderModelOnly",
+      "pos": [
+        4775.880180380794,
+        3195.84604541235
+      ],
+      "size": [
+        518.75,
+        159.484375
+      ],
+      "flags": {},
+      "order": 19,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "link": 14713
+        },
+        {
+          "label": "lora_name",
+          "name": "lora_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "lora_name"
+          },
+          "link": null
+        },
+        {
+          "label": "strength_model",
+          "name": "strength_model",
+          "type": "FLOAT",
+          "widget": {
+            "name": "strength_model"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "model",
+          "name": "model",
+          "type": "MODEL",
+          "links": [
+            14712,
+            14728
+          ]
+        },
+        {
+          "label": "latent_downscale_factor",
+          "name": "latent_downscale_factor",
+          "type": "FLOAT",
+          "links": []
+        }
+      ],
+      "properties": {
+        "cnr_id": "ComfyUI-LTXVideo",
+        "ver": "531512f7286963dc7aff1fd8bf5556e95eae03af",
+        "Node name for S&R": "LTXICLoRALoaderModelOnly",
+        "ue_properties": {
+          "widget_ue_connectable": {},
+          "input_ue_unconnectable": {},
+          "version": "7.5.2"
+        },
+        "widget_ue_connectable": {
+          "lora_name": true,
+          "strength_model": true
+        }
+      },
+      "widgets_values": [
+        "ltx-2.3-22b-ic-lora-deblur-0.9.safetensors",
+        0.5
+      ]
+    },
+    {
+      "id": 5571,
+      "type": "VAEDecodeTiled",
+      "pos": [
+        4307.970821887195,
+        3721.582301829511
+      ],
+      "size": [
+        337.5,
+        103.515625
+      ],
+      "flags": {},
+      "order": 55,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14764
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14719
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14720
+          ]
+        }
+      ],
+      "title": "Stage 2 Stable Refine · Tiled Decode",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "VAEDecodeTiled"
+      },
+      "widgets_values": [
+        512,
+        128,
+        64,
+        16
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 5572,
+      "type": "LTXReframeTargetFit",
+      "pos": [
+        4312.216589678334,
+        3904.6347656976636
+      ],
+      "size": [
+        340.625,
+        119.515625
+      ],
+      "flags": {},
+      "order": 56,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14720
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "link": 14721
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "link": 14722
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            14775
+          ]
+        }
+      ],
+      "title": "Stage 2 Stable Refine · Exact Output Fit",
+      "properties": {
+        "Node name for S&R": "LTXReframeTargetFit"
+      },
+      "widgets_values": [],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 5573,
+      "type": "LazySwitchKJ",
+      "pos": [
+        4198.716221962536,
+        4102.913726457322
+      ],
+      "size": [
+        462.5,
+        135.625
+      ],
+      "flags": {},
+      "order": 58,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "on_false",
+          "type": "*",
+          "link": 14739
+        },
+        {
+          "name": "on_true",
+          "type": "*",
+          "link": 14776
+        }
+      ],
+      "outputs": [
+        {
+          "name": "*",
+          "type": "*",
+          "links": [
+            14741
+          ]
+        }
+      ],
+      "title": "Stable Stage 2? · Off = Known-Good Stage 1 Baseline",
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "LazySwitchKJ"
+      },
+      "widgets_values": [
+        true
+      ],
+      "color": "#263423",
+      "bgcolor": "#395034"
+    },
+    {
+      "id": 5574,
+      "type": "RandomNoise",
+      "pos": [
+        3240.714780798453,
+        3879.0174108322
+      ],
+      "size": [
+        393.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 10,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "NOISE",
+          "type": "NOISE",
+          "links": [
+            14731,
+            14757
+          ]
+        }
+      ],
+      "title": "Stage 2 Noise · Shared Seed",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "RandomNoise"
+      },
+      "widgets_values": [
+        42,
+        "fixed"
+      ]
+    },
+    {
+      "id": 5575,
+      "type": "CFGGuider",
+      "pos": [
+        3770.756246069819,
+        4109.357417020128
+      ],
+      "size": [
+        353.125,
+        127.515625
+      ],
+      "flags": {},
+      "order": 43,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14728
+        },
+        {
+          "name": "positive",
+          "type": "CONDITIONING",
+          "link": 14729
+        },
+        {
+          "name": "negative",
+          "type": "CONDITIONING",
+          "link": 14730
+        }
+      ],
+      "outputs": [
+        {
+          "name": "GUIDER",
+          "type": "GUIDER",
+          "links": [
+            14732
+          ]
+        }
+      ],
+      "title": "Stage 2 Guider · Distilled Model Only",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "CFGGuider"
+      },
+      "widgets_values": [
+        1
+      ]
+    },
+    {
+      "id": 5576,
+      "type": "KSamplerSelect",
+      "pos": [
+        3288.634542361643,
+        4033.468006419744
+      ],
+      "size": [
+        343.75,
+        79.515625
+      ],
+      "flags": {},
+      "order": 11,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "SAMPLER",
+          "type": "SAMPLER",
+          "links": [
+            14733
+          ]
+        }
+      ],
+      "title": "Stage 2 Sampler · Euler CFG++",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "KSamplerSelect"
+      },
+      "widgets_values": [
+        "euler_ancestral_cfg_pp"
+      ]
+    },
+    {
+      "id": 5578,
+      "type": "LTXVConcatAVLatent",
+      "pos": [
+        3246.961216079445,
+        4388.98830555648
+      ],
+      "size": [
+        418.75,
+        71.53125
+      ],
+      "flags": {},
+      "order": 52,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "link": 14770
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "link": 14727
+        }
+      ],
+      "outputs": [
+        {
+          "name": "latent",
+          "type": "LATENT",
+          "links": [
+            14735
+          ]
+        }
+      ],
+      "title": "Stage 2 Stable Refine · Recombine Video + Audio",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVConcatAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5579,
+      "type": "SamplerCustomAdvanced",
+      "pos": [
+        3765.1519845497905,
+        3699.21671130008
+      ],
+      "size": [
+        371.875,
+        143.53125
+      ],
+      "flags": {},
+      "order": 53,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "noise",
+          "type": "NOISE",
+          "link": 14731
+        },
+        {
+          "name": "guider",
+          "type": "GUIDER",
+          "link": 14732
+        },
+        {
+          "name": "sampler",
+          "type": "SAMPLER",
+          "link": 14733
+        },
+        {
+          "name": "sigmas",
+          "type": "SIGMAS",
+          "link": 14774
+        },
+        {
+          "name": "latent_image",
+          "type": "LATENT",
+          "link": 14735
+        }
+      ],
+      "outputs": [
+        {
+          "name": "output",
+          "type": "LATENT",
+          "links": [
+            14736
+          ]
+        },
+        {
+          "name": "denoised_output",
+          "type": "LATENT",
+          "links": []
+        }
+      ],
+      "title": "Stage 2 Stable Refine · Full-Clip Synthesis",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1",
+        "Node name for S&R": "SamplerCustomAdvanced"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5580,
+      "type": "LTXVSeparateAVLatent",
+      "pos": [
+        3780.396372827632,
+        3552.238041634808
+      ],
+      "size": [
+        334.375,
+        71.53125
+      ],
+      "flags": {},
+      "order": 54,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "av_latent",
+          "type": "LATENT",
+          "link": 14736
+        }
+      ],
+      "outputs": [
+        {
+          "name": "video_latent",
+          "type": "LATENT",
+          "links": [
+            14764
+          ]
+        },
+        {
+          "name": "audio_latent",
+          "type": "LATENT",
+          "links": []
+        }
+      ],
+      "title": "Stage 2 Stable Refine · Separate Video",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.5.1",
+        "Node name for S&R": "LTXVSeparateAVLatent"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 5581,
+      "type": "VAEDecode",
+      "pos": [
+        2563.0096890718787,
+        3216.9706252991077
+      ],
+      "size": [
+        225,
+        71.53125
+      ],
+      "flags": {},
+      "order": 44,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14742
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14743
+        }
+      ],
+      "outputs": [
+        {
+          "name": "IMAGE",
+          "type": "IMAGE",
+          "links": [
+            14744
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "VAEDecode"
+      },
+      "widgets_values": []
+    },
+    {
+      "id": 9040,
+      "type": "LTXReframeTargetFit",
+      "pos": [
+        3256.1281699471074,
+        3515.7784422245127
+      ],
+      "size": [
+        378.125,
+        119.515625
+      ],
+      "flags": {},
+      "order": 48,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14758
+        },
+        {
+          "name": "target_width",
+          "type": "INT",
+          "link": 14759
+        },
+        {
+          "name": "target_height",
+          "type": "INT",
+          "link": 14760
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            14761
+          ]
+        }
+      ],
+      "title": "Stage 2 Stable Initializer · Exact 2x Pixel Fit",
+      "properties": {
+        "Node name for S&R": "LTXReframeTargetFit"
+      },
+      "widgets_values": [],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9041,
+      "type": "VAEEncodeTiled",
+      "pos": [
+        3699.524488081439,
+        3923.746749711841
+      ],
+      "size": [
+        521.875,
+        103.515625
+      ],
+      "flags": {},
+      "order": 50,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "pixels",
+          "type": "IMAGE",
+          "link": 14761
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14762
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14763
+          ]
+        }
+      ],
+      "title": "Stage 2 Stable Initializer · Spatial Tiles / Whole-Clip Temporal Encode",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "VAEEncodeTiled"
+      },
+      "widgets_values": [
+        512,
+        128,
+        4096,
+        4
+      ]
+    },
+    {
+      "id": 9042,
+      "type": "LatentUpscaleModelLoader",
+      "pos": [
+        3244.4080617481095,
+        3714.340044785345
+      ],
+      "size": [
+        387.5,
+        79.515625
+      ],
+      "flags": {},
+      "order": 12,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [
+        {
+          "name": "LATENT_UPSCALE_MODEL",
+          "type": "LATENT_UPSCALE_MODEL",
+          "links": [
+            14766
+          ]
+        }
+      ],
+      "title": "Native LTX Upscale Model · 1.5× / 2×",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LatentUpscaleModelLoader"
+      },
+      "widgets_values": [
+        "ltx-2.3-spatial-upscaler-x2-1.1.safetensors"
+      ],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9043,
+      "type": "LTXVLatentUpsampler",
+      "pos": [
+        3770.9772327459195,
+        4319.6484045027
+      ],
+      "size": [
+        309.375,
+        120
+      ],
+      "flags": {},
+      "order": 45,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "samples",
+          "type": "LATENT",
+          "link": 14765
+        },
+        {
+          "name": "upscale_model",
+          "type": "LATENT_UPSCALE_MODEL",
+          "link": 14766
+        },
+        {
+          "name": "vae",
+          "type": "VAE",
+          "link": 14767
+        }
+      ],
+      "outputs": [
+        {
+          "name": "LATENT",
+          "type": "LATENT",
+          "links": [
+            14769
+          ]
+        }
+      ],
+      "title": "Stage 2 Initializer · Native Latent Upscale",
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "LTXVLatentUpsampler"
+      },
+      "widgets_values": [],
+      "color": "#223",
+      "bgcolor": "#335"
+    },
+    {
+      "id": 9044,
+      "type": "LazySwitchKJ",
+      "pos": [
+        4268.9362335743,
+        3514.407950576732
+      ],
+      "size": [
+        378.125,
+        135.625
+      ],
+      "flags": {},
+      "order": 51,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "on_false",
+          "type": "*",
+          "link": 14763
+        },
+        {
+          "name": "on_true",
+          "type": "*",
+          "link": 14769
+        }
+      ],
+      "outputs": [
+        {
+          "name": "*",
+          "type": "*",
+          "links": [
+            14770
+          ]
+        }
+      ],
+      "title": "Native Latent Upscale? · Off = v3.3 Stable Initializer",
+      "properties": {
+        "cnr_id": "comfyui-kjnodes",
+        "ver": "e44c61785708f3d917fe26a0ddd810f6aa598f7f",
+        "Node name for S&R": "LazySwitchKJ"
+      },
+      "widgets_values": [
+        true
+      ],
+      "color": "#263423",
+      "bgcolor": "#395034"
+    },
+    {
+      "id": 9045,
+      "type": "BasicScheduler",
+      "pos": [
+        3292.597732113014,
+        4181.631940671632
+      ],
+      "size": [
+        268.75,
+        143.46875
+      ],
+      "flags": {},
+      "order": 20,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "model",
+          "type": "MODEL",
+          "link": 14771
+        }
+      ],
+      "outputs": [
+        {
+          "name": "SIGMAS",
+          "type": "SIGMAS",
+          "links": [
+            14774
+          ]
+        }
+      ],
+      "properties": {
+        "cnr_id": "comfy-core",
+        "ver": "0.24.0",
+        "Node name for S&R": "BasicScheduler"
+      },
+      "widgets_values": [
+        "simple",
+        3,
+        0.2
+      ]
+    },
+    {
+      "id": 9046,
+      "type": "Note",
+      "pos": [
+        5905.749921382536,
+        3770.973382423348
+      ],
+      "size": [
+        484.515625,
+        105
+      ],
+      "flags": {},
+      "order": 13,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "properties": {},
+      "widgets_values": [
+        "GAUSSIAN SPLAT AS CAMERA CONTROL + ALTERNATE CAMERA LoRA"
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
+    },
+    {
+      "id": 9047,
+      "type": "LTXReframeRectilinearCorrection",
+      "pos": [
+        4276.238198854122,
+        4391.626197800758
+      ],
+      "size": [
+        331.25,
+        347.359375
+      ],
+      "flags": {},
+      "order": 57,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "link": 14775
+        },
+        {
+          "name": "source_mask",
+          "shape": 7,
+          "type": "MASK",
+          "link": 14777
+        }
+      ],
+      "outputs": [
+        {
+          "name": "images",
+          "type": "IMAGE",
+          "links": [
+            14776
+          ]
+        },
+        {
+          "name": "correction_mask",
+          "type": "MASK",
+          "links": null
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "LTXReframeRectilinearCorrection"
+      },
+      "widgets_values": [
+        true,
+        0.065,
+        0.5,
+        0.5,
+        0.45,
+        192,
+        "bicubic",
+        8
+      ]
+    }
+  ],
+  "links": [
+    [
+      14640,
+      5527,
+      0,
+      5521,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14641,
+      5528,
+      0,
+      5521,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14642,
+      5553,
+      5,
+      5521,
+      2,
+      "FLOAT"
+    ],
+    [
+      14643,
+      5553,
+      0,
+      5523,
+      0,
+      "IMAGE"
+    ],
+    [
+      14644,
+      5553,
+      1,
+      5524,
+      0,
+      "MASK"
+    ],
+    [
+      14645,
+      5523,
+      0,
+      5524,
+      1,
+      "IMAGE"
+    ],
+    [
+      14646,
+      5523,
+      0,
+      5525,
+      0,
+      "IMAGE"
+    ],
+    [
+      14647,
+      5524,
+      0,
+      5525,
+      1,
+      "MASK"
+    ],
+    [
+      14648,
+      5523,
+      0,
+      5526,
+      0,
+      "IMAGE"
+    ],
+    [
+      14649,
+      5540,
+      0,
+      5526,
+      1,
+      "IMAGE,MASK"
+    ],
+    [
+      14650,
+      5563,
+      0,
+      5527,
+      0,
+      "CLIP"
+    ],
+    [
+      14651,
+      5563,
+      0,
+      5528,
+      0,
+      "CLIP"
+    ],
+    [
+      14654,
+      5522,
+      0,
+      5532,
+      0,
+      "IMAGE"
+    ],
+    [
+      14655,
+      5567,
+      0,
+      5534,
+      0,
+      "MODEL"
+    ],
+    [
+      14656,
+      5544,
+      0,
+      5534,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14657,
+      5544,
+      1,
+      5534,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      14658,
+      5548,
+      0,
+      5535,
+      0,
+      "IMAGE"
+    ],
+    [
+      14659,
+      5544,
+      0,
+      5536,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14660,
+      5544,
+      1,
+      5536,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14661,
+      5547,
+      0,
+      5536,
+      2,
+      "LATENT"
+    ],
+    [
+      14663,
+      5548,
+      0,
+      5537,
+      1,
+      "IMAGE"
+    ],
+    [
+      14664,
+      5554,
+      0,
+      5537,
+      2,
+      "MASK"
+    ],
+    [
+      14666,
+      5553,
+      4,
+      5539,
+      1,
+      "AUDIO"
+    ],
+    [
+      14667,
+      5524,
+      0,
+      5540,
+      0,
+      "MASK"
+    ],
+    [
+      14668,
+      5565,
+      0,
+      5541,
+      0,
+      "VAE"
+    ],
+    [
+      14669,
+      5542,
+      0,
+      5541,
+      1,
+      "IMAGE"
+    ],
+    [
+      14670,
+      5546,
+      0,
+      5541,
+      2,
+      "LATENT"
+    ],
+    [
+      14671,
+      5531,
+      0,
+      5541,
+      3,
+      "BOOLEAN"
+    ],
+    [
+      14672,
+      5532,
+      0,
+      5542,
+      0,
+      "IMAGE"
+    ],
+    [
+      14674,
+      5534,
+      0,
+      5543,
+      1,
+      "GUIDER"
+    ],
+    [
+      14675,
+      5538,
+      0,
+      5543,
+      2,
+      "SAMPLER"
+    ],
+    [
+      14677,
+      5545,
+      0,
+      5543,
+      4,
+      "LATENT"
+    ],
+    [
+      14678,
+      5521,
+      0,
+      5544,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14679,
+      5521,
+      1,
+      5544,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14680,
+      5565,
+      0,
+      5544,
+      2,
+      "VAE"
+    ],
+    [
+      14681,
+      5541,
+      0,
+      5544,
+      3,
+      "LATENT"
+    ],
+    [
+      14682,
+      5548,
+      0,
+      5544,
+      4,
+      "IMAGE"
+    ],
+    [
+      14683,
+      5544,
+      2,
+      5545,
+      0,
+      "LATENT"
+    ],
+    [
+      14684,
+      5550,
+      2,
+      5545,
+      1,
+      "LATENT"
+    ],
+    [
+      14685,
+      5535,
+      0,
+      5546,
+      0,
+      "INT"
+    ],
+    [
+      14686,
+      5535,
+      1,
+      5546,
+      1,
+      "INT"
+    ],
+    [
+      14687,
+      5535,
+      2,
+      5546,
+      2,
+      "INT"
+    ],
+    [
+      14688,
+      5543,
+      1,
+      5547,
+      0,
+      "LATENT"
+    ],
+    [
+      14689,
+      5526,
+      0,
+      5548,
+      0,
+      "IMAGE"
+    ],
+    [
+      14690,
+      5540,
+      0,
+      5548,
+      1,
+      "MASK"
+    ],
+    [
+      14691,
+      5553,
+      4,
+      5549,
+      0,
+      "AUDIO"
+    ],
+    [
+      14692,
+      5564,
+      0,
+      5549,
+      1,
+      "VAE"
+    ],
+    [
+      14693,
+      5544,
+      0,
+      5550,
+      0,
+      "CONDITIONING"
+    ],
+    [
+      14694,
+      5544,
+      1,
+      5550,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14695,
+      5549,
+      0,
+      5550,
+      2,
+      "LATENT"
+    ],
+    [
+      14696,
+      5562,
+      0,
+      5552,
+      0,
+      "VIDEO"
+    ],
+    [
+      14697,
+      5540,
+      0,
+      5554,
+      0,
+      "MASK"
+    ],
+    [
+      14698,
+      5535,
+      2,
+      5554,
+      1,
+      "INT"
+    ],
+    [
+      14699,
+      5553,
+      2,
+      5555,
+      0,
+      "MASK"
+    ],
+    [
+      14700,
+      5523,
+      0,
+      5555,
+      1,
+      "IMAGE"
+    ],
+    [
+      14701,
+      5555,
+      0,
+      5556,
+      0,
+      "MASK"
+    ],
+    [
+      14702,
+      5535,
+      2,
+      5556,
+      1,
+      "INT"
+    ],
+    [
+      14703,
+      5556,
+      0,
+      5557,
+      0,
+      "MASK"
+    ],
+    [
+      14705,
+      5523,
+      0,
+      5558,
+      1,
+      "IMAGE"
+    ],
+    [
+      14706,
+      5557,
+      0,
+      5558,
+      2,
+      "MASK"
+    ],
+    [
+      14708,
+      5562,
+      0,
+      5559,
+      1,
+      "VIDEO"
+    ],
+    [
+      14710,
+      5537,
+      0,
+      5561,
+      0,
+      "IMAGE"
+    ],
+    [
+      14711,
+      5539,
+      0,
+      5562,
+      0,
+      "VIDEO"
+    ],
+    [
+      14712,
+      5568,
+      0,
+      5567,
+      0,
+      "MODEL"
+    ],
+    [
+      14713,
+      5566,
+      0,
+      5568,
+      0,
+      "MODEL"
+    ],
+    [
+      14714,
+      5553,
+      6,
+      5559,
+      0,
+      "VIDEO"
+    ],
+    [
+      14719,
+      5565,
+      0,
+      5571,
+      1,
+      "VAE"
+    ],
+    [
+      14720,
+      5571,
+      0,
+      5572,
+      0,
+      "IMAGE"
+    ],
+    [
+      14721,
+      5553,
+      7,
+      5572,
+      1,
+      "INT"
+    ],
+    [
+      14722,
+      5553,
+      8,
+      5572,
+      2,
+      "INT"
+    ],
+    [
+      14727,
+      5547,
+      1,
+      5578,
+      1,
+      "LATENT"
+    ],
+    [
+      14728,
+      5568,
+      0,
+      5575,
+      0,
+      "MODEL"
+    ],
+    [
+      14729,
+      5536,
+      0,
+      5575,
+      1,
+      "CONDITIONING"
+    ],
+    [
+      14730,
+      5536,
+      1,
+      5575,
+      2,
+      "CONDITIONING"
+    ],
+    [
+      14731,
+      5574,
+      0,
+      5579,
+      0,
+      "NOISE"
+    ],
+    [
+      14732,
+      5575,
+      0,
+      5579,
+      1,
+      "GUIDER"
+    ],
+    [
+      14733,
+      5576,
+      0,
+      5579,
+      2,
+      "SAMPLER"
+    ],
+    [
+      14735,
+      5578,
+      0,
+      5579,
+      4,
+      "LATENT"
+    ],
+    [
+      14736,
+      5579,
+      0,
+      5580,
+      0,
+      "LATENT"
+    ],
+    [
+      14738,
+      5561,
+      0,
+      5558,
+      0,
+      "IMAGE"
+    ],
+    [
+      14739,
+      5558,
+      0,
+      5573,
+      0,
+      "*"
+    ],
+    [
+      14741,
+      5573,
+      0,
+      5539,
+      0,
+      "*"
+    ],
+    [
+      14742,
+      5536,
+      2,
+      5581,
+      0,
+      "LATENT"
+    ],
+    [
+      14743,
+      5565,
+      0,
+      5581,
+      1,
+      "VAE"
+    ],
+    [
+      14744,
+      5581,
+      0,
+      5537,
+      0,
+      "IMAGE"
+    ],
+    [
+      14757,
+      5574,
+      0,
+      5543,
+      0,
+      "NOISE"
+    ],
+    [
+      14758,
+      5537,
+      0,
+      9040,
+      0,
+      "IMAGE"
+    ],
+    [
+      14759,
+      5553,
+      7,
+      9040,
+      1,
+      "INT"
+    ],
+    [
+      14760,
+      5553,
+      8,
+      9040,
+      2,
+      "INT"
+    ],
+    [
+      14761,
+      9040,
+      0,
+      9041,
+      0,
+      "IMAGE"
+    ],
+    [
+      14762,
+      5565,
+      0,
+      9041,
+      1,
+      "VAE"
+    ],
+    [
+      14763,
+      9041,
+      0,
+      9044,
+      0,
+      "*"
+    ],
+    [
+      14764,
+      5580,
+      0,
+      5571,
+      0,
+      "LATENT"
+    ],
+    [
+      14765,
+      5536,
+      2,
+      9043,
+      0,
+      "LATENT"
+    ],
+    [
+      14766,
+      9042,
+      0,
+      9043,
+      1,
+      "LATENT_UPSCALE_MODEL"
+    ],
+    [
+      14767,
+      5565,
+      0,
+      9043,
+      2,
+      "VAE"
+    ],
+    [
+      14769,
+      9043,
+      0,
+      9044,
+      1,
+      "*"
+    ],
+    [
+      14770,
+      9044,
+      0,
+      5578,
+      0,
+      "*"
+    ],
+    [
+      14771,
+      5566,
+      0,
+      9045,
+      0,
+      "MODEL"
+    ],
+    [
+      14773,
+      5533,
+      0,
+      5543,
+      3,
+      "SIGMAS"
+    ],
+    [
+      14774,
+      9045,
+      0,
+      5579,
+      3,
+      "SIGMAS"
+    ],
+    [
+      14775,
+      5572,
+      0,
+      9047,
+      0,
+      "IMAGE"
+    ],
+    [
+      14776,
+      9047,
+      0,
+      5573,
+      1,
+      "IMAGE"
+    ],
+    [
+      14777,
+      5553,
+      2,
+      9047,
+      1,
+      "MASK"
+    ]
+  ],
+  "groups": [
+    {
+      "id": 201,
+      "title": "Group",
+      "bounding": [
+        1458.2785201328422,
+        2848.0784825089822,
+        3944.615148267745,
+        4135.355437363503
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 202,
+      "title": "Set prompts     -> use API nodes to save on memory",
+      "bounding": [
+        4731.171036326067,
+        2915.351177250605,
+        605.366317151188,
+        2131.4844169185094
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 203,
+      "title": "Load Models",
+      "bounding": [
+        2172.8266537553436,
+        3623.491307185998,
+        1011.1492910007833,
+        544.0496711190453
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 208,
+      "title": "1. Reframe Source Video",
+      "bounding": [
+        1505.2090210219183,
+        2912.515412362328,
+        626.94336357631,
+        2135.610539796812
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 204,
+      "title": "Load Image",
+      "bounding": [
+        1551.8221498090081,
+        4541.245346832556,
+        538.3863234400123,
+        433.69496177544534
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 205,
+      "title": "Generate Low Resolution",
+      "bounding": [
+        2178.2659396318495,
+        4195.285580595269,
+        1001.5313918737324,
+        851.1188490429977
+      ],
+      "color": "#8AA",
+      "flags": {}
+    },
+    {
+      "id": 206,
+      "title": "2. Prepare Two-Stage Outpaint Conditions",
+      "bounding": [
+        2170.955648902097,
+        2912.5166137074807,
+        1014.8970403549247,
+        683.4223204333011
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 207,
+      "title": "Decode and Blend (STAGE 1)",
+      "bounding": [
+        2515.0491720527093,
+        3113.801213334338,
+        315.930959737791,
+        451.6717601420296
+      ],
+      "color": "#8AA",
+      "flags": {}
+    },
+    {
+      "id": 209,
+      "title": "Stage 2 Full-Resolution Mask",
+      "bounding": [
+        2861.7647267829916,
+        2959.8688774736343,
+        297.44324413688855,
+        612.3911306184623
+      ],
+      "color": "#b06634",
+      "flags": {}
+    },
+    {
+      "id": 210,
+      "title": "Stage 1 Half-Resolution Mask",
+      "bounding": [
+        2187.616467707624,
+        2965.230138881732,
+        296.76323989940556,
+        598.112162338969
+      ],
+      "color": "#8AA",
+      "flags": {}
+    },
+    {
+      "id": 211,
+      "title": "3. Stage 1 Direct Finish + Source Restoration",
+      "bounding": [
+        3218.23530977505,
+        2917.263795434737,
+        1460.9116896957812,
+        453.04834719718474
+      ],
+      "color": "#4f8a5b",
+      "flags": {}
+    },
+    {
+      "id": 212,
+      "title": "4. Stable Stage 2 · Optional Native 1.5× / 2× Upscale + 4-Step Full-Clip Synthesis",
+      "bounding": [
+        3220.4463588700673,
+        3405.5163745545105,
+        1458.2551524967785,
+        1393.321900353215
+      ],
+      "color": "#b06634",
+      "flags": {}
+    },
+    {
+      "id": 213,
+      "title": "REPAIR FISH-EYE LENS",
+      "bounding": [
+        4214.13605790708,
+        4317.499519863217,
+        445.1294100462537,
+        448.34089612390835
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    },
+    {
+      "id": 214,
+      "title": "Final Stage",
+      "bounding": [
+        3218.217838918279,
+        4818.880812191698,
+        1459.7007727389987,
+        228.48115549868362
+      ],
+      "color": "#3f789e",
+      "flags": {}
+    }
+  ],
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "c1aa79cc-b9f8-43f1-8d88-3d7406881f62",
+        "version": 1,
+        "state": {
+          "lastGroupId": 214,
+          "lastNodeId": 9050,
+          "lastLinkId": 14777,
+          "lastRerouteId": 228
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Video Stitch",
+        "description": "Stitches multiple video clips into a single sequential video file.",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            -6810,
+            2580,
+            143.55859375,
+            160
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            -4770,
+            2600,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "85555afe-c7a1-4f6e-b073-7c37f7bace7f",
+            "name": "video",
+            "type": "VIDEO",
+            "linkIds": [
+              253
+            ],
+            "localized_name": "video",
+            "label": "Before Video",
+            "pos": [
+              -6686.44140625,
+              2600
+            ]
+          },
+          {
+            "id": "022773ee-6b4f-4e3d-bead-68b3e75e2d20",
+            "name": "video_1",
+            "type": "VIDEO",
+            "linkIds": [
+              254
+            ],
+            "localized_name": "video_1",
+            "label": "After Video",
+            "pos": [
+              -6686.44140625,
+              2620
+            ]
+          },
+          {
+            "id": "7bcd7cbc-e918-472a-a0cf-2e0900545372",
+            "name": "direction",
+            "type": "COMBO",
+            "linkIds": [
+              259
+            ],
+            "pos": [
+              -6686.44140625,
+              2640
+            ]
+          },
+          {
+            "id": "9a00389d-c1c8-40d5-87fe-f41019b61fbc",
+            "name": "match_image_size",
+            "type": "BOOLEAN",
+            "linkIds": [
+              260
+            ],
+            "pos": [
+              -6686.44140625,
+              2660
+            ]
+          },
+          {
+            "id": "b95e0440-3ea8-4ae0-887e-12e75701042a",
+            "name": "spacing_width",
+            "type": "INT",
+            "linkIds": [
+              261
+            ],
+            "pos": [
+              -6686.44140625,
+              2680
+            ]
+          },
+          {
+            "id": "83ab9382-0a70-4169-b26a-66ab026b43c4",
+            "name": "spacing_color",
+            "type": "COMBO",
+            "linkIds": [
+              262
+            ],
+            "pos": [
+              -6686.44140625,
+              2700
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "09707f43-7552-4a6e-bd23-d962d31801c2",
+            "name": "VIDEO",
+            "type": "VIDEO",
+            "linkIds": [
+              255
+            ],
+            "localized_name": "VIDEO",
+            "pos": [
+              -4750,
+              2620
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 78,
+            "type": "GetVideoComponents",
+            "pos": [
+              -6390,
+              2600
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 254
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  249
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": null
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "GetVideoComponents"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 77,
+            "type": "GetVideoComponents",
+            "pos": [
+              -6390,
+              2420
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 253
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  248
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": [
+                  251
+                ]
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": [
+                  252
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "GetVideoComponents"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 90,
+            "type": "GetImageSize",
+            "pos": [
+              -6390,
+              3030
+            ],
+            "size": [
+              230,
+              120
+            ],
+            "flags": {},
+            "order": 4,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "link": 266
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "width",
+                "name": "width",
+                "type": "INT",
+                "links": [
+                  274
+                ]
+              },
+              {
+                "localized_name": "height",
+                "name": "height",
+                "type": "INT",
+                "links": [
+                  276
+                ]
+              },
+              {
+                "localized_name": "batch_size",
+                "name": "batch_size",
+                "type": "INT",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "GetImageSize"
+            },
+            "widgets_values": []
+          },
+          {
+            "id": 80,
+            "type": "CreateVideo",
+            "pos": [
+              -5190,
+              2420
+            ],
+            "size": [
+              270,
+              130
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 282
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": 251
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "fps"
+                },
+                "link": 252
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "VIDEO",
+                "name": "VIDEO",
+                "type": "VIDEO",
+                "links": [
+                  255
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "CreateVideo"
+            },
+            "widgets_values": [
+              30
+            ]
+          },
+          {
+            "id": 95,
+            "type": "ComfyMathExpression",
+            "pos": [
+              -6040,
+              3020
+            ],
+            "size": [
+              400,
+              200
+            ],
+            "flags": {},
+            "order": 5,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "a",
+                "localized_name": "values.a",
+                "name": "values.a",
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": 274
+              },
+              {
+                "label": "b",
+                "localized_name": "values.b",
+                "name": "values.b",
+                "shape": 7,
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": null
+              },
+              {
+                "localized_name": "INT",
+                "name": "INT",
+                "type": "INT",
+                "links": [
+                  279
+                ]
+              },
+              {
+                "localized_name": "BOOL",
+                "name": "BOOL",
+                "type": "BOOLEAN",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ComfyMathExpression"
+            },
+            "widgets_values": [
+              "a & ~1"
+            ]
+          },
+          {
+            "id": 96,
+            "type": "ComfyMathExpression",
+            "pos": [
+              -6040,
+              3290
+            ],
+            "size": [
+              400,
+              200
+            ],
+            "flags": {},
+            "order": 6,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "a",
+                "localized_name": "values.a",
+                "name": "values.a",
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": 276
+              },
+              {
+                "label": "b",
+                "localized_name": "values.b",
+                "name": "values.b",
+                "shape": 7,
+                "type": "FLOAT,INT,BOOLEAN",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": null
+              },
+              {
+                "localized_name": "INT",
+                "name": "INT",
+                "type": "INT",
+                "links": [
+                  280
+                ]
+              },
+              {
+                "localized_name": "BOOL",
+                "name": "BOOL",
+                "type": "BOOLEAN",
+                "links": null
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ComfyMathExpression"
+            },
+            "widgets_values": [
+              "a & ~1"
+            ]
+          },
+          {
+            "id": 79,
+            "type": "ImageStitch",
+            "pos": [
+              -6390,
+              2780
+            ],
+            "size": [
+              270,
+              160
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "image1",
+                "name": "image1",
+                "type": "IMAGE",
+                "link": 248
+              },
+              {
+                "localized_name": "image2",
+                "name": "image2",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": 249
+              },
+              {
+                "localized_name": "direction",
+                "name": "direction",
+                "type": "COMBO",
+                "widget": {
+                  "name": "direction"
+                },
+                "link": 259
+              },
+              {
+                "localized_name": "match_image_size",
+                "name": "match_image_size",
+                "type": "BOOLEAN",
+                "widget": {
+                  "name": "match_image_size"
+                },
+                "link": 260
+              },
+              {
+                "localized_name": "spacing_width",
+                "name": "spacing_width",
+                "type": "INT",
+                "widget": {
+                  "name": "spacing_width"
+                },
+                "link": 261
+              },
+              {
+                "localized_name": "spacing_color",
+                "name": "spacing_color",
+                "type": "COMBO",
+                "widget": {
+                  "name": "spacing_color"
+                },
+                "link": 262
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  266,
+                  281
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.13.0",
+              "Node name for S&R": "ImageStitch"
+            },
+            "widgets_values": [
+              "down",
+              true,
+              0,
+              "white"
+            ]
+          },
+          {
+            "id": 97,
+            "type": "ResizeImageMaskNode",
+            "pos": [
+              -5560,
+              2790
+            ],
+            "size": [
+              270,
+              154
+            ],
+            "flags": {},
+            "order": 7,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "input",
+                "name": "input",
+                "type": "IMAGE,MASK",
+                "link": 281
+              },
+              {
+                "localized_name": "width",
+                "name": "resize_type.width",
+                "type": "INT",
+                "widget": {
+                  "name": "resize_type.width"
+                },
+                "link": 279
+              },
+              {
+                "localized_name": "height",
+                "name": "resize_type.height",
+                "type": "INT",
+                "widget": {
+                  "name": "resize_type.height"
+                },
+                "link": 280
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "resized",
+                "name": "resized",
+                "type": "IMAGE",
+                "links": [
+                  282
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.24.0",
+              "Node name for S&R": "ResizeImageMaskNode"
+            },
+            "widgets_values": [
+              "scale dimensions",
+              512,
+              512,
+              "center",
+              "area"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 248,
+            "origin_id": 77,
+            "origin_slot": 0,
+            "target_id": 79,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 249,
+            "origin_id": 78,
+            "origin_slot": 0,
+            "target_id": 79,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 251,
+            "origin_id": 77,
+            "origin_slot": 1,
+            "target_id": 80,
+            "target_slot": 1,
+            "type": "AUDIO"
+          },
+          {
+            "id": 252,
+            "origin_id": 77,
+            "origin_slot": 2,
+            "target_id": 80,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 253,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 77,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 254,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 78,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 255,
+            "origin_id": 80,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 259,
+            "origin_id": -10,
+            "origin_slot": 2,
+            "target_id": 79,
+            "target_slot": 2,
+            "type": "COMBO"
+          },
+          {
+            "id": 260,
+            "origin_id": -10,
+            "origin_slot": 3,
+            "target_id": 79,
+            "target_slot": 3,
+            "type": "BOOLEAN"
+          },
+          {
+            "id": 261,
+            "origin_id": -10,
+            "origin_slot": 4,
+            "target_id": 79,
+            "target_slot": 4,
+            "type": "INT"
+          },
+          {
+            "id": 262,
+            "origin_id": -10,
+            "origin_slot": 5,
+            "target_id": 79,
+            "target_slot": 5,
+            "type": "COMBO"
+          },
+          {
+            "id": 266,
+            "origin_id": 79,
+            "origin_slot": 0,
+            "target_id": 90,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 274,
+            "origin_id": 90,
+            "origin_slot": 0,
+            "target_id": 95,
+            "target_slot": 0,
+            "type": "INT"
+          },
+          {
+            "id": 276,
+            "origin_id": 90,
+            "origin_slot": 1,
+            "target_id": 96,
+            "target_slot": 0,
+            "type": "INT"
+          },
+          {
+            "id": 279,
+            "origin_id": 95,
+            "origin_slot": 1,
+            "target_id": 97,
+            "target_slot": 1,
+            "type": "INT"
+          },
+          {
+            "id": 280,
+            "origin_id": 96,
+            "origin_slot": 1,
+            "target_id": 97,
+            "target_slot": 2,
+            "type": "INT"
+          },
+          {
+            "id": 281,
+            "origin_id": 79,
+            "origin_slot": 0,
+            "target_id": 97,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 282,
+            "origin_id": 97,
+            "origin_slot": 0,
+            "target_id": 80,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        }
+      }
+    ]
+  },
+  "config": {},
+  "extra": {
+    "ds": {
+      "scale": 0.19277164471476788,
+      "offset": [
+        1092.6567032383577,
+        -2142.8455922547687
+      ]
+    },
+    "reroutes": [
+      {
+        "id": 227,
+        "pos": [
+          8184.132758122472,
+          4913.258187542512
+        ],
+        "linkIds": [
+          14655
+        ]
+      },
+      {
+        "id": 228,
+        "parentId": 227,
+        "pos": [
+          8591.204920209277,
+          4924.530979220256
+        ],
+        "linkIds": [
+          14655
+        ]
+      }
+    ],
+    "frontendVersion": "1.45.15",
+    "VHS_latentpreview": false,
+    "VHS_latentpreviewrate": 0,
+    "VHS_MetadataImage": true,
+    "VHS_KeepIntermediate": true,
+    "linkExtensions": [
+      {
+        "id": 14655,
+        "parentId": 228
+      }
+    ]
+  },
+  "version": 0.4
+}
\ No newline at end of file
diff --git a/ltx_reframe.py b/ltx_reframe.py
new file mode 100644
index 0000000..2b99fa8
--- /dev/null
+++ b/ltx_reframe.py
@@ -0,0 +1,1049 @@
+# Copyright (c) 2026 exportAnything. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+"""Interactive source-video layout preparation for LTX outpainting workflows."""
+
+from __future__ import annotations
+
+import json
+import hashlib
+import math
+import os
+from dataclasses import asdict, dataclass
+from fractions import Fraction
+from typing import Any
+
+import folder_paths
+import torch
+from aiohttp import web
+from comfy_api.latest import InputImpl, Types
+from comfy_extras.nodes_audio import load as load_audio_file
+import comfy.utils
+
+
+CATEGORY = "Lightricks/video"
+ASPECT_PRESETS = ["16:9", "9:16", "1:1", "4:5", "5:4", "4:3", "3:4", "21:9", "Custom"]
+DIMENSION_MULTIPLE = 64
+MIN_SOURCE_SCALE = 0.1
+MAX_SOURCE_SCALE = 4.0
+SILENT_AUDIO_SAMPLE_RATE = 44100
+SILENT_AUDIO_CHANNELS = 2
+
+
+@dataclass(frozen=True)
+class ReframePlacement:
+    target_width: int
+    target_height: int
+    source_width: int
+    source_height: int
+    box_x: int
+    box_y: int
+    box_width: int
+    box_height: int
+    position_x: float
+    position_y: float
+    source_scale: float
+
+
+@dataclass(frozen=True)
+class ReframeFrameRange:
+    range_enabled: bool
+    source_frame_count: int
+    requested_first_frame: int
+    requested_last_frame: int
+    effective_first_frame: int
+    effective_last_frame: int
+    selected_frame_count: int
+    start_time: float
+    duration: float
+
+
+def _clamp(value: float, low: float, high: float) -> float:
+    return max(low, min(high, float(value)))
+
+
+def snap_dimension(value: int, multiple: int = DIMENSION_MULTIPLE) -> int:
+    """Snap a target dimension to an LTX-friendly multiple."""
+
+    value = max(256, min(8192, int(value)))
+    return max(multiple, int(math.floor((value + multiple / 2) / multiple) * multiple))
+
+
+def resolve_frame_range(
+    source_frame_count: int,
+    fps: float,
+    limit_frames: bool,
+    first_frame: int,
+    last_frame: int,
+) -> ReframeFrameRange:
+    """Resolve an inclusive UI range to an LTX-compatible ``8n+1`` window."""
+
+    source_frame_count = int(source_frame_count)
+    fps = float(fps)
+    if source_frame_count < 1:
+        raise ValueError("The selected video does not contain any frames.")
+    if not math.isfinite(fps) or fps <= 0:
+        raise ValueError("The selected video does not have a valid frame rate.")
+
+    if not bool(limit_frames):
+        selected_frame_count = source_frame_count
+        return ReframeFrameRange(
+            range_enabled=False,
+            source_frame_count=source_frame_count,
+            requested_first_frame=0,
+            requested_last_frame=source_frame_count - 1,
+            effective_first_frame=0,
+            effective_last_frame=source_frame_count - 1,
+            selected_frame_count=selected_frame_count,
+            start_time=0.0,
+            duration=selected_frame_count / fps,
+        )
+
+    requested_first = int(first_frame)
+    requested_last = source_frame_count - 1 if int(last_frame) == -1 else int(last_frame)
+    if requested_first < 0 or requested_first >= source_frame_count:
+        raise ValueError(
+            f"First frame {requested_first} is outside this video's 0-{source_frame_count - 1} range."
+        )
+    if requested_last < 0 or requested_last >= source_frame_count:
+        raise ValueError(
+            f"Last frame {requested_last} is outside this video's 0-{source_frame_count - 1} range."
+        )
+    if requested_last < requested_first:
+        raise ValueError("Last frame must be greater than or equal to first frame.")
+
+    requested_count = requested_last - requested_first + 1
+    if 2 <= requested_count <= 8:
+        raise ValueError(
+            "LTX frame-limited video ranges must contain one frame or at least nine frames."
+        )
+    selected_frame_count = (
+        1 if requested_count == 1 else ((requested_count - 1) // 8) * 8 + 1
+    )
+    effective_last = requested_first + selected_frame_count - 1
+    return ReframeFrameRange(
+        range_enabled=True,
+        source_frame_count=source_frame_count,
+        requested_first_frame=requested_first,
+        requested_last_frame=requested_last,
+        effective_first_frame=requested_first,
+        effective_last_frame=effective_last,
+        selected_frame_count=selected_frame_count,
+        start_time=requested_first / fps,
+        duration=selected_frame_count / fps,
+    )
+
+
+def ensure_audio_track(
+    audio: Any,
+    frame_count: int,
+    fps: float,
+    *,
+    sample_rate: int = SILENT_AUDIO_SAMPLE_RATE,
+    channels: int = SILENT_AUDIO_CHANNELS,
+) -> tuple[Any, bool]:
+    """Return source audio, or duration-matched silence for an audio-less video."""
+
+    if audio is not None:
+        return audio, False
+
+    frame_count = max(1, int(frame_count))
+    fps = float(fps)
+    if not math.isfinite(fps) or fps <= 0:
+        raise ValueError("Cannot synthesize fallback audio without a valid video frame rate.")
+
+    sample_rate = max(1, int(sample_rate))
+    channels = max(1, min(2, int(channels)))
+    duration = frame_count / fps
+    num_samples = max(1, int(round(duration * sample_rate)))
+    return (
+        {
+            "waveform": torch.zeros((1, channels, num_samples), dtype=torch.float32),
+            "sample_rate": sample_rate,
+        },
+        True,
+    )
+
+
+def fit_audio_track(
+    audio: Any,
+    frame_count: int,
+    fps: float,
+    *,
+    sample_rate: int = SILENT_AUDIO_SAMPLE_RATE,
+    channels: int = SILENT_AUDIO_CHANNELS,
+) -> tuple[Any, bool]:
+    """Crop or right-pad audio to the exact selected video duration."""
+
+    if audio is None:
+        return ensure_audio_track(
+            None,
+            frame_count,
+            fps,
+            sample_rate=sample_rate,
+            channels=channels,
+        )
+
+    waveform = audio.get("waveform")
+    source_sample_rate = int(audio.get("sample_rate", 0))
+    if not isinstance(waveform, torch.Tensor) or waveform.ndim != 3:
+        raise ValueError("Source audio must contain a [batch, channels, samples] waveform.")
+    if source_sample_rate <= 0:
+        raise ValueError("Source audio must have a valid sample rate.")
+
+    target_samples = max(1, int(round((int(frame_count) / float(fps)) * source_sample_rate)))
+    if waveform.shape[-1] < target_samples:
+        waveform = torch.nn.functional.pad(waveform, (0, target_samples - waveform.shape[-1]))
+    else:
+        waveform = waveform[..., :target_samples]
+    return {"waveform": waveform, "sample_rate": source_sample_rate}, False
+
+
+def _resolve_input_audio_path(audio_file: str) -> str:
+    """Resolve a custom audio selection strictly inside ComfyUI's input folder."""
+
+    selection = str(audio_file or "").strip()
+    if not selection:
+        raise ValueError(
+            "Custom audio is enabled, but no custom audio file is selected. "
+            "Upload or select an audio file in LTX Reframe Custom Audio Loader."
+        )
+
+    relative_name, annotated_base = folder_paths.annotated_filepath(selection)
+    input_dir = os.path.realpath(folder_paths.get_input_directory())
+    if annotated_base is not None and os.path.realpath(annotated_base) != input_dir:
+        raise ValueError("Custom audio must be selected from ComfyUI's input directory.")
+
+    audio_path = os.path.realpath(os.path.join(input_dir, relative_name))
+    try:
+        inside_input = os.path.commonpath((input_dir, audio_path)) == input_dir
+    except ValueError:
+        inside_input = False
+    if not inside_input:
+        raise ValueError("Custom audio must be selected from ComfyUI's input directory.")
+    if not os.path.isfile(audio_path):
+        raise ValueError(f"Custom audio file was not found: {relative_name}")
+    return audio_path
+
+
+def _sha256_file(path: str) -> str:
+    digest = hashlib.sha256()
+    with open(path, "rb") as handle:
+        for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+class LTXReframeCustomAudioLoader:
+    """Lazy-safe custom audio loader for optional reframe workflow branches."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_dir = folder_paths.get_input_directory()
+        os.makedirs(input_dir, exist_ok=True)
+        files = folder_paths.filter_files_content_types(
+            os.listdir(input_dir),
+            ["audio", "video"],
+        )
+        return {
+            "required": {
+                "audio_file": (
+                    [""] + sorted(files),
+                    {"audio_upload": True},
+                ),
+            },
+        }
+
+    RETURN_TYPES = ("AUDIO",)
+    RETURN_NAMES = ("audio",)
+    FUNCTION = "load_audio"
+    CATEGORY = "Lightricks/audio"
+    DESCRIPTION = (
+        "Loads custom audio only when its lazy workflow branch is selected. "
+        "An empty selection remains valid while source-audio mode is active."
+    )
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, audio_file):
+        # ComfyUI validates both sides of a lazy switch. File validation is
+        # intentionally deferred until this branch actually executes.
+        return True
+
+    @classmethod
+    def IS_CHANGED(cls, audio_file):
+        try:
+            return _sha256_file(_resolve_input_audio_path(audio_file))
+        except (OSError, ValueError):
+            return f"unselected:{str(audio_file or '').strip()}"
+
+    def load_audio(self, audio_file):
+        audio_path = _resolve_input_audio_path(audio_file)
+        try:
+            waveform, sample_rate = load_audio_file(audio_path)
+        except (OSError, RuntimeError, ValueError) as error:
+            raise ValueError(f"Custom audio could not be decoded: {error}") from error
+        return ({"waveform": waveform.unsqueeze(0), "sample_rate": sample_rate},)
+
+
+def _resolve_input_video_path(source_file: str) -> str:
+    """Resolve a source name while preventing metadata reads outside ComfyUI input."""
+
+    relative_name, annotated_base = folder_paths.annotated_filepath(str(source_file))
+    input_dir = os.path.realpath(folder_paths.get_input_directory())
+    if annotated_base is not None and os.path.realpath(annotated_base) != input_dir:
+        raise ValueError("Only ComfyUI input videos may be inspected.")
+    video_path = os.path.realpath(os.path.join(input_dir, relative_name))
+    try:
+        inside_input = os.path.commonpath((input_dir, video_path)) == input_dir
+    except ValueError:
+        inside_input = False
+    if not inside_input or not os.path.isfile(video_path):
+        raise ValueError("The selected input video does not exist.")
+    return video_path
+
+
+def probe_video_metadata(source_file: str) -> dict[str, Any]:
+    video_path = _resolve_input_video_path(source_file)
+    source_video = InputImpl.VideoFromFile(video_path)
+    width, height = source_video.get_dimensions()
+    fps = float(source_video.get_frame_rate())
+    frame_count = int(source_video.get_frame_count())
+    duration = float(source_video.get_duration())
+    return {
+        "source_file": source_file,
+        "width": int(width),
+        "height": int(height),
+        "duration": duration,
+        "fps": fps,
+        "frame_count": frame_count,
+    }
+
+
+def compute_reframe_placement(
+    source_width: int,
+    source_height: int,
+    target_width: int,
+    target_height: int,
+    position_x: float,
+    position_y: float,
+    source_scale: float,
+) -> ReframePlacement:
+    """Resolve normalized UI controls into an aspect-locked pixel box."""
+
+    source_width = max(1, int(source_width))
+    source_height = max(1, int(source_height))
+    target_width = snap_dimension(target_width)
+    target_height = snap_dimension(target_height)
+    position_x = _clamp(position_x, 0.0, 1.0)
+    position_y = _clamp(position_y, 0.0, 1.0)
+    source_scale = _clamp(source_scale, MIN_SOURCE_SCALE, MAX_SOURCE_SCALE)
+
+    fit = min(target_width / source_width, target_height / source_height)
+    box_width = max(2, int(round(source_width * fit * source_scale)))
+    box_height = max(2, int(round(source_height * fit * source_scale)))
+    box_x = int(round(position_x * (target_width - box_width)))
+    box_y = int(round(position_y * (target_height - box_height)))
+
+    return ReframePlacement(
+        target_width=target_width,
+        target_height=target_height,
+        source_width=source_width,
+        source_height=source_height,
+        box_x=box_x,
+        box_y=box_y,
+        box_width=box_width,
+        box_height=box_height,
+        position_x=position_x,
+        position_y=position_y,
+        source_scale=source_scale,
+    )
+
+
+def build_reframe_masks(
+    placement: ReframePlacement,
+    feather: int,
+    *,
+    device: torch.device | str = "cpu",
+    dtype: torch.dtype = torch.float32,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Return an LTX outpaint mask and a binary source-preservation mask."""
+
+    height = placement.target_height
+    width = placement.target_width
+    x0 = placement.box_x
+    y0 = placement.box_y
+    x1 = x0 + placement.box_width
+    y1 = y0 + placement.box_height
+
+    source_mask = torch.zeros((1, height, width), dtype=dtype, device=device)
+    visible_x0 = max(0, x0)
+    visible_y0 = max(0, y0)
+    visible_x1 = min(width, x1)
+    visible_y1 = min(height, y1)
+    if visible_x1 > visible_x0 and visible_y1 > visible_y0:
+        source_mask[:, visible_y0:visible_y1, visible_x0:visible_x1] = 1.0
+
+    feather = max(0, int(feather))
+    if feather == 0:
+        return 1.0 - source_mask, source_mask
+
+    yy = torch.arange(height, dtype=dtype, device=device).view(height, 1)
+    xx = torch.arange(width, dtype=dtype, device=device).view(1, width)
+    dx = torch.maximum(torch.maximum(x0 - xx, xx - (x1 - 1)), torch.zeros_like(xx))
+    dy = torch.maximum(torch.maximum(y0 - yy, yy - (y1 - 1)), torch.zeros_like(yy))
+    distance = torch.sqrt(dx.square() + dy.square())
+    outpaint_mask = torch.clamp(distance / float(feather), 0.0, 1.0).unsqueeze(0)
+    return outpaint_mask, source_mask
+
+
+def prepare_reframe_canvas(
+    frames: torch.Tensor,
+    placement: ReframePlacement,
+) -> torch.Tensor:
+    """Sample the visible source region into the target canvas without distortion."""
+
+    if frames.ndim != 4 or frames.shape[0] == 0:
+        raise ValueError("The selected video did not decode any frames.")
+    if frames.shape[-1] < 3:
+        raise ValueError("Video frames must have at least three color channels.")
+
+    source = frames[..., :3]
+    canvas = torch.zeros(
+        (
+            source.shape[0],
+            placement.target_height,
+            placement.target_width,
+            3,
+        ),
+        dtype=source.dtype,
+        device=source.device,
+    )
+
+    if (
+        placement.box_x >= 0
+        and placement.box_y >= 0
+        and placement.box_x + placement.box_width <= placement.target_width
+        and placement.box_y + placement.box_height <= placement.target_height
+    ):
+        resized = torch.nn.functional.interpolate(
+            source.movedim(-1, 1),
+            size=(placement.box_height, placement.box_width),
+            mode="bilinear",
+            align_corners=False,
+            antialias=True,
+        ).movedim(1, -1)
+        canvas[
+            :,
+            placement.box_y : placement.box_y + placement.box_height,
+            placement.box_x : placement.box_x + placement.box_width,
+            :,
+        ] = resized
+        return canvas
+
+    x0 = max(0, placement.box_x)
+    y0 = max(0, placement.box_y)
+    x1 = min(placement.target_width, placement.box_x + placement.box_width)
+    y1 = min(placement.target_height, placement.box_y + placement.box_height)
+    if x1 <= x0 or y1 <= y0:
+        return canvas
+
+    # Sampling only the visible intersection prevents a zoomed-in placement from
+    # allocating a potentially enormous intermediate image before it is cropped.
+    destination_x = torch.arange(x0, x1, dtype=torch.float32, device=source.device)
+    destination_y = torch.arange(y0, y1, dtype=torch.float32, device=source.device)
+    grid_x = 2.0 * (
+        (destination_x + 0.5 - placement.box_x) / float(placement.box_width)
+    ) - 1.0
+    grid_y = 2.0 * (
+        (destination_y + 0.5 - placement.box_y) / float(placement.box_height)
+    ) - 1.0
+    yy, xx = torch.meshgrid(grid_y, grid_x, indexing="ij")
+    grid = torch.stack((xx, yy), dim=-1).unsqueeze(0).expand(source.shape[0], -1, -1, -1)
+    sampled = torch.nn.functional.grid_sample(
+        source.movedim(-1, 1),
+        grid,
+        mode="bilinear",
+        padding_mode="border",
+        align_corners=False,
+    ).movedim(1, -1)
+    canvas[:, y0:y1, x0:x1, :] = sampled
+    return canvas
+
+
+def _normalize_source_mask(
+    source_mask: torch.Tensor,
+    *,
+    batch_size: int,
+    height: int,
+    width: int,
+    device: torch.device,
+) -> torch.Tensor:
+    """Normalize a ComfyUI MASK to ``[1|batch, 1, height, width]``."""
+
+    if not isinstance(source_mask, torch.Tensor):
+        raise ValueError("Source protection requires a MASK tensor.")
+    if source_mask.ndim == 2:
+        source_mask = source_mask.unsqueeze(0)
+    elif source_mask.ndim == 4:
+        if source_mask.shape[1] == 1:
+            source_mask = source_mask[:, 0]
+        elif source_mask.shape[-1] == 1:
+            source_mask = source_mask[..., 0]
+        else:
+            raise ValueError("Source protection MASK must have one channel.")
+    if source_mask.ndim != 3:
+        raise ValueError("Source protection MASK must have shape [batch, height, width].")
+    if source_mask.shape[0] not in (1, batch_size):
+        raise ValueError(
+            "Source protection MASK batch must contain one mask or match the IMAGE batch."
+        )
+
+    normalized = source_mask.to(device=device, dtype=torch.float32).unsqueeze(1)
+    if normalized.shape[-2:] != (height, width):
+        normalized = torch.nn.functional.interpolate(
+            normalized,
+            size=(height, width),
+            mode="bilinear",
+            align_corners=False,
+        )
+    return normalized.clamp(0.0, 1.0)
+
+
+def _feather_mask(mask: torch.Tensor, feather: int) -> torch.Tensor:
+    """Apply a separable Gaussian feather without allocating a 2-D kernel."""
+
+    radius = max(0, int(feather))
+    if radius == 0:
+        return mask
+
+    sigma = max(float(radius) / 3.0, 0.5)
+    coordinates = torch.arange(
+        -radius,
+        radius + 1,
+        device=mask.device,
+        dtype=torch.float32,
+    )
+    kernel = torch.exp(-(coordinates.square()) / (2.0 * sigma * sigma))
+    kernel = kernel / kernel.sum()
+    horizontal = kernel.view(1, 1, 1, -1)
+    vertical = kernel.view(1, 1, -1, 1)
+    blurred = torch.nn.functional.conv2d(
+        torch.nn.functional.pad(mask, (radius, radius, 0, 0), mode="replicate"),
+        horizontal,
+    )
+    blurred = torch.nn.functional.conv2d(
+        torch.nn.functional.pad(blurred, (0, 0, radius, radius), mode="replicate"),
+        vertical,
+    )
+    return blurred.clamp(0.0, 1.0)
+
+
+def rectilinear_correct_images(
+    images: torch.Tensor,
+    correction_strength: float,
+    center_x: float,
+    center_y: float,
+    protect_source: float,
+    protection_feather: int,
+    interpolation: str,
+    chunk_size: int,
+    source_mask: torch.Tensor | None = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Apply one temporally stable barrel-to-rectilinear remap to an IMAGE batch."""
+
+    if images.ndim != 4 or images.shape[0] == 0 or images.shape[-1] < 1:
+        raise ValueError("Rectilinear correction requires a non-empty IMAGE batch.")
+
+    batch_size, height, width, _channels = images.shape
+    correction_strength = _clamp(correction_strength, 0.0, 0.25)
+    center_x = _clamp(center_x, 0.0, 1.0)
+    center_y = _clamp(center_y, 0.0, 1.0)
+    protect_source = _clamp(protect_source, 0.0, 1.0)
+    chunk_size = max(1, min(64, int(chunk_size)))
+    interpolation = str(interpolation).lower()
+    if interpolation not in {"bilinear", "bicubic"}:
+        raise ValueError("Rectilinear correction interpolation must be bilinear or bicubic.")
+
+    device = images.device
+    center_pixel_x = center_x * max(0, width - 1)
+    center_pixel_y = center_y * max(0, height - 1)
+    yy, xx = torch.meshgrid(
+        torch.arange(height, device=device, dtype=torch.float32),
+        torch.arange(width, device=device, dtype=torch.float32),
+        indexing="ij",
+    )
+    delta_x = xx - center_pixel_x
+    delta_y = yy - center_pixel_y
+    corner_distances = torch.tensor(
+        [
+            center_pixel_x**2 + center_pixel_y**2,
+            (width - 1 - center_pixel_x) ** 2 + center_pixel_y**2,
+            center_pixel_x**2 + (height - 1 - center_pixel_y) ** 2,
+            (width - 1 - center_pixel_x) ** 2
+            + (height - 1 - center_pixel_y) ** 2,
+        ],
+        device=device,
+        dtype=torch.float32,
+    )
+    maximum_radius_squared = corner_distances.max().clamp_min(1.0)
+    radial_weight = ((delta_x.square() + delta_y.square()) / maximum_radius_squared).clamp(
+        0.0, 1.0
+    )
+
+    # Positive strength samples progressively nearer the optical center. This
+    # counteracts barrel/fisheye expansion while naturally cropping the outermost
+    # pixels, so the remap never creates empty borders.
+    radial_scale = 1.0 - correction_strength * radial_weight
+    sample_x = center_pixel_x + delta_x * radial_scale
+    sample_y = center_pixel_y + delta_y * radial_scale
+    grid_x = (2.0 * (sample_x + 0.5) / float(width)) - 1.0
+    grid_y = (2.0 * (sample_y + 0.5) / float(height)) - 1.0
+    grid = torch.stack((grid_x, grid_y), dim=-1).unsqueeze(0)
+
+    protection = None
+    if source_mask is not None and protect_source > 0.0:
+        protection = _normalize_source_mask(
+            source_mask,
+            batch_size=batch_size,
+            height=height,
+            width=width,
+            device=device,
+        )
+        protection = _feather_mask(protection, protection_feather) * protect_source
+
+    correction_mask = radial_weight.unsqueeze(0)
+    if protection is not None:
+        correction_mask = correction_mask * (1.0 - protection[:, 0])
+    if correction_strength == 0.0:
+        correction_mask = torch.zeros_like(correction_mask)
+
+    source = images.movedim(-1, 1)
+    corrected_chunks: list[torch.Tensor] = []
+    for start in range(0, batch_size, chunk_size):
+        end = min(batch_size, start + chunk_size)
+        source_chunk = source[start:end].to(dtype=torch.float32)
+        corrected_chunk = torch.nn.functional.grid_sample(
+            source_chunk,
+            grid.expand(end - start, -1, -1, -1),
+            mode=interpolation,
+            padding_mode="border",
+            align_corners=False,
+        )
+        if protection is not None:
+            protection_chunk = protection if protection.shape[0] == 1 else protection[start:end]
+            corrected_chunk = (
+                corrected_chunk * (1.0 - protection_chunk)
+                + source_chunk * protection_chunk
+            )
+        corrected_chunks.append(corrected_chunk.to(dtype=images.dtype).movedim(1, -1))
+
+    corrected = torch.cat(corrected_chunks, dim=0).clamp(0.0, 1.0)
+    return corrected, correction_mask.to(dtype=images.dtype)
+
+
+class LTXReframeLayout:
+    """Upload, position, and mask a source video for downstream LTX outpainting."""
+
+    DESCRIPTION = (
+        "Preview, zoom, and place a source video on an LTX-compatible canvas. "
+        "Outputs deterministic padded frames and masks; connect them to "
+        "LTXVInpaintPreprocess and the LTX IC-LoRA outpaint workflow."
+    )
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        input_dir = folder_paths.get_input_directory()
+        files = [
+            name
+            for name in os.listdir(input_dir)
+            if os.path.isfile(os.path.join(input_dir, name))
+        ]
+        files = folder_paths.filter_files_content_types(files, ["video"])
+        return {
+            "required": {
+                "source_file": (sorted(files), {"video_upload": True}),
+                "aspect_ratio": (ASPECT_PRESETS, {"default": "9:16"}),
+                "target_width": (
+                    "INT",
+                    {"default": 1088, "min": 256, "max": 8192, "step": DIMENSION_MULTIPLE},
+                ),
+                "target_height": (
+                    "INT",
+                    {"default": 1920, "min": 256, "max": 8192, "step": DIMENSION_MULTIPLE},
+                ),
+                "position_x": (
+                    "FLOAT",
+                    {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.001, "round": 0.001},
+                ),
+                "position_y": (
+                    "FLOAT",
+                    {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.001, "round": 0.001},
+                ),
+                "source_scale": (
+                    "FLOAT",
+                    {
+                        "default": 0.7,
+                        "min": MIN_SOURCE_SCALE,
+                        "max": MAX_SOURCE_SCALE,
+                        "step": 0.001,
+                        "round": 0.001,
+                    },
+                ),
+                "feather": (
+                    "INT",
+                    {"default": 32, "min": 0, "max": 512, "step": 1},
+                ),
+            },
+            "optional": {
+                "limit_frames": ("BOOLEAN", {"default": False}),
+                "first_frame": (
+                    "INT",
+                    {"default": 0, "min": 0, "max": 2147483647, "step": 1},
+                ),
+                "last_frame": (
+                    "INT",
+                    {"default": -1, "min": -1, "max": 2147483647, "step": 1},
+                ),
+            },
+        }
+
+    RETURN_TYPES = (
+        "IMAGE",
+        "MASK",
+        "MASK",
+        "IMAGE",
+        "AUDIO",
+        "FLOAT",
+        "VIDEO",
+        "INT",
+        "INT",
+        "STRING",
+    )
+    RETURN_NAMES = (
+        "padded_frames",
+        "outpaint_mask",
+        "source_mask",
+        "source_frames",
+        "audio",
+        "fps",
+        "source_video",
+        "target_width",
+        "target_height",
+        "placement_json",
+    )
+    FUNCTION = "prepare"
+    CATEGORY = CATEGORY
+
+    def prepare(
+        self,
+        source_file: str,
+        aspect_ratio: str,
+        target_width: int,
+        target_height: int,
+        position_x: float,
+        position_y: float,
+        source_scale: float,
+        feather: int,
+        limit_frames: bool = False,
+        first_frame: int = 0,
+        last_frame: int = -1,
+    ) -> tuple[Any, ...]:
+        video_path = _resolve_input_video_path(source_file)
+        original_source_video = InputImpl.VideoFromFile(video_path)
+        source_width, source_height = original_source_video.get_dimensions()
+        fps_fraction = Fraction(original_source_video.get_frame_rate())
+        fps = float(fps_fraction)
+        frame_range = resolve_frame_range(
+            original_source_video.get_frame_count(),
+            fps,
+            limit_frames,
+            first_frame,
+            last_frame,
+        )
+
+        if frame_range.range_enabled:
+            decode_source = InputImpl.VideoFromFile(
+                video_path,
+                start_time=frame_range.start_time,
+                duration=frame_range.duration,
+            )
+            decoded = decode_source.get_components()
+            decoded_count = int(decoded.images.shape[0])
+            if decoded_count < frame_range.selected_frame_count:
+                raise ValueError(
+                    "The selected video range decoded fewer frames than expected "
+                    f"({decoded_count} of {frame_range.selected_frame_count})."
+                )
+            images = decoded.images[: frame_range.selected_frame_count]
+            alpha = (
+                decoded.alpha[: frame_range.selected_frame_count]
+                if decoded.alpha is not None
+                else None
+            )
+            audio, synthesized_silent_audio = fit_audio_track(
+                decoded.audio,
+                frame_range.selected_frame_count,
+                fps,
+            )
+            components = Types.VideoComponents(
+                images=images,
+                alpha=alpha,
+                audio=audio,
+                frame_rate=fps_fraction,
+                metadata=decoded.metadata,
+            )
+            source_video = InputImpl.VideoFromComponents(components)
+        else:
+            components = original_source_video.get_components()
+            audio, synthesized_silent_audio = ensure_audio_track(
+                components.audio,
+                components.images.shape[0],
+                components.frame_rate,
+            )
+            source_video = original_source_video
+
+        placement = compute_reframe_placement(
+            source_width,
+            source_height,
+            target_width,
+            target_height,
+            position_x,
+            position_y,
+            source_scale,
+        )
+        padded_frames = prepare_reframe_canvas(components.images, placement)
+        outpaint_mask, source_mask = build_reframe_masks(
+            placement,
+            feather,
+            device=padded_frames.device,
+            dtype=padded_frames.dtype,
+        )
+        metadata = {
+            "schema": "ltx_reframe_layout/v2",
+            "source_file": source_file,
+            "aspect_ratio_preset": aspect_ratio,
+            "feather": int(feather),
+            "frame_count": int(components.images.shape[0]),
+            "fps": float(components.frame_rate),
+            "source_frame_count": frame_range.source_frame_count,
+            "range_enabled": frame_range.range_enabled,
+            "requested_first_frame": frame_range.requested_first_frame,
+            "requested_last_frame": frame_range.requested_last_frame,
+            "effective_first_frame": frame_range.effective_first_frame,
+            "effective_last_frame": frame_range.effective_last_frame,
+            "selected_frame_count": frame_range.selected_frame_count,
+            "start_time": frame_range.start_time,
+            "duration": frame_range.duration,
+            "source_has_audio": not synthesized_silent_audio,
+            "audio_fallback": "duration_matched_silence" if synthesized_silent_audio else "none",
+            "placement": asdict(placement),
+        }
+        return (
+            padded_frames,
+            outpaint_mask,
+            source_mask,
+            components.images,
+            audio,
+            float(components.frame_rate),
+            source_video,
+            placement.target_width,
+            placement.target_height,
+            json.dumps(metadata, indent=2, sort_keys=True),
+        )
+
+    @classmethod
+    def IS_CHANGED(cls, source_file: str, **_kwargs):
+        video_path = _resolve_input_video_path(source_file)
+        return os.path.getmtime(video_path)
+
+    @classmethod
+    def VALIDATE_INPUTS(cls, source_file: str, **_kwargs):
+        try:
+            _resolve_input_video_path(source_file)
+        except ValueError as error:
+            return f"Invalid video file: {source_file}. {error}"
+        return True
+
+
+class LTXReframeTargetFit:
+    """Center-fit an upscaled frame batch to the reframe node's exact canvas."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "target_width": ("INT", {"forceInput": True}),
+                "target_height": ("INT", {"forceInput": True}),
+            }
+        }
+
+    RETURN_TYPES = ("IMAGE",)
+    RETURN_NAMES = ("images",)
+    FUNCTION = "fit"
+    CATEGORY = CATEGORY
+    DESCRIPTION = (
+        "Center-crop and Lanczos-resize a learned LTX upscale to the exact "
+        "canvas dimensions produced by LTX Reframe Layout."
+    )
+
+    def fit(
+        self,
+        images: torch.Tensor,
+        target_width: int,
+        target_height: int,
+    ) -> tuple[torch.Tensor]:
+        if images.ndim != 4 or images.shape[0] == 0 or images.shape[-1] < 3:
+            raise ValueError("LTX Reframe Target Fit requires a non-empty IMAGE batch.")
+        target_width = max(1, int(target_width))
+        target_height = max(1, int(target_height))
+        if images.shape[2] == target_width and images.shape[1] == target_height:
+            return (images,)
+        fitted = comfy.utils.common_upscale(
+            images.movedim(-1, 1),
+            target_width,
+            target_height,
+            "lanczos",
+            "center",
+        ).movedim(1, -1)
+        return (fitted.clamp(0.0, 1.0),)
+
+
+class LTXReframeRectilinearCorrection:
+    """Temporally stable, mask-aware correction for generated fisheye curvature."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "images": ("IMAGE",),
+                "enabled": ("BOOLEAN", {"default": True}),
+                "correction_strength": (
+                    "FLOAT",
+                    {
+                        "default": 0.05,
+                        "min": 0.0,
+                        "max": 0.25,
+                        "step": 0.005,
+                        "round": 0.005,
+                    },
+                ),
+                "center_x": (
+                    "FLOAT",
+                    {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01},
+                ),
+                "center_y": (
+                    "FLOAT",
+                    {"default": 0.5, "min": 0.0, "max": 1.0, "step": 0.01},
+                ),
+                "protect_source": (
+                    "FLOAT",
+                    {
+                        "default": 1.0,
+                        "min": 0.0,
+                        "max": 1.0,
+                        "step": 0.05,
+                        "round": 0.05,
+                    },
+                ),
+                "protection_feather": (
+                    "INT",
+                    {"default": 96, "min": 0, "max": 512, "step": 8},
+                ),
+                "interpolation": (["bicubic", "bilinear"], {"default": "bicubic"}),
+                "chunk_size": (
+                    "INT",
+                    {"default": 8, "min": 1, "max": 64, "step": 1},
+                ),
+            },
+            "optional": {
+                "source_mask": ("MASK",),
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE", "MASK")
+    RETURN_NAMES = ("images", "correction_mask")
+    FUNCTION = "correct"
+    CATEGORY = CATEGORY
+    DESCRIPTION = (
+        "Apply a deterministic barrel-to-rectilinear correction after the final "
+        "Stage 2 fit. Connect LTX Reframe Layout's source_mask to protect original "
+        "pixels. Positive strength removes fisheye curvature and crops edges "
+        "without creating empty borders. Start at 0.04-0.06."
+    )
+
+    def correct(
+        self,
+        images: torch.Tensor,
+        enabled: bool,
+        correction_strength: float,
+        center_x: float,
+        center_y: float,
+        protect_source: float,
+        protection_feather: int,
+        interpolation: str,
+        chunk_size: int,
+        source_mask: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if images.ndim != 4 or images.shape[0] == 0 or images.shape[-1] < 1:
+            raise ValueError("Rectilinear correction requires a non-empty IMAGE batch.")
+        if not bool(enabled):
+            return (
+                images,
+                torch.zeros(
+                    (1, images.shape[1], images.shape[2]),
+                    device=images.device,
+                    dtype=images.dtype,
+                ),
+            )
+        return rectilinear_correct_images(
+            images,
+            correction_strength,
+            center_x,
+            center_y,
+            protect_source,
+            protection_feather,
+            interpolation,
+            chunk_size,
+            source_mask,
+        )
+
+
+NODE_CLASS_MAPPINGS = {
+    "LTXReframeLayout": LTXReframeLayout,
+    "LTXReframeTargetFit": LTXReframeTargetFit,
+    "LTXReframeRectilinearCorrection": LTXReframeRectilinearCorrection,
+    "LTXReframeCustomAudioLoader": LTXReframeCustomAudioLoader,
+}
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "LTXReframeLayout": "LTX Reframe Layout",
+    "LTXReframeTargetFit": "LTX Reframe Target Fit",
+    "LTXReframeRectilinearCorrection": "LTX Reframe Rectilinear Correction",
+    "LTXReframeCustomAudioLoader": "LTX Reframe Custom Audio Loader",
+}
+
+
+def _register_video_metadata_route() -> None:
+    try:
+        from server import PromptServer
+    except ImportError:
+        return
+
+    prompt_server = getattr(PromptServer, "instance", None)
+    if prompt_server is None or getattr(prompt_server, "_ltx_reframe_metadata_route", False):
+        return
+
+    @prompt_server.routes.get("/ltx_reframe/video_metadata")
+    async def ltx_reframe_video_metadata(request):
+        source_file = request.query.get("source_file", "")
+        if not source_file:
+            return web.json_response({"error": "source_file is required"}, status=400)
+        try:
+            return web.json_response(probe_video_metadata(source_file))
+        except (OSError, RuntimeError, ValueError) as error:
+            return web.json_response({"error": str(error)}, status=400)
+
+    prompt_server._ltx_reframe_metadata_route = True
+
+
+_register_video_metadata_route()
diff --git a/nodes.py b/nodes.py
index bcd40b8..07bc894 100644
--- a/nodes.py
+++ b/nodes.py
@@ -3,6 +3,7 @@
 
 from __future__ import annotations
 
+import copy
 import gc
 import json
 import logging
@@ -13,6 +14,7 @@ import shlex
 import subprocess
 import threading
 import time
+import unicodedata
 from dataclasses import asdict, dataclass, field
 from pathlib import Path
 from typing import Any
@@ -21,6 +23,7 @@ from typing import Any
 CATEGORY = "prompt/diffusiongemma"
 COMPATIBILITY_CATEGORY = f"{CATEGORY}/compatibility"
 OPTIONAL_CATEGORY = f"{CATEGORY}/optional"
+EXPERIMENTAL_MOTION_CATEGORY = f"{CATEGORY}/experimental-motion-planning"
 MODEL_TYPE = "DG_MODEL_CONFIG"
 MEDIA_TYPE = "DG_MEDIA_CONTEXT"
 CONTEXT_TYPE = "DG_CONTEXT"
@@ -67,7 +70,7 @@ REFERENCE_PREP_ASPECT_CHOICES = ["source", "1:1", "16:9", "9:16", "4:3", "3:4",
 MEDIA_SYNTHESIS_MODE_CHOICES = ["video_recreation", "image_identity_video_control"]
 DEFAULT_MASTER_PROMPT = """You are a local ComfyUI prompt architect. Treat the user's text as a content brief: it only needs to say what the video or image should be about. Do not require the user to ask for prompt-writing boilerplate such as detailed subject, background, lighting, camera movement, pacing, composition, or style; infer useful generator-ready detail when it is missing, while preserving every explicit user constraint.
 
-Use the user's text, any backend-consumable pixels, and any supplied visual_description to produce final, practical prompts for generation models. If the active backend is metadata-only for image/video, do not pretend you saw pixels; use only the user's text, the visual_description, and factual media metadata. Preserve requested subjects, exact text, composition, motion, style, camera behavior, visual capture language, audio mode, audio guidance, and constraints. In image_identity_video_control workflows, treat a request to replace the video subject with the image subject as a strict identity-and-appearance transfer: face, hair, body type, wardrobe, accessories, styling, and distinguishing traits come from the reference image, while the video contributes pose, action, blocking, depth/canny/edge structure, camera choreography, timing, composition, and scene geometry. Do not copy the control-video subject's wardrobe, hair, face, body type, accessories, or styling into the final prompt unless the user explicitly asks for those video-subject traits. For LTX source-video work, explicitly observe camera choreography before writing the final prompt: first-frame and ending framing, shot size, camera height/angle, camera support or stability, camera path, zoom/dolly/truck/boom/pan/tilt/roll/orbit/rotation, subject-camera relationship, focus behavior, depth/parallax, and visible lens/capture changes. Transcribe observed camera work precisely instead of reducing it to generic phrases like camera movement, zoom out, or pan left. Visual capture language includes exposure level, overexposed or blown highlights, underexposure, flash/direct light, bloom, halation, lens blur, bokeh, depth of field, focal length impression, perspective compression, lens distortion, motion blur, grain/noise, white balance, contrast, dynamic range, and camera angle. Do not describe invisible intent. Every prompt field must contain the finished prompt text or finished prompt JSON, never instructions about how to write a prompt. Return only valid JSON with ltx_prompt, ideogram_prompt, negative_prompt, scene_segments, and metadata."""
+Use the user's text, any backend-consumable pixels, and any supplied visual_description to produce final, practical prompts for generation models. If the active backend is metadata-only for image/video, do not pretend you saw pixels; use only the user's text, the visual_description, and factual media metadata. Preserve requested subjects, exact text, composition, motion, style, camera behavior, visual capture language, audio mode, audio guidance, and constraints. In image_identity_video_control workflows, treat a request to replace the video subject with the image subject as a strict identity-and-appearance transfer: face, hair, body type, wardrobe, accessories, styling, and distinguishing traits come from the reference image, while the video contributes pose, action, blocking, depth/canny/edge structure, camera choreography, timing, composition, and scene geometry. Do not copy the control-video subject's wardrobe, hair, face, body type, accessories, or styling into the final prompt unless the user explicitly asks for those video-subject traits. For LTX source-video work, explicitly observe camera choreography before writing the final prompt: first-frame and ending framing, shot size, camera height/angle, camera support or stability, camera path, zoom/dolly/truck/boom/pan/tilt/roll/orbit/rotation, subject-camera relationship, focus behavior, depth/parallax, and visible lens/capture changes. Transcribe observed camera work precisely instead of reducing it to generic phrases like camera movement, zoom out, or pan left. Visual capture language includes exposure level, overexposed or blown highlights, underexposure, flash/direct light, bloom, halation, lens blur, bokeh, depth of field, focal length impression, perspective compression, lens distortion, motion blur, grain/noise, white balance, contrast, dynamic range, and camera angle. Do not describe invisible intent. Every prompt field must contain the finished prompt text or finished prompt JSON, never instructions about how to write a prompt. Return only valid JSON with ltx_prompt, ideogram_prompt, minimax_h3_prompt, negative_prompt, scene_segments, and metadata."""
 
 
 @dataclass
@@ -116,6 +119,7 @@ class TargetProfileConfig:
     ideogram_json_output: bool = True
     negative_prompt_mode: str = "auto"
     negative_prompt_guidance: str = ""
+    minimax_h3_mode: str = "t2va"
 
 
 _MODEL_CACHE: dict[str, tuple[Any, Any]] = {}
@@ -906,6 +910,582 @@ def _sanitize_prompt_text(
     return value
 
 
+def _sanitize_structured_prompt_text(
+    text: str,
+    fallback: str = "",
+    max_chars: int = 8000,
+    strip_thinking: bool = True,
+    strip_markdown: bool = True,
+) -> str:
+    value = str(text or "").replace("\r\n", "\n").replace("\r", "\n").strip()
+    if strip_thinking:
+        value = _strip_thinking(value)
+    if strip_markdown:
+        value = _strip_markdown(value)
+    replacements = {
+        "\ufeff": "",
+        "\ufffd": "",
+        "": "",
+    }
+    for artifact, replacement in replacements.items():
+        value = value.replace(artifact, replacement)
+    value = re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f-\x9f]", " ", value)
+    lines = []
+    for line in value.splitlines():
+        line = re.sub(r"[ \t]+", " ", line).strip()
+        lines.append(line)
+    value = "\n".join(lines)
+    value = re.sub(r"\n{3,}", "\n\n", value).strip()
+    if not value:
+        value = str(fallback or "").strip()
+    if max_chars > 0 and len(value) > max_chars:
+        value = value[:max_chars].rstrip()
+    return value
+
+
+_MINIMAX_H3_REF_FIELDS = (
+    "subject_definitions",
+    "summary",
+    "retention_analysis",
+    "detailed_description",
+    "overall_soundscape",
+    "non_diegetic_music",
+)
+
+
+def _minimax_h3_ref_sections(prompt: str) -> dict[str, str] | None:
+    pattern = r"(?ms)^" + r"\s*\n+\s*".join(
+        rf"{re.escape(field_name)}\s*:\s*(.*?)" for field_name in _MINIMAX_H3_REF_FIELDS[:-1]
+    ) + rf"\s*\n+\s*{re.escape(_MINIMAX_H3_REF_FIELDS[-1])}\s*:\s*(.*)$"
+    match = re.fullmatch(pattern, prompt)
+    if not match:
+        return None
+    return {
+        field_name: match.group(index + 1).strip()
+        for index, field_name in enumerate(_MINIMAX_H3_REF_FIELDS)
+    }
+
+
+def _rebuild_minimax_h3_ref_prompt(sections: dict[str, str]) -> str:
+    return "\n\n".join(f"{field_name}:\n{sections.get(field_name, '').strip()}" for field_name in _MINIMAX_H3_REF_FIELDS)
+
+
+def _normalize_minimax_h3_prompt(text: str, minimax_h3_mode: str = "t2va") -> str:
+    prompt = _sanitize_structured_prompt_text(text, "", 0)
+    if not prompt:
+        return ""
+
+    if _normalize_minimax_h3_mode(minimax_h3_mode) == "ref2va":
+        prompt = re.sub(
+            r"<\s*(subject|picture|image|video|audio)\s+(\d+)\s*>",
+            lambda match: (
+                f"<Subject {int(match.group(2))}>"
+                if match.group(1).casefold() == "subject"
+                else f"<{'Picture' if match.group(1).casefold() in {'picture', 'image'} else match.group(1).title()} {int(match.group(2))}>"
+            ),
+            prompt,
+            flags=re.IGNORECASE,
+        )
+        for field_name in _MINIMAX_H3_REF_FIELDS:
+            prompt = re.sub(
+                rf"(?im)^\s*{re.escape(field_name)}\s*:",
+                f"{field_name}:",
+                prompt,
+            )
+        sections = _minimax_h3_ref_sections(prompt)
+        return _rebuild_minimax_h3_ref_prompt(sections) if sections else prompt
+
+    fields = (
+        "integrated_multimodal_description",
+        "overall_soundscape",
+        "non_diegetic_music",
+    )
+    for field_name in fields:
+        prompt = re.sub(
+            rf"(?im)^\s*{re.escape(field_name)}\s*:",
+            f"{field_name}:",
+            prompt,
+        )
+
+    field_matches = {
+        field_name: re.search(rf"(?m)^{re.escape(field_name)}\s*:", prompt)
+        for field_name in fields
+    }
+    if not any(field_matches.values()) and re.match(r"^\[Shot 1\](?=\s|$)", prompt):
+        blocks = [block.strip() for block in re.split(r"\n\s*\n", prompt) if block.strip()]
+        if len(blocks) >= 3:
+            integrated = "\n\n".join(blocks[:-2])
+            prompt = (
+                f"integrated_multimodal_description: {integrated}\n\n"
+                f"overall_soundscape: {blocks[-2]}\n\n"
+                f"non_diegetic_music: {blocks[-1]}"
+            )
+    elif (
+        not field_matches["integrated_multimodal_description"]
+        and field_matches["overall_soundscape"]
+        and field_matches["non_diegetic_music"]
+        and re.match(r"^\[Shot 1\](?=\s|$)", prompt)
+    ):
+        prompt = f"integrated_multimodal_description: {prompt}"
+
+    structured_match = re.fullmatch(
+        r"(?ms)integrated_multimodal_description\s*:\s*(.*?)\s*\n+\s*overall_soundscape\s*:\s*(.*?)\s*\n+\s*non_diegetic_music\s*:\s*(.*)",
+        prompt,
+    )
+    if structured_match:
+        integrated = structured_match.group(1).strip()
+        shot_labels = re.findall(r"\[Shot (\d+)\]", integrated, flags=re.IGNORECASE)
+        if shot_labels == ["1"]:
+            next_shot = 1
+
+            def label_unmarked_cut(match: re.Match[str]) -> str:
+                nonlocal next_shot
+                next_shot += 1
+                return f"[Shot {next_shot}] {match.group(0)}"
+
+            integrated = re.sub(
+                r"\bAt\s+\d{2,}:\d{2}\.\d{3},?\s+(?=the\s+(?:camera|shot|scene)\s+(?:(?:hard\s+)?cuts|transitions|changes|switches|cross-dissolves|fades|wipes)\s+(?:to|into)\b)",
+                label_unmarked_cut,
+                integrated,
+                flags=re.IGNORECASE,
+            )
+        prompt = (
+            f"integrated_multimodal_description: {integrated}\n\n"
+            f"overall_soundscape: {structured_match.group(2).strip()}\n\n"
+            f"non_diegetic_music: {structured_match.group(3).strip()}"
+        )
+
+    return re.sub(
+        r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3})(?!,)(?=\s+the\s+(?:camera|shot|scene)\b)",
+        r"\1,",
+        prompt,
+        flags=re.IGNORECASE,
+    )
+
+
+def _minimax_h3_ref_definition_items(text: str) -> list[tuple[str, str]]:
+    return [
+        (f"<{kind.title()} {int(ordinal)}>", description.strip())
+        for kind, ordinal, description in re.findall(
+            r"(?im)^\s*<\s*(Subject|Picture|Video|Audio)\s+(\d+)\s*>\s*(?::|=|[-–—]|\bis\b)\s*(\S.*)$",
+            text or "",
+        )
+    ]
+
+
+def _minimax_h3_ref_summary_task_types(user_prompt: str, reference_manifest: str) -> list[str]:
+    manifest = _normalize_minimax_h3_reference_manifest(reference_manifest)
+    definitions = _minimax_h3_reference_definitions(manifest)
+    tags = [tag for tag, _description in definitions]
+    combined = f"{user_prompt} {manifest}"
+    task_types: list[str] = []
+
+    if _has_unnegated_prompt_phrase(
+        combined,
+        r"\b(?:keyframes?|first[-\s]+frames?|last[-\s]+frames?|first\s+and\s+last\s+frames?|interpolat(?:e|ion|ing))\b",
+    ):
+        task_types.append("keyframe completion")
+    if any(tag.startswith(("<Picture ", "<Video ")) for tag in tags):
+        task_types.append("reference generation")
+    if any(tag.startswith("<Video ") for tag in tags):
+        if re.search(r"\b(?:continu(?:e|ation)|extend(?:ing)?|append|carry\s+on)\b", combined, flags=re.IGNORECASE):
+            task_types.append("video continuation")
+        elif re.search(r"\b(?:edit|replace|remove|modify|change|transform|restyl(?:e|ing)|rework)\b", combined, flags=re.IGNORECASE):
+            task_types.append("video editing")
+    if any(tag.startswith("<Audio ") for tag in tags):
+        if re.search(r"\b(?:copy|reuse|retain|preserve|keep)\b[^.!?\n]{0,80}\b(?:audio|soundtrack|voice|music)\b", combined, flags=re.IGNORECASE):
+            task_types.append("audio reuse")
+        else:
+            task_types.append("audio reference")
+    return list(dict.fromkeys(task_types or ["reference generation"]))
+
+
+def _repair_minimax_h3_ref_detailed_timeline(detailed: str, duration_seconds: float = 0.0) -> str:
+    range_pattern = re.compile(
+        r"\[Shot (?P<shot>\d+)\]\s*\(?\s*"
+        r"(?P<start_major>\d{1,2}):(?P<start_minor>\d{2})(?:\.(?P<start_fraction>\d{1,3}))?\s*[-–—]\s*"
+        r"(?P<end_major>\d{1,2}):(?P<end_minor>\d{2})(?:\.(?P<end_fraction>\d{1,3}))?\s*\)?\s*:?\s*",
+        flags=re.IGNORECASE,
+    )
+    range_matches = list(range_pattern.finditer(detailed))
+    duration = max(0.0, _safe_float(duration_seconds, 0.0))
+
+    def range_time(match: re.Match[str], prefix: str, mode: str) -> float:
+        major = int(match.group(f"{prefix}_major"))
+        minor = int(match.group(f"{prefix}_minor"))
+        fraction_text = match.group(f"{prefix}_fraction") or ""
+        if mode == "minutes_seconds":
+            if minor >= 60:
+                return -1.0
+            fraction = int(fraction_text.ljust(3, "0")) / 1000.0 if fraction_text else 0.0
+            return major * 60.0 + minor + fraction
+        if fraction_text or minor >= 100:
+            return -1.0
+        return major + minor / 100.0
+
+    def valid_range_mode(mode: str) -> bool:
+        previous_end: float | None = None
+        for match in range_matches:
+            start = range_time(match, "start", mode)
+            end = range_time(match, "end", mode)
+            if start < 0 or end <= start:
+                return False
+            if int(match.group("shot")) == 1 and start != 0:
+                return False
+            if previous_end is not None and abs(start - previous_end) > 0.01:
+                return False
+            if duration > 0 and (start >= duration or end > duration + 0.05):
+                return False
+            previous_end = end
+        return bool(range_matches)
+
+    range_mode = ""
+    if valid_range_mode("minutes_seconds"):
+        range_mode = "minutes_seconds"
+    elif duration > 0 and valid_range_mode("seconds_hundredths"):
+        range_mode = "seconds_hundredths"
+    if range_mode:
+        def replace_range(match: re.Match[str]) -> str:
+            shot = int(match.group("shot"))
+            start = range_time(match, "start", range_mode)
+            if shot == 1:
+                return "[Shot 1] "
+            minutes = int(start // 60)
+            seconds = int(start % 60)
+            milliseconds = int(round((start - int(start)) * 1000))
+            if milliseconds == 1000:
+                seconds += 1
+                milliseconds = 0
+                if seconds == 60:
+                    minutes += 1
+                    seconds = 0
+            following = match.string[match.end() :]
+            article = "a " if re.match(
+                r"(?:close[- ]up|medium(?:[- ]wide)? shot|wide shot|long shot|high[- ]angle|low[- ]angle|bird['’]s[- ]eye)",
+                following,
+                flags=re.IGNORECASE,
+            ) else ""
+            return f"[Shot {shot}] At {minutes:02d}:{seconds:02d}.{milliseconds:03d}, the camera cuts to {article}"
+
+        detailed = range_pattern.sub(replace_range, detailed)
+        detailed = re.sub(
+            r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3},\s*the camera cuts to (?:a )?)([A-Z])",
+            lambda match: f"{match.group(1)}{match.group(2).casefold()}",
+            detailed,
+        )
+    detailed = re.sub(
+        r"(\[Shot 1\])\s+At\s+00:00\.000,?\s*",
+        r"\1 ",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+    detailed = re.sub(
+        r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3})(?!,)(?=\s+\S)",
+        r"\1,",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+    detailed = re.sub(
+        r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3},\s*)the\s+camera\s+(?:quick|clean)\s+cuts?\s+to\s+",
+        r"\1the camera cuts to ",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+    detailed = re.sub(
+        r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3},\s*)(?:a\s+)?(?:quick|hard|clean)?\s*cut\s+to\s+",
+        r"\1the camera cuts to ",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+    detailed = re.sub(
+        r"(\[Shot [2-9]\d*\]\s+At\s+\d{2,}:\d{2}\.\d{3},)(?!\s*the\s+(?:camera|shot|scene)\s+(?:(?:hard\s+)?cuts|transitions|changes|switches|cross-dissolves|fades|wipes)\s+(?:to|into)\b)(\s*)",
+        r"\1\2the camera cuts to a static wide view as ",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+    return re.sub(
+        r"(the camera cuts to a static wide view as )(The|An?)\b",
+        lambda match: f"{match.group(1)}{match.group(2).casefold()}",
+        detailed,
+        flags=re.IGNORECASE,
+    )
+
+
+def _repair_minimax_h3_ref_subject_timeline_labels(sections: dict[str, str]) -> None:
+    detailed = sections["detailed_description"]
+    shot1_match = re.search(r"\[Shot 1\]", detailed, flags=re.IGNORECASE)
+    if not shot1_match:
+        return
+
+    retention = sections["retention_analysis"]
+    timeline = detailed[shot1_match.start() :]
+    repairs: list[tuple[str, str]] = []
+    for tag, description in _minimax_h3_ref_definition_items(sections["subject_definitions"]):
+        if not tag.startswith("<Subject ") or tag in timeline:
+            continue
+        if not re.search(
+            r"\b(?:environment|landscape|setting|world|location|background|scenery|production design|visual style|palette|"
+            r"lighting|town|city|street|square|field|forest|room|interior|exterior)\b",
+            description,
+            flags=re.IGNORECASE,
+        ) or re.search(
+            r"\b(?:person|character|courier|protagonist|antagonist|creature|performer|actor|man|woman|boy|girl|animal|monster)\b",
+            description,
+            flags=re.IGNORECASE,
+        ):
+            return
+        source_tags = _minimax_h3_reference_tags(description)
+        if tag not in retention or len(source_tags) != 1 or source_tags[0] not in timeline:
+            return
+        repairs.append((tag, source_tags[0]))
+
+    if not repairs or len({source_tag for _tag, source_tag in repairs}) != len(repairs):
+        return
+    for tag, source_tag in repairs:
+        timeline = timeline.replace(source_tag, f"{tag} (derived from {source_tag})", 1)
+    sections["detailed_description"] = f"{detailed[: shot1_match.start()]}{timeline}"
+
+
+def _repair_minimax_h3_ref2va_structure(
+    prompt_text: str,
+    user_prompt: str = "",
+    reference_manifest: str = "",
+    duration_seconds: float = 0.0,
+) -> str:
+    prompt = _normalize_minimax_h3_prompt(prompt_text, "ref2va")
+    if not prompt:
+        return ""
+    timeline_duration = max(
+        0.0,
+        _safe_float(duration_seconds, 0.0) or _explicit_prompt_duration_seconds(user_prompt),
+    )
+
+    canonical_sections = _minimax_h3_ref_sections(prompt)
+    if canonical_sections is not None:
+        summary = canonical_sections["summary"].strip()
+        summary_match = re.match(r"^\[([^\]\n]+)\]\s*(.*)$", summary, flags=re.DOTALL)
+        summary_body = summary_match.group(2).strip() if summary_match else summary
+        actual_task_types = [item.strip() for item in summary_match.group(1).split("+")] if summary_match else []
+        expected_task_types = _minimax_h3_ref_summary_task_types(user_prompt, reference_manifest)
+        if actual_task_types != expected_task_types:
+            canonical_sections["summary"] = f"[{' + '.join(expected_task_types)}] {summary_body}".rstrip()
+        _repair_minimax_h3_ref_subject_timeline_labels(canonical_sections)
+        canonical_sections["detailed_description"] = _repair_minimax_h3_ref_detailed_timeline(
+            canonical_sections["detailed_description"],
+            timeline_duration,
+        )
+        repaired_prompt = _rebuild_minimax_h3_ref_prompt(canonical_sections)
+        return repaired_prompt if repaired_prompt != prompt else prompt
+
+    heading_pattern = re.compile(
+        r"(?im)^[ \t]*(subject[ _-]*definitions|summary(?:[ \t]*\[[^\]\n]*\])?|retention[ _-]*analysis|"
+        r"detail(?:ed)?[ _-]*description|overall[ _-]*soundscape|non[ _-]*diegetic[ _-]*music)"
+        r"[ \t]*(?P<delimiter>:[ \t]*|(?=\r?$))"
+    )
+    matches = list(heading_pattern.finditer(prompt))
+    if not matches:
+        return prompt
+
+    heading_names = [
+        re.sub(
+            r"[\s-]+",
+            "_",
+            re.sub(r"\s*\[[^\]\n]*\]\s*$", "", match.group(1)).strip().casefold(),
+        )
+        for match in matches
+    ]
+    heading_names = ["detailed_description" if name == "detail_description" else name for name in heading_names]
+    prefix = prompt[: matches[0].start()].strip()
+    has_colonless_heading = any(not match.group("delimiter").startswith(":") for match in matches)
+    if has_colonless_heading and not (
+        (not prefix and heading_names == list(_MINIMAX_H3_REF_FIELDS))
+        or (
+            prefix
+            and heading_names
+            == ["retention_analysis", "detailed_description", "overall_soundscape", "non_diegetic_music"]
+        )
+    ):
+        return prompt
+    if prefix and "subject_definitions" not in heading_names and "summary" not in heading_names:
+        if heading_names not in (
+            ["retention_analysis", "detailed_description", "overall_soundscape", "non_diegetic_music"],
+            ["overall_soundscape", "non_diegetic_music"],
+        ):
+            return prompt
+    if (
+        not any(name.startswith("summary") for name in heading_names)
+        and "subject_definitions" not in heading_names
+        and not _minimax_h3_ref_definition_items(prefix)
+    ):
+        return prompt
+
+    sections = {field_name: "" for field_name in _MINIMAX_H3_REF_FIELDS}
+    for index, match in enumerate(matches):
+        raw_name = re.sub(r"\s*\[[^\]\n]*\]\s*$", "", match.group(1))
+        field_name = re.sub(r"[\s-]+", "_", raw_name.strip().casefold())
+        if field_name == "detail_description":
+            field_name = "detailed_description"
+        value_end = matches[index + 1].start() if index + 1 < len(matches) else len(prompt)
+        value = prompt[match.end() : value_end].strip()
+        if field_name in sections:
+            sections[field_name] = "\n".join(item for item in (sections[field_name], value) if item).strip()
+    if has_colonless_heading and any(not sections[field_name] for field_name in heading_names):
+        return prompt
+
+    prefix_recovered = False
+    if prefix and heading_names in (
+        ["retention_analysis", "detailed_description", "overall_soundscape", "non_diegetic_music"],
+        ["overall_soundscape", "non_diegetic_music"],
+    ):
+        prefix_blocks = [block.strip() for block in re.split(r"\n\s*\n", prefix) if block.strip()]
+        retention_markers = _MINIMAX_H3_REF_VISUAL_RETENTION | _MINIMAX_H3_REF_AUDIO_RETENTION
+        retention_pattern = re.compile(
+            r"(?im)^\s*<\s*(?:Subject|Picture|Video|Audio)\s+\d+\s*>"
+            rf"(?:\s*(?:\([^\n)]*\)|\[[^\n\]]*\]))*\s*:\s*(?:{'|'.join(sorted(retention_markers))})\s*-\s*\S"
+        )
+
+        def is_retention_block(block: str) -> bool:
+            tagged_items = _minimax_h3_ref_definition_items(block)
+            lines = [line for line in block.splitlines() if line.strip()]
+            return bool(tagged_items) and len(retention_pattern.findall(block)) == len(tagged_items) == len(lines)
+
+        def is_definition_block(block: str) -> bool:
+            lines = [line for line in block.splitlines() if line.strip()]
+            return bool(lines) and len(_minimax_h3_ref_definition_items(block)) == len(lines)
+
+    if prefix and heading_names == ["retention_analysis", "detailed_description", "overall_soundscape", "non_diegetic_music"]:
+        summary_index = next(
+            (
+                index
+                for index, block in enumerate(prefix_blocks)
+                if re.match(r"^\[(?!Shot\b)[^\]\n]+\]\s+\S", block, flags=re.IGNORECASE | re.DOTALL)
+            ),
+            None,
+        )
+        definition_blocks = prefix_blocks[:summary_index] if summary_index is not None else []
+        if (
+            summary_index
+            and summary_index == len(prefix_blocks) - 1
+            and all(is_definition_block(block) and not retention_pattern.search(block) for block in definition_blocks)
+            and is_retention_block(sections["retention_analysis"])
+            and re.search(r"\[Shot 1\]", sections["detailed_description"], flags=re.IGNORECASE)
+            and sections["overall_soundscape"]
+            and sections["non_diegetic_music"]
+        ):
+            sections["subject_definitions"] = "\n".join(definition_blocks)
+            sections["summary"] = prefix_blocks[summary_index]
+            prefix_recovered = True
+        else:
+            return prompt
+
+    if prefix and heading_names == ["overall_soundscape", "non_diegetic_music"]:
+        summary_index = next(
+            (
+                index
+                for index, block in enumerate(prefix_blocks)
+                if re.match(r"^\[(?!Shot\b)[^\]\n]+\]\s+\S", block, flags=re.IGNORECASE | re.DOTALL)
+            ),
+            None,
+        )
+        definition_blocks = prefix_blocks[:summary_index] if summary_index is not None else []
+        retention_start = summary_index + 1 if summary_index is not None else 0
+        retention_end = retention_start
+        while retention_end < len(prefix_blocks) and is_retention_block(prefix_blocks[retention_end]):
+            retention_end += 1
+        detailed_blocks = prefix_blocks[retention_end:]
+        detailed_shot_index = next(
+            (
+                index
+                for index, block in enumerate(detailed_blocks)
+                if re.search(r"\[Shot 1\]", block, flags=re.IGNORECASE)
+            ),
+            None,
+        )
+        stray_tagged_detail = bool(
+            detailed_shot_index is not None
+            and any(_minimax_h3_ref_definition_items(block) for block in detailed_blocks[:detailed_shot_index])
+        )
+        if (
+            summary_index
+            and all(is_definition_block(block) and not retention_pattern.search(block) for block in definition_blocks)
+            and retention_end > retention_start
+            and detailed_blocks
+            and not stray_tagged_detail
+            and re.search(r"\[Shot 1\]", "\n\n".join(detailed_blocks), flags=re.IGNORECASE)
+        ):
+            sections["subject_definitions"] = "\n".join(definition_blocks)
+            sections["summary"] = prefix_blocks[summary_index]
+            sections["retention_analysis"] = "\n".join(prefix_blocks[retention_start:retention_end])
+            sections["detailed_description"] = "\n\n".join(detailed_blocks)
+            prefix_recovered = True
+        else:
+            return prompt
+
+    if prefix and not prefix_recovered:
+        if not sections["subject_definitions"]:
+            sections["subject_definitions"] = prefix
+        elif _minimax_h3_ref_definition_items(prefix):
+            sections["subject_definitions"] = f"{prefix}\n{sections['subject_definitions']}".strip()
+
+    definition_items: list[tuple[str, str]] = []
+    for tag, description in _minimax_h3_ref_definition_items(sections["subject_definitions"]):
+        if tag not in {item_tag for item_tag, _item_description in definition_items}:
+            definition_items.append((tag, description.strip(" []")))
+
+    definition_text = "\n".join(f"{tag}: {description}" for tag, description in definition_items)
+    for tag, description in _minimax_h3_reference_definitions(reference_manifest):
+        if tag not in _minimax_h3_reference_tags(definition_text):
+            definition_items.append((tag, description.strip(" []")))
+            definition_text = "\n".join(f"{item_tag}: {item_description}" for item_tag, item_description in definition_items)
+    if definition_items:
+        sections["subject_definitions"] = definition_text
+
+    summary = sections["summary"].strip()
+    summary_match = re.match(r"^\[([^\]\n]+)\]\s*(.*)$", summary, flags=re.DOTALL)
+    if summary_match:
+        summary = summary_match.group(2).strip()
+    task_prefix = " + ".join(_minimax_h3_ref_summary_task_types(user_prompt, reference_manifest))
+    sections["summary"] = f"[{task_prefix}] {summary}".rstrip()
+
+    retention = sections["retention_analysis"]
+    retention_lines: list[str] = []
+    for tag, definition in definition_items:
+        tag_match = re.fullmatch(r"<(Subject|Picture|Video|Audio) (\d+)>", tag)
+        if not tag_match:
+            continue
+        kind, ordinal = tag_match.groups()
+        existing_match = re.search(
+            rf"(?is)(?:<|\[)\s*{kind}\s+{ordinal}\s*(?:>|\])(?:\s*(?:\([^\n)]*\)|\[[^\n\]]*\]))*\s*:\s*"
+            r"(.*?)(?=(?:<|\[)\s*(?:Subject|Picture|Video|Audio)\s+\d+\s*(?:>|\])|$)",
+            retention,
+        )
+        existing = existing_match.group(1).strip().strip("[]").strip() if existing_match else ""
+        marker_match = re.match(
+            r"^(fully_preserved|partially_preserved|attribute_transfer|weak_reference|fully_copy|partially_copy|reference)\s*-\s*(.*)$",
+            existing,
+            flags=re.IGNORECASE | re.DOTALL,
+        )
+        if marker_match:
+            marker = marker_match.group(1).casefold()
+            retention_detail = marker_match.group(2).strip().strip("[]").strip()
+        else:
+            marker = "fully_preserved" if kind == "Subject" else ("reference" if kind == "Audio" else "attribute_transfer")
+            retention_detail = re.sub(r"^[^:]{0,48}:\s*", "", existing).strip().strip("[]").strip() or definition
+        allowed_markers = _MINIMAX_H3_REF_AUDIO_RETENTION if kind == "Audio" else _MINIMAX_H3_REF_VISUAL_RETENTION
+        if marker not in allowed_markers:
+            marker = "reference" if kind == "Audio" else ("fully_preserved" if kind == "Subject" else "attribute_transfer")
+        retention_lines.append(f"{tag}: {marker} - {retention_detail}")
+    if retention_lines:
+        sections["retention_analysis"] = "\n".join(retention_lines)
+
+    _repair_minimax_h3_ref_subject_timeline_labels(sections)
+    sections["detailed_description"] = _repair_minimax_h3_ref_detailed_timeline(
+        sections["detailed_description"],
+        timeline_duration,
+    )
+    return _rebuild_minimax_h3_ref_prompt(sections)
+
+
 def _extract_json_object(text: str) -> dict[str, Any] | None:
     if not text:
         return None
@@ -992,6 +1572,13 @@ def _is_image_identity_video_control(media_metadata: dict[str, Any] | None) -> b
     ) == "image_identity_video_control"
 
 
+def _is_minimax_h3_reference_context(media_metadata: dict[str, Any] | None) -> bool:
+    if not isinstance(media_metadata, dict):
+        return False
+    value = str(media_metadata.get("minimax_h3_mode", "")).strip().lower().replace("-", "_")
+    return value in {"ref2va", "ref2v", "r2v", "reference", "reference_to_video", "reference_to_audio_video"}
+
+
 def _synthesis_contract_text(media_metadata: dict[str, Any] | None) -> str:
     if _is_image_identity_video_control(media_metadata):
         return (
@@ -1092,6 +1679,17 @@ def _sampled_video_frame_sequence_intro(media_metadata: dict[str, Any] | None, f
     if video_frame_count > 24:
         shown = f"{shown}, ..."
     source = str(media_metadata.get("source", "video")) if isinstance(media_metadata, dict) else "video"
+    if _is_minimax_h3_reference_context(media_metadata):
+        picture_intro = (
+            f"The first {reference_count} attached image(s) are ordered MiniMax H3 reference pictures <Picture 1> through <Picture {reference_count}>. "
+            if reference_count
+            else "No reference pictures are attached for pixel analysis. "
+        )
+        return (
+            f"{picture_intro}The remaining images are ordered sampled frames from <Video 1> at "
+            f"{shown or 'unknown'}; they are analysis evidence, not additional Picture labels. Describe each reference only for the role "
+            "declared in the H3 reference manifest, and do not cross-wire identity, style, motion, camera, composition, or temporal roles."
+        )
     if source == "image+video" and _is_image_identity_video_control(media_metadata):
         lead = (
             "The first attached image is the identity reference for the output subject. "
@@ -1131,6 +1729,15 @@ def _media_grounding_instruction(media_metadata: dict[str, Any] | None) -> str:
         source = str(media_metadata.get("source", "none")).lower() if isinstance(media_metadata, dict) else "none"
         transport = str(media_metadata.get("transformers_video_transport", _video_transport_mode())) if isinstance(media_metadata, dict) else _video_transport_mode()
         if source in {"video", "image+video"} and transport == "sampled_frame_images":
+            if _is_minimax_h3_reference_context(media_metadata):
+                picture_count = int(max(0.0, _safe_float(media_metadata.get("minimax_h3_reference_image_batch_count"), 0.0)))
+                return (
+                    f"MiniMax H3 full-reference evidence is attached in declared order: the first {picture_count} image(s) map to "
+                    "<Picture N>, followed by sampled analysis frames from <Video 1>. Use each asset only for its manifest role; do not "
+                    "cross-wire subject identity, wardrobe, environment, style, action, camera, cut rhythm, composition, or temporal structure. "
+                    "The sampled video frames are not additional Picture labels. "
+                    f"{capture_facts}"
+                )
             if _is_image_identity_video_control(media_metadata):
                 return (
                     "Image+video synthesis mode is active. The first attached image is the identity reference for the output subject; "
@@ -1354,12 +1961,839 @@ def _verified_prompt_claims(
     return not reasons, reasons
 
 
+def _has_unnegated_prompt_phrase(text: str, pattern: str) -> bool:
+    for match in re.finditer(pattern, text, flags=re.IGNORECASE):
+        clause_start = max(
+            text.rfind(".", 0, match.start()),
+            text.rfind(";", 0, match.start()),
+            text.rfind("!", 0, match.start()),
+            text.rfind("?", 0, match.start()),
+            text.rfind("\n", 0, match.start()),
+        )
+        prefix = text[clause_start + 1 : match.start()]
+        redirects = list(
+            re.finditer(
+                r"\b(?:but|instead|rather)\b|,\s*(?:use|make|keep|choose|render)\b",
+                prefix,
+                flags=re.IGNORECASE,
+            )
+        )
+        if redirects:
+            prefix = prefix[redirects[-1].end() :]
+        if re.search(
+            r"\b(?:no|not|never|without|avoid(?:ing)?|exclude(?:d|s|ing)?|do\s+not|don't|must\s+not|mustn't|cannot|can't)\b",
+            prefix,
+            flags=re.IGNORECASE,
+        ):
+            continue
+        return True
+    return False
+
+
+def _has_negated_prompt_phrase(text: str, pattern: str) -> bool:
+    for negation in re.finditer(
+        r"\b(?:no|not|never|without|avoid(?:ing)?|exclude(?:d|s|ing)?|do\s+not|don't|must\s+not|mustn't|cannot|can't)\b",
+        text,
+        flags=re.IGNORECASE,
+    ):
+        if re.match(r"\s+only\b", text[negation.end() :], flags=re.IGNORECASE):
+            continue
+        sentence_end = re.search(r"[.!?\n]", text[negation.end() :])
+        span_end = negation.end() + (sentence_end.start() if sentence_end else 220)
+        if re.search(pattern, text[negation.end() : min(len(text), span_end, negation.end() + 220)], flags=re.IGNORECASE):
+            return True
+    return False
+
+
+_MINIMAX_H3_LIVE_ACTION_PATTERN = r"\b(?:live[- ]action|practical\s+(?:film|photography)|photoreal(?:istic|ism)?|realistic\s+film\s+photography)\b"
+_MINIMAX_H3_STOP_MOTION_PATTERN = r"\b(?:stop[- ]motion|claymation|clay[- ]animation|puppet[- ]animation)\b"
+_MINIMAX_H3_3D_ANIMATION_PATTERN = r"\b(?:3[ -]?d(?:\s+(?:cg|cgi|animation|animated|render(?:ed|ing)?))?|cgi?(?:[- ](?:animated|animation|render(?:ed|ing)?))?|computer[- ](?:animated|generated)|cg[- ]animation)\b"
+_MINIMAX_H3_2D_ANIMATION_PATTERN = r"\b(?:2[ -]?d(?:\s+(?:animation|animated|art))?|hand[- ]drawn|cel[- ](?:animation|animated|shaded)|traditional[- ]animation|anime|cartoon)\b"
+_MINIMAX_H3_ANIMATION_PATTERN = r"\b(?:animated|animation)\b"
+
+
+def _minimax_h3_requested_visual_medium(user_prompt: str) -> str:
+    text = _single_paragraph(_sanitize_prompt_text(user_prompt, "", 8000))
+    if not text:
+        return ""
+    if _has_unnegated_prompt_phrase(text, _MINIMAX_H3_3D_ANIMATION_PATTERN):
+        return "3d_animation"
+    if _has_unnegated_prompt_phrase(text, _MINIMAX_H3_STOP_MOTION_PATTERN):
+        return "stop_motion"
+    if _has_unnegated_prompt_phrase(text, _MINIMAX_H3_LIVE_ACTION_PATTERN):
+        return "live_action"
+    if _has_unnegated_prompt_phrase(text, _MINIMAX_H3_2D_ANIMATION_PATTERN):
+        return "2d_animation"
+    if _has_unnegated_prompt_phrase(text, _MINIMAX_H3_ANIMATION_PATTERN):
+        if re.search(r"\bpok[eé]mon(?:[- ]inspired)?\b", text, flags=re.IGNORECASE):
+            return "2d_animation"
+        return "animation"
+    return ""
+
+
+def _minimax_h3_requested_style_locks(user_prompt: str) -> list[str]:
+    text = _single_paragraph(_sanitize_prompt_text(user_prompt, "", 8000))
+    locks: list[str] = []
+    for match in re.finditer(r"\b([A-Za-z0-9][A-Za-z0-9'’]*)[- ]inspired\b", text, flags=re.IGNORECASE):
+        locks.append(f"{match.group(1)}-inspired")
+    for match in re.finditer(
+        r"\b(?:the\s+)?style\s+(?:(?:should|must|will)\s+be|is|:)\s*"
+        r"(?:reimagined\s+)?(?:into|as|in)?\s*([^.!?\n]+)",
+        text,
+        flags=re.IGNORECASE,
+    ):
+        lock = re.sub(r"^(?:a|an|the)\s+", "", match.group(1).strip(" ,:;"), flags=re.IGNORECASE)
+        if lock:
+            locks.append(lock)
+    deduped: list[str] = []
+    for lock in locks:
+        if lock.casefold() not in {item.casefold() for item in deduped}:
+            deduped.append(lock)
+    return deduped
+
+
+def _minimax_h3_style_lock_tokens(text: str) -> set[str]:
+    tokens: set[str] = set()
+    normalized = "".join(character for character in unicodedata.normalize("NFKD", text) if not unicodedata.combining(character))
+    for token in re.findall(r"[a-z0-9]+", normalized.casefold().replace("’", "'")):
+        if len(token) <= 1 or token in {"the", "and", "with", "style", "visual", "look"}:
+            continue
+        if re.fullmatch(r"\d{3,4}s", token):
+            token = token[:-1]
+        tokens.add(token)
+    return tokens
+
+
+def _minimax_h3_style_locks_present(prompt_text: str, style_locks: list[str]) -> bool:
+    prompt_tokens = _minimax_h3_style_lock_tokens(prompt_text)
+    return all(_minimax_h3_style_lock_tokens(lock).issubset(prompt_tokens) for lock in style_locks)
+
+
+def _apply_minimax_h3_visual_medium_lock(
+    prompt_text: str,
+    user_prompt: str,
+    minimax_h3_mode: str = "t2va",
+) -> str:
+    mode = _normalize_minimax_h3_mode(minimax_h3_mode)
+    prompt = _normalize_minimax_h3_prompt(prompt_text, mode)
+    medium = _minimax_h3_requested_visual_medium(user_prompt)
+    if mode == "ref2va":
+        sections = _minimax_h3_ref_sections(prompt)
+        if not sections or not medium:
+            return prompt
+        detailed = sections["detailed_description"]
+        shot_match = re.search(r"\[Shot 1\]", detailed)
+        if not shot_match:
+            return prompt
+        opening = detailed[: shot_match.start()]
+        anchor_patterns = {
+            "2d_animation": r"(?:\b2[ -]?d\s+hand[- ]drawn\s+cel[- ]animation\b|\b2[ -]?d\b[^.!?\n]{0,80}\b(?:hand[- ]drawn|cel[- ]animation|cartoon|anime)\b|\b(?:hand[- ]drawn|cel[- ]animation|cartoon|anime)\b[^.!?\n]{0,80}\b2[ -]?d\b)",
+            "3d_animation": _MINIMAX_H3_3D_ANIMATION_PATTERN,
+            "stop_motion": _MINIMAX_H3_STOP_MOTION_PATTERN,
+            "live_action": _MINIMAX_H3_LIVE_ACTION_PATTERN,
+            "animation": rf"(?:{_MINIMAX_H3_ANIMATION_PATTERN}|{_MINIMAX_H3_2D_ANIMATION_PATTERN}|{_MINIMAX_H3_3D_ANIMATION_PATTERN}|{_MINIMAX_H3_STOP_MOTION_PATTERN})",
+        }
+        anchors = {
+            "2d_animation": "The target video uses immutable 2D hand-drawn cel animation, with no live action, photorealism, or 3D CGI. ",
+            "3d_animation": "The target video uses immutable stylized 3D computer animation. ",
+            "stop_motion": "The target video uses immutable practical stop-motion animation. ",
+            "live_action": "The target video uses immutable realistic live-action practical film photography. ",
+            "animation": "The target video uses immutable stylized animation, with no live action or photorealism. ",
+        }
+        anchor_match = re.search(anchor_patterns[medium], opening[:400], flags=re.IGNORECASE)
+        if not anchor_match:
+            sections["detailed_description"] = f"{anchors[medium]}{detailed}"
+            return _rebuild_minimax_h3_ref_prompt(sections)
+        additions: list[str] = []
+        if medium == "2d_animation":
+            if not _has_negated_prompt_phrase(detailed, _MINIMAX_H3_LIVE_ACTION_PATTERN):
+                additions.append("no live action or photorealism")
+            if not _has_negated_prompt_phrase(detailed, _MINIMAX_H3_3D_ANIMATION_PATTERN):
+                additions.append("no 3D CGI")
+        elif medium == "animation" and not _has_negated_prompt_phrase(detailed, _MINIMAX_H3_LIVE_ACTION_PATTERN):
+            additions.append("no live action or photorealism")
+        if additions:
+            insert_at = anchor_match.end()
+            opening = f"{opening[:insert_at]}, with {' and '.join(additions)}{opening[insert_at:]}"
+            sections["detailed_description"] = f"{opening}{detailed[shot_match.start():]}"
+            return _rebuild_minimax_h3_ref_prompt(sections)
+        return prompt
+    opening_match = re.match(r"^integrated_multimodal_description\s*:\s*\[Shot 1\]\s*", prompt)
+    if not prompt or not medium or not opening_match:
+        return prompt
+    opening = prompt[opening_match.end() : opening_match.end() + 320]
+    anchor_patterns = {
+        "2d_animation": r"(?:\b2[ -]?d\s+hand[- ]drawn\s+cel[- ]animation\b|\b2[ -]?d\b[^.!?\n]{0,80}\b(?:hand[- ]drawn|cel[- ]animation|cartoon|anime)\b|"
+        r"\b(?:hand[- ]drawn|cel[- ]animation|cartoon|anime)\b[^.!?\n]{0,80}\b2[ -]?d\b)",
+        "3d_animation": _MINIMAX_H3_3D_ANIMATION_PATTERN,
+        "stop_motion": _MINIMAX_H3_STOP_MOTION_PATTERN,
+        "live_action": _MINIMAX_H3_LIVE_ACTION_PATTERN,
+        "animation": rf"(?:{_MINIMAX_H3_ANIMATION_PATTERN}|{_MINIMAX_H3_2D_ANIMATION_PATTERN}|{_MINIMAX_H3_3D_ANIMATION_PATTERN}|{_MINIMAX_H3_STOP_MOTION_PATTERN})",
+    }
+    anchor_match = re.search(anchor_patterns[medium], opening, flags=re.IGNORECASE)
+    if not anchor_match:
+        anchors = {
+            "2d_animation": "2D hand-drawn cel animation, with no live action, photorealism, or 3D CGI. ",
+            "3d_animation": "Stylized 3D computer animation. ",
+            "stop_motion": "Practical stop-motion animation. ",
+            "live_action": "Realistic live-action practical film photography. ",
+            "animation": "Stylized animation, with no live action or photorealism. ",
+        }
+        return f"{prompt[:opening_match.end()]}{anchors[medium]}{prompt[opening_match.end():]}"
+    anchor_end = opening_match.end() + anchor_match.end()
+    integrated = prompt.split("\n\noverall_soundscape:", 1)[0]
+    if medium == "2d_animation":
+        live_action_excluded = _has_negated_prompt_phrase(integrated, _MINIMAX_H3_LIVE_ACTION_PATTERN)
+        three_d_excluded = _has_negated_prompt_phrase(integrated, _MINIMAX_H3_3D_ANIMATION_PATTERN)
+        if not live_action_excluded or not three_d_excluded:
+            separator = "" if prompt[anchor_end : anchor_end + 1] == "," else ","
+            return f"{prompt[:anchor_end]}{separator} with no live action, photorealism, or 3D CGI{prompt[anchor_end:]}"
+    elif medium == "animation" and not _has_negated_prompt_phrase(integrated, _MINIMAX_H3_LIVE_ACTION_PATTERN):
+        separator = "" if prompt[anchor_end : anchor_end + 1] == "," else ","
+        return f"{prompt[:anchor_end]}{separator} with no live action or photorealism{prompt[anchor_end:]}"
+    return prompt
+
+
+def _apply_minimax_h3_style_locks(
+    prompt_text: str,
+    user_prompt: str,
+    minimax_h3_mode: str = "t2va",
+) -> str:
+    mode = _normalize_minimax_h3_mode(minimax_h3_mode)
+    prompt = _normalize_minimax_h3_prompt(prompt_text, mode)
+    style_locks = _minimax_h3_requested_style_locks(user_prompt)
+    if not prompt or not style_locks:
+        return prompt
+    if mode == "ref2va":
+        sections = _minimax_h3_ref_sections(prompt)
+        if not sections:
+            return prompt
+        detailed = sections["detailed_description"]
+        shot_match = re.search(r"\[Shot 1\]", detailed)
+        if not shot_match:
+            return prompt
+        opening = detailed[: shot_match.start()].strip()
+        missing = [lock for lock in style_locks if not _minimax_h3_style_locks_present(opening, [lock])]
+        if not missing:
+            return prompt
+        style_sentence = f"The target video faithfully uses {' and '.join(missing)} visual styling."
+        sections["detailed_description"] = (
+            f"{opening} {style_sentence} {detailed[shot_match.start():]}".strip()
+            if opening
+            else f"{style_sentence} {detailed}"
+        )
+        return _rebuild_minimax_h3_ref_prompt(sections)
+    opening_match = re.match(r"^integrated_multimodal_description\s*:\s*\[Shot 1\]\s*", prompt)
+    if not opening_match:
+        return prompt
+    opening = prompt[opening_match.end() : opening_match.end() + 320]
+    bracketed_style = re.search(r"\[[^\]\n]{1,240}\]", opening)
+    if bracketed_style and _minimax_h3_style_locks_present(bracketed_style.group(0), style_locks):
+        canonical_style = f"faithfully using {' and '.join(style_locks)} visual styling"
+        start = opening_match.end() + bracketed_style.start()
+        end = opening_match.end() + bracketed_style.end()
+        prompt = f"{prompt[:start]}{canonical_style}{prompt[end:]}"
+        opening = prompt[opening_match.end() : opening_match.end() + 320]
+    missing = [lock for lock in style_locks if not _minimax_h3_style_locks_present(prompt[:900], [lock])]
+    if not missing:
+        return prompt
+    medium_anchor = re.search(
+        r"\b(?:2[ -]?d\s+hand[- ]drawn\s+cel[- ]animation|stylized\s+3[ -]?d\s+computer\s+animation|"
+        r"practical\s+stop[- ]motion\s+animation|realistic\s+live[- ]action\s+practical\s+film\s+photography|stylized\s+animation)\b",
+        opening,
+        flags=re.IGNORECASE,
+    )
+    leading_separator = ""
+    if medium_anchor:
+        split_at = opening_match.end() + medium_anchor.end()
+        if prompt[split_at : split_at + 1] == ",":
+            split_at += 1
+        else:
+            leading_separator = ","
+    else:
+        insertion_point = prompt.find(",", opening_match.end(), opening_match.end() + 220)
+        insert_after_delimiter = insertion_point >= 0
+        if insertion_point < 0:
+            first_sentence_end = prompt.find(".", opening_match.end(), opening_match.end() + 220)
+            insertion_point = first_sentence_end if first_sentence_end >= 0 else opening_match.end()
+            insert_after_delimiter = first_sentence_end >= 0
+        split_at = insertion_point + 1 if insert_after_delimiter else insertion_point
+    style_text = " and ".join(missing)
+    suffix = prompt[split_at:]
+    separator = "" if suffix.startswith((" ", "\n")) else " "
+    return f"{prompt[:split_at]}{leading_separator} faithfully using {style_text} visual styling,{separator}{suffix}"
+
+
+_MINIMAX_H3_SOUND_EVENT_FIDELITY = (
+    (
+        "fire_blast",
+        r"\b(?:(?:fire|flame)[- ](?:blast|burst)|(?:crackle|roar|whoosh)\s+of\s+(?:a\s+)?(?:fire|flames?))\b",
+        r"\b(?:fire|flames?|fiery|burns?|burning|ignites?|ignition|blaze)\b",
+    ),
+    (
+        "gunfire",
+        r"\b(?:gunshots?|gunfire|rifle\s+shots?|pistol\s+shots?)\b",
+        r"\b(?:gun|pistol|rifle|revolver|bullet|gunfire|shoots?|fires?\s+(?:a\s+)?(?:round|bullet))\b",
+    ),
+    (
+        "explosion",
+        r"\b(?:explosion|detonation|explosive\s+boom)\b",
+        r"\b(?:explodes?|explosion|detonates?|detonation|blasts?\s+apart)\b",
+    ),
+)
+
+
+def _minimax_h3_orphaned_sound_events(integrated: str, soundscape: str) -> list[str]:
+    return [
+        label
+        for label, sound_pattern, visual_pattern in _MINIMAX_H3_SOUND_EVENT_FIDELITY
+        if re.search(sound_pattern, soundscape, flags=re.IGNORECASE)
+        and not re.search(visual_pattern, integrated, flags=re.IGNORECASE)
+    ]
+
+
+def _apply_minimax_h3_sound_fidelity(
+    prompt_text: str,
+    user_prompt: str,
+    minimax_h3_mode: str = "t2va",
+) -> str:
+    mode = _normalize_minimax_h3_mode(minimax_h3_mode)
+    prompt = _normalize_minimax_h3_prompt(prompt_text, mode)
+    ref_sections = _minimax_h3_ref_sections(prompt) if mode == "ref2va" else None
+    if ref_sections:
+        integrated = ref_sections["detailed_description"]
+        soundscape = ref_sections["overall_soundscape"]
+        music = ref_sections["non_diegetic_music"]
+    else:
+        integrated = ""
+        soundscape = ""
+        music = ""
+    structured_match = re.fullmatch(
+        r"(?ms)integrated_multimodal_description\s*:\s*(.*?)\s*\n+\s*overall_soundscape\s*:\s*(.*?)\s*\n+\s*non_diegetic_music\s*:\s*(.*)",
+        prompt,
+    )
+    if not ref_sections and not structured_match:
+        return prompt
+    if structured_match:
+        integrated = structured_match.group(1).strip()
+        soundscape = structured_match.group(2).strip()
+        music = structured_match.group(3).strip()
+    orphaned = set(_minimax_h3_orphaned_sound_events(integrated, soundscape))
+    removable_patterns = [
+        sound_pattern
+        for label, sound_pattern, _visual_pattern in _MINIMAX_H3_SOUND_EVENT_FIDELITY
+        if label in orphaned and not _has_unnegated_prompt_phrase(user_prompt, sound_pattern)
+    ]
+    if not removable_patterns:
+        return prompt
+    parts = [part.strip() for part in re.split(r",\s*", soundscape) if part.strip()]
+    kept = [
+        re.sub(r"^and\s+", "", part, flags=re.IGNORECASE).strip()
+        for part in parts
+        if not any(re.search(pattern, part, flags=re.IGNORECASE) for pattern in removable_patterns)
+    ]
+    if not kept:
+        repaired_soundscape = "N/A"
+    elif len(kept) == 1:
+        repaired_soundscape = kept[0].rstrip(".") + "."
+    elif len(kept) == 2:
+        repaired_soundscape = f"{kept[0].rstrip(' .')} and {kept[1].rstrip(' .')}."
+    else:
+        repaired_soundscape = f"{', '.join(item.rstrip(' .') for item in kept[:-1])}, and {kept[-1].rstrip(' .')}."
+    if ref_sections:
+        ref_sections["overall_soundscape"] = repaired_soundscape
+        return _rebuild_minimax_h3_ref_prompt(ref_sections)
+    return (
+        f"integrated_multimodal_description: {integrated}\n\n"
+        f"overall_soundscape: {repaired_soundscape}\n\n"
+        f"non_diegetic_music: {music}"
+    )
+
+
+def _minimax_h3_visual_medium_instruction(medium: str) -> str:
+    if medium == "2d_animation":
+        return (
+            'The requested rendering medium is immutable 2D animation. Begin the first sentence immediately after [Shot 1] with '
+            '"2D hand-drawn cel animation," preserve drawn linework, cel shading, stylized proportions, and design continuity across every cut, '
+            'and explicitly state "no live action, photorealism, or 3D CGI." Cinematic, period, and genre direction changes staging, lighting, '
+            "camera language, and editing inside that medium; it never changes the medium."
+        )
+    if medium == "3d_animation":
+        return (
+            'The requested rendering medium is immutable stylized 3D computer animation. Begin the first sentence immediately after [Shot 1] '
+            'with "Stylized 3D computer animation," and preserve the same rendering language and subject designs across every cut.'
+        )
+    if medium == "stop_motion":
+        return (
+            'The requested rendering medium is immutable stop-motion animation. Begin the first sentence immediately after [Shot 1] with '
+            '"Practical stop-motion animation," name the requested material or puppet technique, and preserve it across every cut.'
+        )
+    if medium == "live_action":
+        return (
+            'The requested rendering medium is immutable live action. Begin the first sentence immediately after [Shot 1] with '
+            '"Realistic live-action practical film photography," and preserve the photographic texture across every cut; cinematic language '
+            "must not silently convert it into cartoon or fully animated rendering."
+        )
+    if medium == "animation":
+        return (
+            'The requested rendering medium is immutable animation. Choose the single concrete animation technique best supported by the user\'s '
+            "style cues, name that technique in the first words immediately after [Shot 1], preserve it across every cut, and explicitly exclude "
+            "live action and photorealism."
+        )
+    return ""
+
+
+_MINIMAX_H3_NUMBER_WORDS = {
+    "one": 1,
+    "two": 2,
+    "three": 3,
+    "four": 4,
+    "five": 5,
+    "six": 6,
+    "seven": 7,
+    "eight": 8,
+    "nine": 9,
+    "ten": 10,
+    "eleven": 11,
+    "twelve": 12,
+}
+
+
+def _minimax_h3_requested_shot_count(user_prompt: str) -> int:
+    text = _single_paragraph(_sanitize_prompt_text(user_prompt, "", 8000))
+    labels = sorted({int(value) for value in re.findall(r"\bShot\s+(\d{1,2})\b", text, flags=re.IGNORECASE)})
+    if len(labels) >= 2 and labels == list(range(1, labels[-1] + 1)):
+        return labels[-1]
+    count_match = re.search(
+        r"\b(one|two|three|four|five|six|seven|eight|nine|ten|eleven|twelve|\d{1,2})\s+"
+        r"(?:(?:separate|distinct|individual|rapid[- ]cut|hard[- ]cut)\s+){0,2}shots?\b",
+        text,
+        flags=re.IGNORECASE,
+    )
+    if not count_match:
+        return 0
+    token = count_match.group(1).lower()
+    return _MINIMAX_H3_NUMBER_WORDS.get(token, int(token) if token.isdigit() else 0)
+
+
+def _minimax_h3_shot_has_camera_spec(shot_body: str) -> bool:
+    return bool(
+        re.search(
+            r"\b(?:extreme\s+close[- ]up|close[- ]up|medium(?:[- ](?:close|wide))?\s+shot|wide\s+shot|long\s+shot|full[- ]body\s+shot|"
+            r"two[- ]shot|over[- ]the[- ]shoulder|pov|point[- ]of[- ]view|bird'?s[- ]eye|overhead|high[- ]angle|low[- ]angle|dutch[- ]angle|"
+            r"profile(?:\s+view)?|locked[- ]off|static(?:\s+hold)?|same\s+framing\s+holds|tracking|tracks?|push(?:es)?\s+in|pull(?:s)?\s+back|"
+            r"zooms?|pans?|tilts?|trucks?|doll(?:y|ies)|pedestal|arcs?|orbits?|cranes?|handheld|steadicam|gimbal|"
+            r"camera\s+(?:holds?|shakes?|follows?|moves?|rises?|descends?))\b",
+            shot_body,
+            flags=re.IGNORECASE,
+        )
+    )
+
+
+_MINIMAX_H3_REF_TASK_TYPES = {
+    "keyframe completion",
+    "reference generation",
+    "video editing",
+    "video continuation",
+    "audio reuse",
+    "audio reference",
+}
+_MINIMAX_H3_REF_VISUAL_RETENTION = {
+    "fully_preserved",
+    "partially_preserved",
+    "attribute_transfer",
+    "weak_reference",
+}
+_MINIMAX_H3_REF_AUDIO_RETENTION = {
+    "fully_copy",
+    "partially_copy",
+    "reference",
+    "weak_reference",
+}
+
+
+def _minimax_h3_ref2va_validation_reasons(
+    prompt_text: str,
+    duration_seconds: float,
+    user_prompt: str = "",
+    audio_mode: str = "auto_scene_audio",
+    max_prompt_chars: int = 0,
+    reference_manifest: str = "",
+) -> list[str]:
+    prompt = _sanitize_structured_prompt_text(prompt_text, "", 0)
+    manifest = _normalize_minimax_h3_reference_manifest(reference_manifest)
+    manifest_definitions = _minimax_h3_reference_definitions(manifest)
+    manifest_tags = [tag for tag, _description in manifest_definitions]
+    reasons = _minimax_h3_reference_manifest_validation_reasons(manifest)
+    if max_prompt_chars > 0 and len(prompt) > max_prompt_chars:
+        reasons.append("minimax_h3_prompt_truncated")
+
+    positions: list[int] = []
+    for field_name in _MINIMAX_H3_REF_FIELDS:
+        matches = list(re.finditer(rf"(?m)^{re.escape(field_name)}\s*:", prompt))
+        if not matches:
+            reasons.append(f"minimax_h3_ref_missing_{field_name}")
+            positions.append(-1)
+        else:
+            positions.append(matches[0].start())
+            if len(matches) != 1:
+                reasons.append("minimax_h3_ref_sections_invalid")
+    if all(position >= 0 for position in positions) and positions != sorted(positions):
+        reasons.append("minimax_h3_ref_field_order_invalid")
+    if prompt and not re.match(r"^subject_definitions\s*:", prompt):
+        reasons.append("minimax_h3_ref_sections_invalid")
+    if any(not re.search(rf"\n\n{re.escape(field_name)}\s*:", prompt) for field_name in _MINIMAX_H3_REF_FIELDS[1:]):
+        reasons.append("minimax_h3_ref_field_spacing_invalid")
+    if re.search(
+        r"(?m)^(?!subject_definitions\s*:|summary\s*:|retention_analysis\s*:|detailed_description\s*:|overall_soundscape\s*:|non_diegetic_music\s*:)[A-Za-z][A-Za-z0-9 _-]{0,48}\s*:",
+        prompt,
+    ):
+        reasons.append("minimax_h3_extra_top_level_field")
+
+    sections = _minimax_h3_ref_sections(prompt)
+    if sections is None:
+        reasons.append("minimax_h3_ref_sections_invalid")
+        return list(dict.fromkeys(reasons))
+    for field_name in _MINIMAX_H3_REF_FIELDS:
+        if not sections[field_name]:
+            reasons.append(f"minimax_h3_ref_missing_{field_name}")
+
+    subject_definitions = sections["subject_definitions"]
+    summary = sections["summary"]
+    retention = sections["retention_analysis"]
+    detailed = sections["detailed_description"]
+    soundscape = sections["overall_soundscape"]
+    music = sections["non_diegetic_music"]
+
+    prompt_asset_tags = set(_minimax_h3_reference_tags(prompt))
+    declared_asset_tags = set(manifest_tags)
+    if declared_asset_tags - set(_minimax_h3_reference_tags(subject_definitions)):
+        reasons.append("minimax_h3_ref_missing_asset_tag")
+    if prompt_asset_tags - declared_asset_tags:
+        reasons.append("minimax_h3_ref_undefined_tag")
+
+    definition_items = _minimax_h3_ref_definition_items(subject_definitions)
+    defined_item_tags = [tag for tag, _description in definition_items]
+    defined_subjects = [tag for tag in defined_item_tags if tag.startswith("<Subject ")]
+    all_subjects = {
+        f"<Subject {int(ordinal)}>"
+        for ordinal in re.findall(r"<\s*Subject\s+(\d+)\s*>", prompt, flags=re.IGNORECASE)
+    }
+    if len(defined_item_tags) != len(set(defined_item_tags)) or any(
+        len(re.findall(r"[A-Za-z0-9]{2,}", description)) < 4 for _tag, description in definition_items
+    ):
+        reasons.append("minimax_h3_ref_subject_definition_invalid")
+    if all_subjects - set(defined_subjects):
+        reasons.append("minimax_h3_ref_subject_definition_invalid")
+    subject_ordinals = sorted(int(re.search(r"\d+", tag).group()) for tag in set(defined_subjects))
+    if subject_ordinals and subject_ordinals != list(range(1, max(subject_ordinals) + 1)):
+        reasons.append("minimax_h3_ref_subject_definition_invalid")
+    if any(tag not in retention or tag not in detailed for tag in defined_subjects):
+        reasons.append("minimax_h3_ref_subject_usage_invalid")
+
+    summary_match = re.match(r"^\[([^\]\n]+)\]\s+\S", summary)
+    task_types: list[str] = []
+    if summary_match:
+        task_types = [item.strip() for item in summary_match.group(1).split("+")]
+    if (
+        not summary_match
+        or not task_types
+        or len(task_types) != len(set(task_types))
+        or any(task_type not in _MINIMAX_H3_REF_TASK_TYPES for task_type in task_types)
+    ):
+        reasons.append("minimax_h3_ref_summary_invalid")
+
+    for tag in defined_item_tags:
+        kind = re.match(r"<(Subject|Picture|Video|Audio)\s", tag).group(1)
+        markers = _MINIMAX_H3_REF_AUDIO_RETENTION if kind == "Audio" else _MINIMAX_H3_REF_VISUAL_RETENTION
+        marker_pattern = "|".join(re.escape(marker) for marker in sorted(markers))
+        if not re.search(
+            rf"(?m)^\s*{re.escape(tag)}(?:\s*\([^\n)]*\))?\s*:\s*(?:{marker_pattern})\s+-\s*\S",
+            retention,
+        ):
+            reasons.append("minimax_h3_ref_retention_invalid")
+            break
+
+    shot1_match = re.search(r"\[Shot 1\]", detailed)
+    style_opening = detailed[: shot1_match.start()].strip() if shot1_match else ""
+    if not shot1_match or len(re.findall(r"[A-Za-z0-9]{2,}", style_opening)) < 5:
+        reasons.append("minimax_h3_ref_style_opening_missing")
+    if task_types and "video editing" not in task_types and len(re.findall(r"\b[\w'’-]+\b", detailed)) < 120:
+        reasons.append("minimax_h3_ref_detailed_description_short")
+
+    if shot1_match:
+        timeline = detailed[shot1_match.end() :].lstrip()
+        integrated = f"[Shot 1] {style_opening} {timeline}".strip()
+        integrated = re.sub(r"<\s*(?:Subject|Picture|Video|Audio)\s+\d+\s*>", "the referenced element", integrated)
+        t2va_prompt = (
+            f"integrated_multimodal_description: {integrated}\n\n"
+            f"overall_soundscape: {soundscape}\n\n"
+            f"non_diegetic_music: {music}"
+        )
+        t2va_prompt = re.sub(
+            r"<\s*(?:Subject|Picture|Video|Audio)\s+\d+\s*>",
+            "the referenced element",
+            t2va_prompt,
+            flags=re.IGNORECASE,
+        )
+        reasons.extend(
+            _minimax_h3_prompt_validation_reasons(
+                t2va_prompt,
+                duration_seconds,
+                user_prompt,
+                audio_mode,
+                0,
+                "t2va",
+                "",
+            )
+        )
+
+    return list(dict.fromkeys(reasons))
+
+
+def _minimax_h3_prompt_validation_reasons(
+    prompt_text: str,
+    duration_seconds: float,
+    user_prompt: str = "",
+    audio_mode: str = "auto_scene_audio",
+    max_prompt_chars: int = 0,
+    minimax_h3_mode: str = "t2va",
+    reference_manifest: str = "",
+) -> list[str]:
+    if _normalize_minimax_h3_mode(minimax_h3_mode) == "ref2va":
+        return _minimax_h3_ref2va_validation_reasons(
+            prompt_text,
+            duration_seconds,
+            user_prompt,
+            audio_mode,
+            max_prompt_chars,
+            reference_manifest,
+        )
+    prompt = _sanitize_structured_prompt_text(prompt_text, "", 0)
+    reasons: list[str] = []
+    if max_prompt_chars > 0 and len(prompt) > max_prompt_chars:
+        reasons.append("minimax_h3_prompt_truncated")
+    if re.search(r"<\s*(?:Subject|Picture|Video|Audio)\s+\d+\s*>", prompt, flags=re.IGNORECASE):
+        reasons.append("minimax_h3_t2va_reference_tag")
+    fields = (
+        "integrated_multimodal_description",
+        "overall_soundscape",
+        "non_diegetic_music",
+    )
+    positions: list[int] = []
+    for field_name in fields:
+        matches = list(re.finditer(rf"(?m)^{re.escape(field_name)}\s*:", prompt))
+        if not matches:
+            reasons.append(f"minimax_h3_missing_{field_name}")
+            positions.append(-1)
+        else:
+            positions.append(matches[0].start())
+            if len(matches) != 1:
+                reasons.append(f"minimax_h3_duplicate_{field_name}")
+    if all(position >= 0 for position in positions) and positions != sorted(positions):
+        reasons.append("minimax_h3_field_order_invalid")
+    if prompt and not re.match(r"^integrated_multimodal_description\s*:\s*\[Shot 1\]", prompt):
+        reasons.append("minimax_h3_opening_invalid")
+    if not re.search(r"\n\noverall_soundscape\s*:", prompt) or not re.search(r"\n\nnon_diegetic_music\s*:", prompt):
+        reasons.append("minimax_h3_field_spacing_invalid")
+    soundscape_match = re.search(
+        r"(?ms)^overall_soundscape\s*:\s*(.+?)\n\nnon_diegetic_music\s*:",
+        prompt,
+    )
+    music_match = re.search(r"(?ms)^non_diegetic_music\s*:\s*(.+?)\s*$", prompt)
+    soundscape = soundscape_match.group(1).strip() if soundscape_match else ""
+    music = music_match.group(1).strip() if music_match else ""
+    if not soundscape:
+        reasons.append("minimax_h3_empty_overall_soundscape")
+    if not music:
+        reasons.append("minimax_h3_empty_non_diegetic_music")
+
+    integrated = prompt
+    soundscape_index = positions[1] if len(positions) > 1 else -1
+    if soundscape_index >= 0:
+        integrated = prompt[:soundscape_index].rstrip()
+    if _minimax_h3_orphaned_sound_events(integrated, soundscape):
+        reasons.append("minimax_h3_orphaned_sound_event")
+    if re.search(
+        r"(?m)^(?!integrated_multimodal_description\s*:|overall_soundscape\s*:|non_diegetic_music\s*:)[A-Za-z][A-Za-z0-9 _-]{0,48}\s*:",
+        prompt,
+    ):
+        reasons.append("minimax_h3_extra_top_level_field")
+    if re.search(r"\[\s*\d+(?:\.\d+)?s\s*[-–—]\s*\d+(?:\.\d+)?s\s*\]", integrated, flags=re.IGNORECASE):
+        reasons.append("minimax_h3_noncanonical_time_range")
+
+    shot_pattern = re.compile(r"\[Shot (\d+)\](?:\s+At\s+(\d{2,}):(\d{2})\.(\d{3}),\s*)?")
+    shot_markers = list(re.finditer(r"\[\s*shot\b[^\]\n]*\]", integrated, flags=re.IGNORECASE))
+    shots = list(shot_pattern.finditer(integrated))
+    if len(shot_markers) != len(shots) or any(marker.start() != shot.start() for marker, shot in zip(shot_markers, shots)):
+        reasons.append("minimax_h3_shot_syntax_invalid")
+    timed_cut_markers = re.findall(
+        r"\bAt\s+\d{2,}:\d{2}\.\d{3},?\s+the\s+(?:camera|shot|scene)\s+"
+        r"(?:(?:hard\s+)?cuts|transitions|changes|switches|cross-dissolves|fades|wipes)\s+(?:to|into)\b",
+        integrated,
+        flags=re.IGNORECASE,
+    )
+    if len(timed_cut_markers) > max(0, len(shots) - 1):
+        reasons.append("minimax_h3_unlabeled_cut")
+    shot_numbers = [int(match.group(1)) for match in shots]
+    if not shot_numbers or shot_numbers != list(range(1, len(shot_numbers) + 1)):
+        reasons.append("minimax_h3_shot_numbering_invalid")
+    shot_bodies = [
+        integrated[match.end() : shots[index + 1].start() if index + 1 < len(shots) else len(integrated)].strip()
+        for index, match in enumerate(shots)
+    ]
+    if shots:
+        shot1_body = shot_bodies[0]
+        if shots[0].group(2) is not None or re.match(r"^At\b", shot1_body, flags=re.IGNORECASE):
+            reasons.append("minimax_h3_shot1_timestamp_invalid")
+        if not shot1_body or not re.search(r"[A-Za-z0-9]", shot1_body):
+            reasons.append("minimax_h3_empty_visual_timeline")
+
+        requested_medium = _minimax_h3_requested_visual_medium(user_prompt)
+        opening = shot1_body[:320]
+        requested_style_locks = _minimax_h3_requested_style_locks(user_prompt)
+        if requested_style_locks and not _minimax_h3_style_locks_present(shot1_body[:600], requested_style_locks):
+            reasons.append("minimax_h3_requested_style_missing")
+        medium_present = True
+        if requested_medium == "2d_animation":
+            medium_present = bool(
+                re.search(r"\b2[ -]?d\b", opening, flags=re.IGNORECASE)
+                and re.search(r"\b(?:hand[- ]drawn|cel[- ]animation|cartoon|anime)\b", opening, flags=re.IGNORECASE)
+            )
+        elif requested_medium == "3d_animation":
+            medium_present = bool(re.search(_MINIMAX_H3_3D_ANIMATION_PATTERN, opening, flags=re.IGNORECASE))
+        elif requested_medium == "stop_motion":
+            medium_present = bool(re.search(_MINIMAX_H3_STOP_MOTION_PATTERN, opening, flags=re.IGNORECASE))
+        elif requested_medium == "live_action":
+            medium_present = bool(re.search(_MINIMAX_H3_LIVE_ACTION_PATTERN, opening, flags=re.IGNORECASE))
+        elif requested_medium == "animation":
+            medium_present = bool(
+                re.search(
+                    rf"(?:{_MINIMAX_H3_ANIMATION_PATTERN}|{_MINIMAX_H3_2D_ANIMATION_PATTERN}|{_MINIMAX_H3_3D_ANIMATION_PATTERN}|{_MINIMAX_H3_STOP_MOTION_PATTERN})",
+                    opening,
+                    flags=re.IGNORECASE,
+                )
+            )
+        if not medium_present:
+            reasons.append("minimax_h3_requested_visual_medium_missing")
+        if requested_medium in {"2d_animation", "animation"}:
+            live_action_excluded = _has_negated_prompt_phrase(integrated, _MINIMAX_H3_LIVE_ACTION_PATTERN)
+            three_d_excluded = _has_negated_prompt_phrase(integrated, _MINIMAX_H3_3D_ANIMATION_PATTERN)
+            if not live_action_excluded or (requested_medium == "2d_animation" and not three_d_excluded):
+                reasons.append("minimax_h3_requested_visual_exclusion_missing")
+            if _has_unnegated_prompt_phrase(integrated, _MINIMAX_H3_LIVE_ACTION_PATTERN) or (
+                requested_medium == "2d_animation"
+                and _has_unnegated_prompt_phrase(integrated, _MINIMAX_H3_3D_ANIMATION_PATTERN)
+            ):
+                reasons.append("minimax_h3_requested_visual_medium_conflict")
+
+        requested_shot_count = _minimax_h3_requested_shot_count(user_prompt)
+        if requested_shot_count and len(shots) != requested_shot_count:
+            reasons.append("minimax_h3_requested_shot_count_mismatch")
+        if any(not _minimax_h3_shot_has_camera_spec(body) for body in shot_bodies):
+            reasons.append("minimax_h3_shot_camera_unspecified")
+        final_hold_requested = _has_unnegated_prompt_phrase(
+            user_prompt,
+            r"\b(?:freeze(?:[- ]frame)?|final\s+(?:frame|shot)\s+holds?|hold(?:ing|s)?\s+(?:the\s+)?(?:final|last)\s+(?:frame|shot))\b",
+        )
+        if final_hold_requested and not re.search(
+            r"\b(?:freeze(?:s|[- ]frame)?|static\s+hold|held\s+frame|(?:final|last)\s+frame\s+holds?)\b",
+            shot_bodies[-1],
+            flags=re.IGNORECASE,
+        ):
+            reasons.append("minimax_h3_requested_final_hold_missing")
+        final_frame_match = re.search(r"\bThe final frame holds on\b(.+)$", shot_bodies[-1], flags=re.IGNORECASE | re.DOTALL)
+        if final_frame_match and not final_hold_requested and re.search(
+            r"\b(?:mid[- ]air|mid[- ]clash|mid[- ]fight|mid[- ]attack|airborne|about to|prepares? to|begins? to|starts? to|charges? (?:an? |the )?attack|"
+            r"runs? toward|sprints? toward|disappears?|vanishes?)\b",
+            final_frame_match.group(1),
+            flags=re.IGNORECASE,
+        ):
+            reasons.append("minimax_h3_unresolved_final_state")
+        if _has_unnegated_prompt_phrase(user_prompt, r"\b(?:battle|fight|duel|combat|clash)\b") and not re.search(
+            r"\b(?:knock(?:s|ed|ing)?(?:\s+back(?:ward)?)?|throw(?:s|n|ing)?(?:\s+back(?:ward)?)?|recoil(?:s|ed|ing)?|"
+            r"stagger(?:s|ed|ing)?|stumbl(?:es|ed|ing)|skids?|skidd(?:ing|ed)|slides? back|slid back|falls?|fell|collapses?|"
+            r"blocks?|blocked|dodges?|dodged|evades?|evaded|deflects?|deflected|absorbs?|absorbed|counters?|countered|"
+            r"driven back|forced back|sent (?:backward|sprawling|to the ground)|pushed back|regains? (?:its|his|her|their) footing)\b",
+            " ".join(shot_bodies[1:]),
+            flags=re.IGNORECASE,
+        ):
+            reasons.append("minimax_h3_battle_result_missing")
+
+    previous_cut = 0.0
+    transition_effect_used = False
+    for match in shots[1:]:
+        if match.group(2) is None:
+            reasons.append("minimax_h3_cut_timestamp_invalid")
+            continue
+        minutes = int(match.group(2))
+        seconds = int(match.group(3))
+        milliseconds = int(match.group(4))
+        if seconds >= 60:
+            reasons.append("minimax_h3_cut_timestamp_invalid")
+            continue
+        cut_time = minutes * 60.0 + seconds + milliseconds / 1000.0
+        if cut_time <= previous_cut:
+            reasons.append("minimax_h3_cut_timestamp_invalid")
+        if duration_seconds > 0 and cut_time >= duration_seconds:
+            reasons.append("minimax_h3_cut_timestamp_out_of_range")
+        transition_text = integrated[match.end() :].lstrip()
+        if re.match(
+            r"the (?:camera|shot|scene) (?:cross-dissolves|fades|wipes) (?:to|into)\b",
+            transition_text,
+            flags=re.IGNORECASE,
+        ):
+            transition_effect_used = True
+        if not re.match(
+            r"the (?:camera|shot|scene) (?:(?:hard )?cuts|transitions|changes|switches|cross-dissolves|fades|wipes) (?:to|into)\b",
+            transition_text,
+            flags=re.IGNORECASE,
+        ):
+            reasons.append("minimax_h3_cut_transition_invalid")
+        previous_cut = cut_time
+
+    if transition_effect_used and not re.search(
+        r"\b(?:cross[- ]dissolves?|fades?|wipes?)\b",
+        user_prompt,
+        flags=re.IGNORECASE,
+    ):
+        reasons.append("minimax_h3_unrequested_transition_effect")
+
+    if prompt.count("<d>") != integrated.count("<d>") or prompt.count("</d>") != integrated.count("</d>"):
+        reasons.append("minimax_h3_dialogue_outside_timeline")
+    if prompt.count("<d>") != prompt.count("</d>"):
+        reasons.append("minimax_h3_dialogue_tags_invalid")
+    elif "<d>" in integrated and len(re.findall(r"<d>\[[^\]\n]+\]\s*.*?</d>", integrated, flags=re.DOTALL)) != integrated.count("<d>"):
+        reasons.append("minimax_h3_dialogue_tags_invalid")
+    for dialogue in re.finditer(r"<d>", integrated):
+        preceding = integrated[max(0, dialogue.start() - 500) : dialogue.start()]
+        if not re.search(r"\(S\d+(?:\s*,\s*S\d+)*\)[^<]{0,500}$", preceding, flags=re.IGNORECASE):
+            reasons.append("minimax_h3_dialogue_speaker_invalid")
+
+    if _normalize_audio_mode(audio_mode) == "visual_only":
+        if soundscape.casefold() != "n/a" or music.casefold() != "n/a":
+            reasons.append("minimax_h3_visual_only_audio_fields_invalid")
+        if re.search(
+            r"<d>|</d>|\b(?:we hear|audible|audio|soundscape|foley|ambience|music|score|soundtrack|voiceover|says?|asks?|replies?|shouts?|whispers?|sings?)\b",
+            integrated,
+            flags=re.IGNORECASE,
+        ):
+            reasons.append("minimax_h3_visual_only_audio_cue")
+
+    deduped: list[str] = []
+    for reason in reasons:
+        if reason not in deduped:
+            deduped.append(reason)
+    return deduped
+
+
 def _packet_generation_readiness(
     packet: dict[str, Any],
     context: GemmaContext,
     target_config: TargetProfileConfig | dict[str, Any],
     prompt_text: str,
     media_metadata: dict[str, Any],
+    max_prompt_chars: int = 8000,
 ) -> tuple[bool, list[str]]:
     metadata = packet.get("metadata") if isinstance(packet.get("metadata"), dict) else {}
     reasons: list[str] = []
@@ -1382,6 +2816,25 @@ def _packet_generation_readiness(
         if _safe_float(media_metadata.get("video_sampled_frame_count"), 0.0) <= 0:
             reasons.append("synthesis_missing_control_video")
 
+    target = _target_profile_config_to_dict(target_config)
+    if target["target_profile"] == "minimax_h3":
+        raw_h3_prompt = str(packet.get("minimax_h3_prompt", "") or "")
+        h3_mode = target["minimax_h3_mode"]
+        reference_manifest = _normalize_minimax_h3_reference_manifest(
+            str(media_metadata.get("minimax_h3_reference_manifest", ""))
+        )
+        reasons.extend(
+            _minimax_h3_prompt_validation_reasons(
+                raw_h3_prompt,
+                _minimax_h3_duration_seconds(media_metadata, target["target_duration_seconds"], context.user_prompt),
+                context.user_prompt,
+                target["audio_mode"],
+                int(max_prompt_chars),
+                h3_mode,
+                reference_manifest,
+            )
+        )
+
     claims_ok, claim_reasons = _verified_prompt_claims(
         prompt_text=prompt_text,
         context=context,
@@ -1513,9 +2966,32 @@ def _contains_audio_terms(prompt: str) -> bool:
     )
 
 
-def _audio_instruction(audio_mode: str, audio_guidance: str) -> str:
+def _audio_instruction(audio_mode: str, audio_guidance: str, target_profile: str = "ltx") -> str:
     mode = _normalize_audio_mode(audio_mode)
     guidance = _audio_guidance_text(audio_guidance)
+    if _normalize_target_profile(target_profile) == "minimax_h3":
+        if mode == "visual_only":
+            return (
+                "Set overall_soundscape and non_diegetic_music to N/A. Do not invent dialogue, singing, ambience, Foley, or music."
+            )
+        if guidance:
+            return (
+                "Place synchronized dialogue, diegetic music, and event-specific sounds at the exact visible beat inside "
+                "integrated_multimodal_description; summarize only ambience, Foley, and non-verbal sounds in overall_soundscape; "
+                f"describe audience-only score in non_diegetic_music. Apply this requested sound design: {guidance}."
+            )
+        if mode == "explicit_sound_design":
+            return (
+                "Build concrete native audio with the video: synchronize dialogue and event-specific sounds inside the shot timeline, "
+                "write 1-4 sentences of ambience/Foley in overall_soundscape, and write 1-3 sentences of audience-only score in "
+                "non_diegetic_music with instrumentation, tempo/rhythm, and dynamic development. Avoid abstract mood or explanations "
+                "of the score's emotional purpose. Do not duplicate dialogue or music across fields."
+            )
+        return (
+            "Infer restrained, scene-specific native audio from visible actions and setting. Put synchronized events in the shot timeline, "
+            "ambient/Foley sound in overall_soundscape, and use non_diegetic_music: N/A unless score is requested or clearly implied. "
+            "If the user explicitly requests complete silence, set both audio fields to N/A."
+        )
     if mode == "visual_only":
         return "Do not include sound, music, dialogue, ambience, or other audio cues in the LTX prompt."
     if guidance:
@@ -1545,6 +3021,56 @@ def _effective_duration_seconds(media_metadata: dict[str, Any], target_duration_
     return _safe_float(media_metadata.get("duration_seconds"), 0.0)
 
 
+def _duration_unit_seconds(value: str, unit: str) -> float:
+    multiplier = 3600.0 if unit.lower().startswith(("h", "hr")) else 60.0 if unit.lower().startswith(("m", "min")) else 1.0
+    return _safe_float(value, 0.0) * multiplier
+
+
+def _explicit_prompt_duration_seconds(user_prompt: str) -> float:
+    text = _single_paragraph(_sanitize_prompt_text(user_prompt, "", 4000)).lower()
+    if not text:
+        return 0.0
+    combined = re.search(
+        r"\b(\d+(?:\.\d+)?)\s*(?:minutes?|mins?)\s*(?:and\s*)?(\d+(?:\.\d+)?)\s*(?:seconds?|secs?)\b",
+        text,
+    )
+    if combined:
+        return _safe_float(combined.group(1), 0.0) * 60.0 + _safe_float(combined.group(2), 0.0)
+    timecode = re.search(r"\b(?:duration|runtime|length)\s*(?:is|of|:)?\s*(\d{1,3}):(\d{2}(?:\.\d+)?)\b", text)
+    if timecode and _safe_float(timecode.group(2), 60.0) < 60.0:
+        return _safe_float(timecode.group(1), 0.0) * 60.0 + _safe_float(timecode.group(2), 0.0)
+    unit = r"hours?|hrs?|hr|h|minutes?|mins?|min|m|seconds?|secs?|sec|s"
+    patterns = (
+        rf"\b(?:duration|runtime|length)\s*(?:is|of|:)?\s*(\d+(?:\.\d+)?)\s*({unit})\b",
+        rf"\b(?:lasting|lasts?|runs?|running|for)\s+(\d+(?:\.\d+)?)\s*({unit})\b",
+        rf"\b(?:make|create|generate|render|produce)\b.{{0,48}}?\b(\d+(?:\.\d+)?)\s*[- ]?\s*({unit})\b",
+        rf"\b(\d+(?:\.\d+)?)\s*-\s*({unit})\b",
+        rf"\b(\d+(?:\.\d+)?)\s*[- ]?\s*({unit})\s+(?:video|clip|sequence|animation|film)\b",
+    )
+    for pattern in patterns:
+        match = re.search(pattern, text)
+        if match:
+            return _duration_unit_seconds(match.group(1), match.group(2))
+    return 0.0
+
+
+def _minimax_h3_duration_seconds(
+    media_metadata: dict[str, Any],
+    target_duration_seconds: float,
+    user_prompt: str = "",
+) -> float:
+    effective = _effective_duration_seconds(media_metadata, target_duration_seconds)
+    if effective > 0:
+        return effective
+    return _explicit_prompt_duration_seconds(user_prompt) or 5.0
+
+
+def _minimax_h3_duration_advisory(duration_seconds: float) -> str:
+    if 4.0 <= duration_seconds <= 15.0:
+        return ""
+    return "Outside MiniMax's published 4-15 second range; preserved without clamping for local generation."
+
+
 def _target_context(
     media_metadata: dict[str, Any],
     target_duration_seconds: float,
@@ -1558,8 +3084,10 @@ def _target_context(
     negative_prompt_mode: str = "auto",
     negative_prompt_guidance: str = "",
 ) -> dict[str, Any]:
+    effective_duration = round(_effective_duration_seconds(media_metadata, target_duration_seconds), 3)
     return {
-        "ltx_duration_seconds": round(_effective_duration_seconds(media_metadata, target_duration_seconds), 3),
+        "target_duration_seconds": effective_duration,
+        "ltx_duration_seconds": effective_duration,
         "ltx_style": _clean_control_text(ltx_style, 240),
         "ideogram_aspect_ratio": ideogram_aspect_ratio,
         "ideogram_render_style": _clean_control_text(ideogram_render_style, 240),
@@ -1618,8 +3146,26 @@ def _creativity_instruction(creativity_mode: str, creative_strength: float) -> s
     )
 
 
-def _target_structure_anchor(target_profile: str, ideogram_json_output: bool) -> str:
+def _target_structure_anchor(
+    target_profile: str,
+    ideogram_json_output: bool,
+    minimax_h3_mode: str = "t2va",
+) -> str:
     profile = _normalize_target_profile(target_profile)
+    if profile == "minimax_h3":
+        if _normalize_minimax_h3_mode(minimax_h3_mode) == "ref2va":
+            return (
+                "MiniMax H3 Ref2VA prompt: exactly six English sections in this order: subject_definitions, summary, "
+                "retention_analysis, detailed_description, overall_soundscape, non_diegetic_music. Define semantic <Subject N> "
+                "labels from the declared <Picture N>, <Video N>, and <Audio N> assets; begin summary with the official "
+                "square-bracketed task types; put one or two style sentences before [Shot 1] in detailed_description; and use "
+                "the exact declared reference labels wherever their roles apply. Preserve these field names, order, and blank lines."
+            )
+        return (
+            "MiniMax H3 T2VA prompt: integrated_multimodal_description begins directly with [Shot 1] and no timestamp; "
+            "later cuts use [Shot N] At MM:SS.mmm, the camera cuts to... with strictly increasing cut times; then one blank line, "
+            "overall_soundscape; then one blank line, non_diegetic_music. Preserve these exact field names, order, and line breaks."
+        )
     if profile == "ideogram4":
         if ideogram_json_output:
             return (
@@ -1687,16 +3233,88 @@ def _creative_art_direction_phrase(creativity_mode: str, creative_strength: floa
 
 
 def _normalize_target_profile(target_profile: str) -> str:
-    value = (target_profile or "").strip().lower().replace("-", "_").replace(".", "_")
+    value = (target_profile or "").strip().lower().replace("-", "_").replace(".", "_").replace(" ", "_")
     if value in {"ltx2", "ltx2_3", "ltx_2_3", "ltx"}:
         return "ltx"
     if value in {"ideogram", "ideogram4", "ideogram4_json", "ideogram_4"}:
         return "ideogram4"
+    if value in {"minimax", "minimaxh3", "minimax_h3", "h3", "hailuo_h3", "hailuo_03"}:
+        return "minimax_h3"
     if value == "both":
         return "ltx"
     return "ltx"
 
 
+def _normalize_minimax_h3_mode(mode: str) -> str:
+    value = (mode or "").strip().lower().replace("-", "_").replace(" ", "_")
+    if value in {"ref2va", "ref2v", "r2v", "reference", "reference_to_video", "reference_to_audio_video"}:
+        return "ref2va"
+    return "t2va"
+
+
+def _normalize_minimax_h3_reference_manifest(manifest: str) -> str:
+    value = _sanitize_structured_prompt_text(manifest, "", 6000)
+    return re.sub(
+        r"<\s*(picture|image|video|audio)\s+(\d+)\s*>",
+        lambda match: f"<{'Picture' if match.group(1).casefold() in {'picture', 'image'} else match.group(1).title()} {int(match.group(2))}>",
+        value,
+        flags=re.IGNORECASE,
+    )
+
+
+def _minimax_h3_reference_tags(text: str) -> list[str]:
+    tags: list[str] = []
+    for kind, ordinal in re.findall(r"<\s*(Picture|Video|Audio)\s+(\d+)\s*>", text or "", flags=re.IGNORECASE):
+        tag = f"<{kind.title()} {int(ordinal)}>"
+        if tag not in tags:
+            tags.append(tag)
+    return tags
+
+
+def _minimax_h3_reference_definitions(manifest: str) -> list[tuple[str, str]]:
+    definitions: list[tuple[str, str]] = []
+    for match in re.finditer(
+        r"(?m)^\s*<\s*(Picture|Video|Audio)\s+(\d+)\s*>\s*[:=\-]\s*(\S.*)$",
+        _normalize_minimax_h3_reference_manifest(manifest),
+        flags=re.IGNORECASE,
+    ):
+        definitions.append((f"<{match.group(1).title()} {int(match.group(2))}>", match.group(3).strip()))
+    return definitions
+
+
+def _minimax_h3_reference_manifest_validation_reasons(manifest: str) -> list[str]:
+    normalized = _normalize_minimax_h3_reference_manifest(manifest)
+    definitions = _minimax_h3_reference_definitions(normalized)
+    reasons: list[str] = []
+    if not definitions:
+        return ["minimax_h3_ref_manifest_missing"]
+
+    tags = [tag for tag, _description in definitions]
+    if len(tags) != len(set(tags)):
+        reasons.append("minimax_h3_ref_manifest_duplicate_tag")
+    limits = {"Picture": 9, "Video": 3, "Audio": 6}
+    for kind, limit in limits.items():
+        ordinals = [
+            int(match.group(1))
+            for tag in tags
+            if (match := re.fullmatch(rf"<{kind} (\d+)>", tag))
+        ]
+        if ordinals and sorted(set(ordinals)) != list(range(1, max(ordinals) + 1)):
+            reasons.append("minimax_h3_ref_manifest_nonconsecutive_tags")
+        if ordinals and max(ordinals) > limit:
+            reasons.append("minimax_h3_ref_manifest_limit_exceeded")
+    if not any(tag.startswith(("<Picture ", "<Video ")) for tag in tags):
+        reasons.append("minimax_h3_ref_manifest_audio_only")
+    if any(len(re.findall(r"[A-Za-z0-9]{2,}", description)) < 3 for _tag, description in definitions):
+        reasons.append("minimax_h3_ref_manifest_role_missing")
+
+    defined = set(tags)
+    mentioned = set(_minimax_h3_reference_tags(normalized))
+    if mentioned - defined:
+        reasons.append("minimax_h3_ref_manifest_undefined_tag")
+    return list(dict.fromkeys(reasons))
+
+
 def _normalize_hex_palette(*texts: str, max_colors: int = 16) -> list[str]:
     colors: list[str] = []
     for text in texts:
@@ -2098,6 +3716,7 @@ def _target_profile_config_to_dict(config: TargetProfileConfig | dict[str, Any]
         "ideogram_json_output": bool(data.get("ideogram_json_output", True)),
         "negative_prompt_mode": _normalize_negative_prompt_mode(str(data.get("negative_prompt_mode", "auto"))),
         "negative_prompt_guidance": _negative_prompt_guidance_text(str(data.get("negative_prompt_guidance", ""))),
+        "minimax_h3_mode": _normalize_minimax_h3_mode(str(data.get("minimax_h3_mode", "t2va"))),
     }
 
 
@@ -2113,6 +3732,7 @@ def _make_target_profile_config(
     ideogram_json_output: bool = True,
     negative_prompt_mode: str = "auto",
     negative_prompt_guidance: str = "",
+    minimax_h3_mode: str = "t2va",
 ) -> TargetProfileConfig:
     data = _target_profile_config_to_dict(
         {
@@ -2127,6 +3747,7 @@ def _make_target_profile_config(
             "ideogram_json_output": ideogram_json_output,
             "negative_prompt_mode": negative_prompt_mode,
             "negative_prompt_guidance": negative_prompt_guidance,
+            "minimax_h3_mode": minimax_h3_mode,
         }
     )
     return TargetProfileConfig(**data)
@@ -2299,6 +3920,192 @@ def _fit_reference_image_to_video_frames(reference_image: Any, video_frames: Any
         return first_reference, report
 
 
+_MINIMAX_H3_REFINABLE_REASONS = {
+    "minimax_h3_prompt_truncated",
+    "minimax_h3_extra_top_level_field",
+    "minimax_h3_noncanonical_time_range",
+    "minimax_h3_shot_syntax_invalid",
+    "minimax_h3_unlabeled_cut",
+    "minimax_h3_shot_numbering_invalid",
+    "minimax_h3_shot1_timestamp_invalid",
+    "minimax_h3_empty_visual_timeline",
+    "minimax_h3_cut_timestamp_invalid",
+    "minimax_h3_cut_timestamp_out_of_range",
+    "minimax_h3_requested_visual_medium_missing",
+    "minimax_h3_requested_visual_exclusion_missing",
+    "minimax_h3_requested_visual_medium_conflict",
+    "minimax_h3_requested_style_missing",
+    "minimax_h3_requested_shot_count_mismatch",
+    "minimax_h3_shot_camera_unspecified",
+    "minimax_h3_requested_final_hold_missing",
+    "minimax_h3_unresolved_final_state",
+    "minimax_h3_battle_result_missing",
+    "minimax_h3_cut_transition_invalid",
+    "minimax_h3_orphaned_sound_event",
+    "minimax_h3_ref_sections_invalid",
+    "minimax_h3_ref_field_order_invalid",
+    "minimax_h3_ref_field_spacing_invalid",
+    "minimax_h3_ref_missing_subject_definitions",
+    "minimax_h3_ref_missing_summary",
+    "minimax_h3_ref_missing_retention_analysis",
+    "minimax_h3_ref_missing_detailed_description",
+    "minimax_h3_ref_missing_overall_soundscape",
+    "minimax_h3_ref_missing_non_diegetic_music",
+    "minimax_h3_ref_missing_asset_tag",
+    "minimax_h3_ref_undefined_tag",
+    "minimax_h3_ref_subject_definition_invalid",
+    "minimax_h3_ref_subject_usage_invalid",
+    "minimax_h3_ref_summary_invalid",
+    "minimax_h3_ref_retention_invalid",
+    "minimax_h3_ref_style_opening_missing",
+    "minimax_h3_ref_detailed_description_short",
+}
+
+
+def _build_minimax_h3_refinement_prompt(
+    user_prompt: str,
+    candidate_prompt: str,
+    duration_seconds: float,
+    audio_mode: str,
+    reasons: list[str],
+    minimax_h3_mode: str = "t2va",
+    reference_manifest: str = "",
+    strict_retry: bool = False,
+) -> str:
+    minimax_h3_mode = _normalize_minimax_h3_mode(minimax_h3_mode)
+    reference_manifest = _normalize_minimax_h3_reference_manifest(reference_manifest)
+    requested_medium = _minimax_h3_requested_visual_medium(user_prompt)
+    style_locks = _minimax_h3_requested_style_locks(user_prompt)
+    corrections: list[str] = []
+    if requested_medium and minimax_h3_mode == "ref2va":
+        corrections.append(
+            f"State the immutable {requested_medium.replace('_', ' ')} medium in the one-or-two-sentence style opening before [Shot 1], "
+            "and keep that medium across every cut."
+        )
+    elif requested_medium:
+        corrections.append(_minimax_h3_visual_medium_instruction(requested_medium))
+    if style_locks:
+        style_location = "in the style opening before [Shot 1]" if minimax_h3_mode == "ref2va" else "near the start of Shot 1"
+        corrections.append(f"Copy these exact style phrases {style_location}: {'; '.join(style_locks)}.")
+    if "minimax_h3_requested_shot_count_mismatch" in reasons:
+        corrections.append(f"Use exactly {_minimax_h3_requested_shot_count(user_prompt)} consecutively numbered shots.")
+    if "minimax_h3_shot_camera_unspecified" in reasons:
+        corrections.append("Give every shot a concrete camera move or an explicit locked/static hold, not only an angle or shot size.")
+    if "minimax_h3_requested_final_hold_missing" in reasons:
+        corrections.append("Put the requested freeze or hold in the final shot and name the held composition.")
+    if "minimax_h3_requested_visual_exclusion_missing" in reasons:
+        corrections.append(
+            "State the required medium exclusions explicitly in the style opening: for 2D animation write exactly "
+            "'no live action, photorealism, or 3D CGI'; for other animation write 'no live action or photorealism'."
+        )
+    if any(
+        reason in reasons
+        for reason in (
+            "minimax_h3_noncanonical_time_range",
+            "minimax_h3_shot1_timestamp_invalid",
+            "minimax_h3_cut_timestamp_invalid",
+            "minimax_h3_cut_timestamp_out_of_range",
+            "minimax_h3_cut_transition_invalid",
+        )
+    ):
+        corrections.append(
+            "Shot 1 must begin exactly with '[Shot 1]' and no timestamp. Never write start-end ranges. Begin every later shot exactly "
+            "with one start instant: '[Shot N] At MM:SS.mmm, the camera cuts to ...'; put camera motion after that cut phrase."
+        )
+    if "minimax_h3_battle_result_missing" in reasons:
+        corrections.append(
+            "Repair the battle causality: establish both opponents and their screen positions; every attack must name attacker, target, "
+            "contact or counter, the target's visible physical reaction, and the resulting positions carried into the next shot. Use an "
+            "unambiguous reaction such as 'is knocked backward,' 'blocks and slides two steps,' 'dodges and lands,' 'staggers,' or 'falls'; "
+            "a glow, charge, lunge, shockwave, impact sound, or mid-clash pose without the opponent's reaction is not a result."
+        )
+    if "minimax_h3_unresolved_final_state" in reasons:
+        corrections.append(
+            "Replace the unresolved ending with both relevant subjects in an observable post-action state; do not end mid-air, charging, "
+            "approaching, beginning an attack, mid-clash, or disappearing unless the brief explicitly requests that cliffhanger."
+        )
+    if "minimax_h3_orphaned_sound_event" in reasons:
+        corrections.append(
+            "Rebuild overall_soundscape from the final visible timeline after rewriting it. Delete any fire blast, gunshot, explosion, or "
+            "other event sound whose visible cause no longer exists; keep requested ambience and synchronized action sounds only."
+        )
+    if "minimax_h3_ref_missing_asset_tag" in reasons:
+        corrections.append(
+            "Keep every declared <Picture N>, <Video N>, and <Audio N> tag in subject_definitions. For identity-, appearance-, environment-, "
+            "or style-only assets, cite the asset inside the matching <Subject N> definition instead of creating a redundant standalone asset role."
+        )
+    if "minimax_h3_ref_subject_usage_invalid" in reasons:
+        corrections.append(
+            "Every defined <Subject N> must have its own retention_analysis line and must appear by that exact Subject tag in the shot timeline. "
+            "Never substitute its source <Picture N> or <Video N> tag for the Subject tag; at first appearance, write '<Subject N> (derived from "
+            "<Picture N>)' when provenance must also be explicit."
+        )
+    output_shape = {
+        "ltx_prompt": "",
+        "ideogram_prompt": "",
+        "minimax_h3_prompt": "REWRITTEN THREE-FIELD MINIMAX H3 PROMPT",
+        "negative_prompt": "",
+        "scene_segments": [],
+        "metadata": {"refined": True},
+    }
+    if minimax_h3_mode == "ref2va":
+        output_shape["minimax_h3_prompt"] = "REWRITTEN SIX-SECTION MINIMAX H3 REF2VA PROMPT"
+        retry_instruction = ""
+        if strict_retry:
+            retry_instruction = (
+                "This is the one bounded retry after an incomplete rewrite. Return no reasoning. Copy the candidate in full and make only "
+                "the listed corrections. Emit each exact heading once and end every heading line with its literal colon: subject_definitions:, "
+                "summary:, retention_analysis:, detailed_description:, overall_soundscape:, and non_diegetic_music:. Never write "
+                "detail_description, leave a section blank, "
+                "drop a shot, stop mid-sentence, or place any content after the final JSON boundary.\n\n"
+            )
+        return (
+            "You are the final quality-control rewrite pass for a MiniMax H3 Ref2VA full-reference prompt. Rewrite the candidate, preserving "
+            "its good concrete details, while fixing every listed failure. Do not explain or critique. Treat both the user brief and reference "
+            "manifest as binding; never invent, omit, renumber, merge, or cross-wire an asset label.\n\n"
+            f"User brief: {user_prompt.strip()}\n"
+            f"Exact duration: {duration_seconds:.3f} seconds\n"
+            f"Audio mode: {_normalize_audio_mode(audio_mode)}\n"
+            f"Reference manifest:\n{reference_manifest}\n"
+            f"Requested visual medium: {requested_medium or 'unspecified'}\n"
+            f"Requested style locks: {'; '.join(style_locks) if style_locks else 'none'}\n"
+            f"Failures to repair: {_json_dumps(reasons)}\n"
+            f"Required corrections: {' '.join(item for item in corrections if item)}\n\n"
+            "The rewritten minimax_h3_prompt must contain exactly these literal heading lines in order, each ending with its colon: "
+            "subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, and non_diegetic_music:. "
+            "Put one blank line between sections. In subject_definitions, map the "
+            "manifest assets to consecutive semantic <Subject N> labels or independently defined whole-asset roles. Write every definition as "
+            "'<Subject N>: description' (a colon immediately follows the tag). Begin summary with the exact "
+            "official bracketed task types. Give every independently defined visible label one fixed visual retention marker and every independently "
+            "defined Audio label one fixed audio marker. Once an element is assigned <Subject N>, use that exact Subject tag in the shot timeline and "
+            "use its <Picture N> or <Video N> source only as provenance, never as a replacement label. Put one or two concrete medium/style sentences before [Shot 1], then a detailed shot timeline "
+            "with stable reference labels, consecutive shots, increasing cut instants below the duration, explicit camera paths, causal continuity, "
+            "synchronized sound, and a resolved final state. Do not use integrated_multimodal_description or a negative section.\n\n"
+            f"{retry_instruction}"
+            f"Candidate prompt: {_json_dumps(candidate_prompt)}\n\n"
+            f"Return only {FINAL_JSON_OPEN}{_json_dumps(output_shape)}{FINAL_JSON_CLOSE} with the placeholder replaced by the complete rewritten prompt."
+        )
+    return (
+        "You are the final quality-control rewrite pass for a MiniMax H3 T2VA prompt. Rewrite the candidate, preserving its good concrete "
+        "details, while fixing every listed failure. Do not explain or critique. Do not add dialogue, visible text, subjects, or cuts that the "
+        "brief does not request. Treat the user brief as binding.\n\n"
+        f"User brief: {user_prompt.strip()}\n"
+        f"Exact duration: {duration_seconds:.3f} seconds\n"
+        f"Audio mode: {_normalize_audio_mode(audio_mode)}\n"
+        f"Requested visual medium: {requested_medium or 'unspecified'}\n"
+        f"Requested style locks: {'; '.join(style_locks) if style_locks else 'none'}\n"
+        f"Failures to repair: {_json_dumps(reasons)}\n"
+        f"Required corrections: {' '.join(item for item in corrections if item)}\n\n"
+        "The rewritten minimax_h3_prompt must contain only integrated_multimodal_description, overall_soundscape, and non_diegetic_music in "
+        "that order with one blank line between fields. Begin directly with integrated_multimodal_description: [Shot 1]. Shot 1 has no "
+        "timestamp; every later cut begins [Shot N] At MM:SS.mmm, the camera cuts to... with consecutive labels and increasing times below "
+        "the exact duration. Keep stable subject labels, designs, props, and screen geography. Put event sounds beside their visible causes. "
+        "End with 'The final frame holds on ...' followed by a resolved, observable subject state and camera composition.\n\n"
+        f"Candidate prompt: {_json_dumps(candidate_prompt)}\n\n"
+        f"Return only {FINAL_JSON_OPEN}{_json_dumps(output_shape)}{FINAL_JSON_CLOSE} with the placeholder replaced by the complete rewritten prompt."
+    )
+
+
 def _build_model_prompt(
     user_prompt: str,
     master_prompt: str,
@@ -2317,14 +4124,36 @@ def _build_model_prompt(
     negative_prompt_mode: str = "auto",
     negative_prompt_guidance: str = "",
     thinking_mode: str = "off",
+    minimax_h3_mode: str = "",
+    minimax_h3_reference_manifest: str = "",
 ) -> str:
     target_profile = _normalize_target_profile(target_profile)
+    minimax_h3_mode = _normalize_minimax_h3_mode(
+        minimax_h3_mode or str(media_metadata.get("minimax_h3_mode", "t2va"))
+    )
+    minimax_h3_reference_manifest = _normalize_minimax_h3_reference_manifest(
+        minimax_h3_reference_manifest or str(media_metadata.get("minimax_h3_reference_manifest", ""))
+    )
+    minimax_h3_reference_definitions = _minimax_h3_reference_definitions(minimax_h3_reference_manifest)
+    minimax_h3_reference_tags = [tag for tag, _description in minimax_h3_reference_definitions]
     audio_mode = _normalize_audio_mode(audio_mode)
     audio_guidance = _audio_guidance_text(audio_guidance)
-    audio_instruction = _audio_instruction(audio_mode, audio_guidance)
+    audio_instruction = _audio_instruction(audio_mode, audio_guidance, target_profile)
     negative_prompt_mode = _normalize_negative_prompt_mode(negative_prompt_mode)
     negative_prompt_guidance = _negative_prompt_guidance_text(negative_prompt_guidance)
     negative_instruction = _negative_prompt_instruction(negative_prompt_mode, negative_prompt_guidance)
+    if target_profile == "minimax_h3":
+        h3_description_field = "detailed_description" if minimax_h3_mode == "ref2va" else "integrated_multimodal_description"
+        if negative_prompt_mode == "empty":
+            negative_instruction = (
+                "Return negative_prompt as an empty string and ignore negative_prompt_guidance. Preserve only exclusions stated explicitly "
+                f"in the user's creative brief, inside {h3_description_field}."
+            )
+        else:
+            negative_instruction = (
+                "Return negative_prompt as an empty string. Fold only relevant requested exclusions into concise natural prose inside "
+                f"{h3_description_field}; do not create a negative section or Stable Diffusion-style defect list."
+            )
     thinking_mode = _normalize_thinking_mode(thinking_mode)
     thinking_instruction = _thinking_instruction(thinking_mode)
     controls = _target_context(
@@ -2340,8 +4169,44 @@ def _build_model_prompt(
         negative_prompt_mode,
         negative_prompt_guidance,
     )
+    minimax_h3_duration = _minimax_h3_duration_seconds(media_metadata, target_duration_seconds, user_prompt)
+    minimax_h3_requested_medium = _minimax_h3_requested_visual_medium(user_prompt)
+    minimax_h3_medium_instruction = _minimax_h3_visual_medium_instruction(minimax_h3_requested_medium)
+    minimax_h3_style_locks = _minimax_h3_requested_style_locks(user_prompt)
+    minimax_h3_style_instruction = ""
+    if minimax_h3_style_locks:
+        style_location = (
+            "inside the one-or-two-sentence style opening before [Shot 1]"
+            if minimax_h3_mode == "ref2va"
+            else "within the first 600 characters after [Shot 1]"
+        )
+        minimax_h3_style_instruction = (
+            f"Requested style locks: {'; '.join(minimax_h3_style_locks)}. Copy each phrase, allowing only spelling normalization, "
+            f"{style_location}; do not leave a requested aesthetic, era, or genre merely implied by props."
+        )
+    if target_profile == "minimax_h3":
+        controls["minimax_h3_mode"] = minimax_h3_mode
+        controls["minimax_h3_planning_duration_seconds"] = round(minimax_h3_duration, 3)
+        controls["minimax_h3_duration_advisory"] = _minimax_h3_duration_advisory(minimax_h3_duration)
+        if minimax_h3_mode == "ref2va":
+            controls["minimax_h3_reference_manifest"] = minimax_h3_reference_manifest
+            controls["minimax_h3_reference_tags"] = minimax_h3_reference_tags
+            controls["minimax_h3_reference_manifest_reasons"] = _minimax_h3_reference_manifest_validation_reasons(
+                minimax_h3_reference_manifest
+            )
+        if minimax_h3_requested_medium:
+            controls["minimax_h3_requested_visual_medium"] = minimax_h3_requested_medium
+            controls["minimax_h3_visual_medium_lock"] = minimax_h3_medium_instruction
+        if minimax_h3_style_locks:
+            controls["minimax_h3_requested_style_locks"] = minimax_h3_style_locks
     creativity_instruction = _creativity_instruction(creativity_mode, creative_strength)
-    structure_anchor = _target_structure_anchor(target_profile, ideogram_json_output)
+    if target_profile == "minimax_h3":
+        creativity_instruction = (
+            f"{creativity_instruction} For MiniMax H3, every added detail must be visibly or audibly realizable inside the selected duration. "
+            "Do not add dialogue, singing, visible text, logos, extra subjects, or extra cuts unless requested. "
+            f"{minimax_h3_medium_instruction} {minimax_h3_style_instruction}"
+        )
+    structure_anchor = _target_structure_anchor(target_profile, ideogram_json_output, minimax_h3_mode)
     media_grounding_mode = _media_grounding_mode(media_metadata)
     media_grounding_instruction = _media_grounding_instruction(media_metadata)
     ideogram_shape: Any
@@ -2379,6 +4244,8 @@ def _build_model_prompt(
         )
     if target_profile == "ltx":
         ideogram_shape = "Return an empty string for LTX target profile."
+    elif target_profile == "minimax_h3":
+        ideogram_shape = "Return an empty string for MiniMax H3 target profile."
     ltx_shape: Any = (
         "Final LTX positive prompt only. Single continuous natural-English paragraph, usually 80-180 words. "
         "Start directly with style if useful, then the action. Use active present/progressive tense, chronological flow, literal visible/audible detail, subject appearance, setting, motion, lighting, and camera behavior. "
@@ -2393,11 +4260,91 @@ def _build_model_prompt(
     )
     if target_profile == "ideogram4":
         ltx_shape = "Return an empty string for Ideogram4 target profile."
+    elif target_profile == "minimax_h3":
+        ltx_shape = "Return an empty string for MiniMax H3 target profile."
+    minimax_h3_shape: Any = "Return an empty string unless the target profile is minimax_h3."
+    if target_profile == "minimax_h3" and minimax_h3_mode == "ref2va":
+        minimax_h3_shape = (
+            f"Final MiniMax H3 Ref2VA full-reference prompt only, planned for exactly {minimax_h3_duration:.3f} seconds. The JSON string must "
+            "contain exactly six English sections, in this exact order with one blank line between sections: subject_definitions, summary, "
+            "retention_analysis, detailed_description, overall_soundscape, non_diegetic_music. Every section heading must be its own literal "
+            "colon-terminated line: subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, and "
+            "non_diegetic_music:. Do not use the T2VA "
+            "integrated_multimodal_description field. The declared reference manifest is authoritative: use every declared asset tag with exactly "
+            "the listed spelling and number, never invent an undeclared asset tag, never omit an asset, and never silently exchange the jobs of two assets. "
+            "In subject_definitions, put each separately tracked item on its own line. Create consecutive semantic <Subject N> labels for visible "
+            "people, creatures, objects, environments, costumes, styles, actions, or effects that will recur in the target. A <Subject N> denotes "
+            "reusable visible content, not an attachment slot; cite the <Picture N> or <Video N> that supplies each part of its identity, appearance, "
+            "motion, environment, or style. Once visible content has a <Subject N> label, use that exact Subject label in every shot where it appears; "
+            "the source asset tag may accompany its first appearance as provenance but must never replace the Subject label. Define a standalone <Picture N> only when the image itself is a concrete composition or storyboard anchor. "
+            "Define a standalone <Video N> for whole-video editing, continuation, camera, cut, rhythm, or temporal-structure roles. Define every "
+            "<Audio N> that is copied or referenced, and bind a voice reference to its <Subject N> (Sx) when it belongs to a visible speaker. "
+            "For this local Ref2VA path, a Picture used only for identity, wardrobe, environment, or style is a semantic reference, not a guaranteed "
+            "first-frame or last-frame pixel lock; claim keyframe completion only when the manifest explicitly assigns a concrete frame-anchor role. "
+            "In summary, write one short paragraph beginning with one square-bracketed combination of only these exact task types, joined by ' + ': "
+            "keyframe completion, reference generation, video editing, video continuation, audio reuse, audio reference. Select task types from actual "
+            "roles rather than asset presence, then name the main <Subject N> and asset relationships without adding new labels. "
+            "In retention_analysis, write one line for every independently defined label. For <Subject N>, <Picture N>, and <Video N>, use exactly one "
+            "of fully_preserved, partially_preserved, attribute_transfer, weak_reference, followed by ' - ' and a concrete explanation of what remains "
+            "and where it appears. For <Audio N>, use exactly one of fully_copy, partially_copy, reference, weak_reference. Do not put speaker IDs here. "
+            "In detailed_description, first establish the immutable rendering medium and requested visual style in one or two natural-English sentences "
+            "before [Shot 1]. Then write a normally 350-500-word shot-by-shot playback description for generation tasks, scaling video edits to source "
+            "complexity. Shot 1 has no timestamp. Every later cut begins [Shot N] At MM:SS.mmm, followed by a natural-English cut phrase, with "
+            "consecutive labels and strictly increasing cut instants below the selected duration; never use time ranges. Insert each reference label at "
+            "its first clear appearance and wherever its declared role takes effect. At first appearance, concretely restate each important referenced "
+            "subject's visible identity, position, action, lighting, and interaction rather than relying on the analysis sections as a plot summary. "
+            "Silently track stable subject identity, design, wardrobe, props, screen geography, causal state changes, and final state across cuts. Give every "
+            "shot a concrete camera path or an explicit locked/static hold. Preserve exact requested dialogue in <d>[Language] ...</d> with stable (Sx) "
+            "speakers; cite <Audio N> only where its copy/reference relationship becomes audible. Put synchronized actions and sounds beside their visible causes. "
+            "overall_soundscape summarizes ambience and physical sounds without repeating dialogue or audience-only music; non_diegetic_music specifies "
+            "audience-only instrumentation, tempo, and dynamic development. Cite an <Audio N> in the matching audio section when that layer is copied or "
+            "referenced. Use N/A for an absent layer. End on a resolved, observable final subject state and composition. Do not emit Markdown, extra headings, "
+            "a negative section, schema explanations, legacy Director commands, or any reference label that the manifest does not support."
+        )
+    elif target_profile == "minimax_h3":
+        minimax_h3_shape = (
+            f"Final MiniMax H3 T2VA prompt only, planned for exactly {minimax_h3_duration:.3f} seconds. The JSON string must literally begin "
+            "integrated_multimodal_description: [Shot 1] and contain exactly this three-field skeleton with one blank line between fields: "
+            "integrated_multimodal_description: [Shot 1] ...; overall_soundscape: ...; non_diegetic_music: ... . "
+            "Shot 1 has no timestamp. Later shots must be consecutive and begin "
+            "[Shot N] At MM:SS.mmm, followed by 'the camera cuts to' or another official natural-English cut phrase; default to hard cuts, and use a "
+            "cross-dissolve, fade, or wipe only when the user explicitly requests that effect. Use strictly increasing cut instants below "
+            "the selected duration, never time ranges. Convert user-supplied ranges into cut instants and resolve overlaps chronologically. Use the fewest shots "
+            "that fully express the request, and scale action/dialogue density to the available time, including "
+            "user-selected durations longer than 15 seconds without shortening or clamping them. Treat the user brief as a binding directing specification, "
+            "not a loose theme list. Silently make a continuity ledger before writing: immutable rendering medium; exact subject names or stable labels; each "
+            "subject's type, silhouette, face, colors, markings, wardrobe, and props; initial spatial relationship; location state; causal state after every beat; "
+            "and the requested final state. Do not output the ledger. Preserve every requested aesthetic, era, and genre phrase; an 'X-inspired' request is an "
+            "aesthetic lock, not permission to replace invented or fantasy creatures with generic real animals. Every recurring subject keeps the same label and "
+            "defining traits after every cut. Each shot must begin from the prior shot's resulting position and state, then show a chronological cause, preparation, "
+            "path of action, impact or counteraction, physical reaction, and observable result as appropriate. For a battle, establish both opponents and their "
+            "screen geography before the first attack, show who targets whom, and carry every attack's consequence into the next beat. "
+            "Embed one coherent camera path per shot using zoom, push or pull, pan or truck, tilt or pedestal, arc, tracking, static hold, shake, POV, or roll, "
+            "including direction, speed, and amplitude when meaningful. Shot size, angle, or the word handheld alone is not a camera path. Use a cut only when it "
+            "reveals genuinely new subject, space, state, viewpoint, or time. "
+            "Preserve exact user dialogue with stable speaker IDs such as (S1) and tags such as <d>[English] Exact words.</d>; do not invent dialogue. For "
+            "voiceover, use 'says in an off-screen voiceover' and state that the corresponding on-screen lips remain closed. Use <scenetrans> only when speech "
+            "continues across a cut and <cutoff> only for deliberate truncation at the end. "
+            "Put synchronized dialogue, diegetic music, and every event-specific sound immediately beside its visible cause inside the integrated timeline; "
+            "overall_soundscape may summarize ambience and Foley but cannot be the only place an action sound appears. Write 1-4 sentences of "
+            "ambience, Foley, and non-verbal sounds in overall_soundscape, without repeating dialogue or music. Write 1-3 sentences in non_diegetic_music "
+            "that specify audience-only instrumentation, tempo/rhythm, and dynamic changes tied to the requested cut or action beats, or N/A when there is no "
+            "score; avoid abstract mood and explanations of the score's emotional purpose. End the last shot with a sentence beginning exactly 'The final frame "
+            "holds on ...' that names the observable resulting subject state and camera composition. Resolve the final action instead of starting a new action, "
+            "vaguely disappearing, or ending on 'about to,' 'prepares to,' or 'begins to' unless the user explicitly requests an unresolved cliffhanger. "
+            "Preserve intentional visible text verbatim in double quotes. Fold only relevant exclusions "
+            "into concise natural prose inside integrated_multimodal_description. Do not emit extra headings, Markdown, reference tags, a negative section, "
+            "plot-summary filler, overlapping ranges, legacy Hailuo Director bracket commands, or explanations. Before returning, self-check that every requested "
+            "subject, action, camera beat, sound beat, exclusion, dialogue line, and visible text item is represented once in the correct timeline position."
+        )
     contract = {
         "ltx_prompt": ltx_shape,
         "ideogram_prompt": ideogram_shape,
-        "negative_prompt": "Final negative prompt string following the requested negative prompt mode.",
-        "scene_segments": [
+        "minimax_h3_prompt": minimax_h3_shape,
+        "negative_prompt": "Return an empty string for minimax_h3; otherwise follow the requested negative prompt mode.",
+        "scene_segments": []
+        if target_profile == "minimax_h3"
+        else [
             {
                 "index": 0,
                 "duration_seconds": 0,
@@ -2411,36 +4358,71 @@ def _build_model_prompt(
             "controls": controls,
         },
     }
+    if target_profile == "ideogram4":
+        profile_rules = (
+            "- Ideogram JSON must follow its caption schema: top-level high_level_description, style_description, compositional_deconstruction; compositional_deconstruction contains background then elements; elements use type=obj/text and desc/text keys; no unknown layout/aspect_ratio keys inside the caption.\n"
+            "- Ideogram color_palette values must be uppercase #RRGGBB hex strings if included; omit color_palette when no hex palette is known.\n"
+            "- Use Ideogram aspect ratio as layout guidance only; do not include aspect_ratio as a JSON caption key.\n"
+            "- Preserve exact quoted or user-specified text verbatim for Ideogram text elements.\n"
+            "- Respect the selected negative prompt mode.\n"
+        )
+    elif target_profile == "minimax_h3" and minimax_h3_mode == "ref2va":
+        profile_rules = (
+            "- Use duration as pacing guidance. Never clamp the selected or explicitly written duration to 15 seconds; the locally exposed frame count is user-controlled, while the model's published trained range remains an advisory. Every cut instant must fit the actual planning duration.\n"
+            "- Treat the reference manifest as an exact asset-role contract. Asset ordinals are one-based per type. Never invent, drop, renumber, merge, or cross-wire <Picture N>, <Video N>, or <Audio N> labels.\n"
+            "- Use <Subject N> for semantic visible content and use asset labels for provenance or whole-asset roles. An image that only supplies identity or style belongs inside a Subject definition; do not falsely call it a first/last-frame lock.\n"
+            "- Every defined <Subject N> must appear by that exact tag in retention_analysis and in the shot timeline. A source asset tag may accompany the Subject at first appearance, but it never replaces the Subject tag.\n"
+            "- Return exactly subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, and non_diegetic_music in that order, separated by one blank line. End every literal heading line with its colon: subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, non_diegetic_music:.\n"
+            "- Use only official summary task types and fixed retention markers. Explain exact preservation, transfer, copy, or reference behavior for each independently defined item.\n"
+            "- Put one or two medium/style sentences before [Shot 1]. Then keep consecutive shot labels, natural-English camera direction, strictly increasing cut instants, stable subjects, causal continuity, and a resolved final state.\n"
+            "- Return negative_prompt empty and fold only relevant user-requested exclusions into detailed_description without a negative section.\n"
+            "- Preserve exact dialogue and visible text verbatim. Dialogue belongs only inside <d>[Language] ...</d> with a stable (Sx) speaker ID; soundscape must not repeat dialogue or music.\n"
+            "- For visual_only audio mode, set both overall_soundscape and non_diegetic_music to N/A and do not use reference audio in the target timeline.\n"
+        )
+    elif target_profile == "minimax_h3":
+        profile_rules = (
+            "- Use duration as pacing guidance. Never clamp the selected or explicitly written duration to 15 seconds; 4-15 seconds is advisory only, and every cut instant must fit the actual planning duration.\n"
+            "- Treat the requested rendering medium, aesthetic, era, and genre as immutable. State the medium in the first words after [Shot 1]; cinematic direction controls staging and camera work inside that medium and never overrides it.\n"
+            "- Plan a silent continuity ledger for subject designs, props, spatial relationships, causal state changes, and the final state. Reuse stable subject labels and carry each shot's observable result into the next shot.\n"
+            "- Return negative_prompt empty and fold only relevant requested exclusions into integrated_multimodal_description without a negative section.\n"
+            "- Preserve exactly one integrated_multimodal_description field, one overall_soundscape field, and one non_diegetic_music field in that order, separated by one blank line.\n"
+            "- Shot 1 has no timestamp. Every later shot uses a strictly increasing MM:SS.mmm cut instant inside the selected duration; do not emit start-end ranges or overlapping ranges. Default to hard cuts; use a cross-dissolve, fade, or wipe only when explicitly requested.\n"
+            "- Give every shot a concrete camera path or explicit static hold; framing, angle, handheld, and 'the camera cuts' do not substitute for the camera path. End the last shot with a concrete final-frame composition and resolved visible state.\n"
+            "- Preserve exact dialogue and visible text verbatim. Dialogue belongs only inside <d>[Language] ...</d> with a stable (Sx) speaker ID; soundscape must not repeat dialogue or music.\n"
+            "- Place each event-specific sound beside the visible event that causes it inside integrated_multimodal_description; overall_soundscape is only the scene-wide ambience and Foley summary.\n"
+            "- non_diegetic_music specifies instrumentation, tempo/rhythm, and dynamic development, not abstract mood or the score's emotional purpose; use N/A when no score is present.\n"
+            "- For visual_only audio mode, set both overall_soundscape and non_diegetic_music to N/A and do not add dialogue, singing, ambience, Foley, or music.\n"
+        )
+    else:
+        profile_rules = (
+            "- Respect the selected negative prompt mode.\n"
+            "- If media_synthesis_mode is image_identity_video_control, the reference image controls subject identity and appearance: face, hair, body type, wardrobe, accessories, styling, and distinguishing visual traits. The video controls pose, depth, canny/edge layout, composition, motion, timing, camera choreography, blocking, and scene geometry. The LTX prompt must describe the reference-image subject performing the video action/control structure.\n"
+            "- In image_identity_video_control, do not copy the control-video subject's wardrobe, hair, face, body type, accessories, or styling into the LTX prompt unless the user explicitly asks for those video-subject traits. If the image subject and video subject differ, replace the video subject appearance with the reference-image appearance.\n"
+            "- For LTX with source video evidence, perform camera-choreography extraction before finalizing: first-frame framing, shot size, camera height/angle, camera support or stability, subject-camera relationship, movement type and direction, zoom/dolly/truck/boom/pan/tilt/roll/orbit/rotation, focus behavior, lens/depth changes, parallax, and final frame.\n"
+            "- Preserve complex visible camera moves as continuous relationships, not generic labels. For example, if the source zooms out while rotating 180 degrees and tracking the actress to the other side of the room, say plainly where the camera starts, how it moves relative to her, and where it settles.\n"
+            "- Preserve the clip chronologically by keeping the opening framing, any visible camera motion or stable hold, and the closing framing. If the source does not visibly zoom, pan, reveal, pull back, widen, rotate, orbit, or stage a new action beat, do not invent one.\n"
+        )
     return (
         f"{master_prompt.strip()}\n\n"
         "Return the final answer as JSON matching this shape:\n"
         f"{_json_dumps(contract)}\n\n"
         "Critical output rules:\n"
-        "- Prompt fields must be ready to paste directly into LTX-2.3 or Ideogram4.\n"
+        "- Prompt fields must be ready to paste directly into LTX-2.3, Ideogram4, or MiniMax H3.\n"
         "- Interpret User raw prompt as the creative brief for the target output, even when it is only a subject, moment, style, or rough idea. Supply the missing generator-prompt details yourself.\n"
-        "- Only write the selected target profile. For ltx, return ideogram_prompt as an empty string. For ideogram4, return ltx_prompt as an empty string and scene_segments as an empty list.\n"
+        "- Only write the selected target profile. For ltx, return ideogram_prompt and minimax_h3_prompt as empty strings. For ideogram4, return ltx_prompt and minimax_h3_prompt as empty strings and scene_segments as an empty list. For minimax_h3, return ltx_prompt, ideogram_prompt, and negative_prompt as empty strings and scene_segments as an empty list.\n"
         "- Do not put writing instructions, schema explanations, or adapter instructions inside any prompt field.\n"
         f"- Final response boundary: start the final answer with {FINAL_JSON_OPEN}, write one valid JSON object, close with {FINAL_JSON_CLOSE}, and write nothing after the closing marker.\n"
-        "- Use duration as pacing guidance for the LTX prompt, but do not write 'Target duration' or duration metadata in the prompt text.\n"
-        "- Ideogram JSON must follow its caption schema: top-level high_level_description, style_description, compositional_deconstruction; compositional_deconstruction contains background then elements; elements use type=obj/text and desc/text keys; no unknown layout/aspect_ratio keys inside the caption.\n"
-        "- Ideogram color_palette values must be uppercase #RRGGBB hex strings if included; omit color_palette when no hex palette is known.\n"
-        "- Use Ideogram aspect ratio as layout guidance only; do not include aspect_ratio as a JSON caption key.\n"
-        "- Preserve exact quoted or user-specified text verbatim for Ideogram text elements.\n"
         "- Apply creative direction by adding specific visual details, not by adding meta commentary about creativity.\n"
-        "- Respect negative prompt mode exactly; custom negatives must be returned unchanged, and empty mode must return an empty string.\n"
-        "- Do not invent dialogue unless the user asked for speech, talking, singing, or conversation.\n\n"
+        "- Do not invent dialogue unless the user asked for speech, talking, singing, or conversation.\n"
+        f"{profile_rules}\n"
         "- Preserve source lighting and time of day as non-negotiable visual identity when media or visual_description is supplied. If the source is night, the final prompt must repeatedly and concretely say nighttime/after dark, visible artificial light sources, deep shadows, and no daylight/sunlit/blue-sky conditions unless those are visibly present.\n"
         "- Preserve source camera/capture language as non-negotiable visual identity when media or visual_description is supplied. Use concrete industry terms when visible or strongly implied: overexposed highlights, clipped whites, underexposed shadows, direct flash, bloom, halation, shallow depth of field, soft focus, lens blur, bokeh, motion blur, focal length impression, wide-angle distortion, telephoto compression, high ISO grain/noise, white balance, high contrast, low dynamic range, and camera angle.\n"
-        "- If media_synthesis_mode is image_identity_video_control, the reference image controls subject identity and appearance: face, hair, body type, wardrobe, accessories, styling, and distinguishing visual traits. The video controls pose, depth, canny/edge layout, composition, motion, timing, camera choreography, blocking, and scene geometry. The LTX prompt must describe the reference-image subject performing the video action/control structure.\n"
-        "- In image_identity_video_control, do not copy the control-video subject's wardrobe, hair, face, body type, accessories, or styling into the LTX prompt unless the user explicitly asks for those video-subject traits. If the image subject and video subject differ, replace the video subject appearance with the reference-image appearance.\n"
-        "- For LTX with source video evidence, perform camera-choreography extraction before finalizing: first-frame framing, shot size, camera height/angle, camera support or stability, subject-camera relationship, movement type and direction, zoom/dolly/truck/boom/pan/tilt/roll/orbit/rotation, focus behavior, lens/depth changes, parallax, and final frame.\n"
-        "- Preserve complex visible camera moves as continuous relationships, not generic labels. For example, if the source zooms out while rotating 180 degrees and tracking the actress to the other side of the room, the LTX prompt should say that plainly with where the camera starts, how it moves relative to her, and where it settles.\n"
-        "- For LTX with source video evidence, preserve the clip chronologically by keeping the opening framing, any visible camera motion or stable hold, and the closing framing. If the source does not visibly zoom, pan, reveal, pull back, widen, rotate, orbit, or stage a new action beat, do not invent one.\n"
         "- If Media grounding mode is metadata_only, do not claim to see or describe the supplied image/video pixels. Dimensions, frame counts, and sample indices are not visual evidence.\n"
         "- If Media grounding mode is visual_description, treat the visual_description as the only visual evidence and make the final prompt specific to it.\n"
         "- If Media grounding mode is pixels, ground observations in the attached pixels.\n\n"
         f"Target profile: {target_profile}\n"
         f"Target controls: {_json_dumps(controls)}\n"
+        f"MiniMax H3 reference manifest: {minimax_h3_reference_manifest if minimax_h3_mode == 'ref2va' else 'not used'}\n"
         f"Media grounding mode: {media_grounding_mode}\n"
         f"Media grounding instruction: {media_grounding_instruction}\n"
         f"Creative direction: {creativity_instruction}\n"
@@ -2518,6 +4500,10 @@ def _template_packet(
     )
     art_direction = _creative_art_direction_phrase(creativity_mode, creative_strength)
     effective_duration = _effective_duration_seconds(media_metadata, target_duration_seconds) or duration_seconds
+    if target_profile == "minimax_h3":
+        h3_duration = _minimax_h3_duration_seconds(media_metadata, target_duration_seconds, user_prompt)
+        controls["minimax_h3_planning_duration_seconds"] = round(h3_duration, 3)
+        controls["minimax_h3_duration_advisory"] = _minimax_h3_duration_advisory(h3_duration)
     media_hint = ""
     if media_grounding_mode == "visual_description":
         media_hint = "The prompt is grounded in the supplied visual description rather than direct pixel analysis."
@@ -2561,6 +4547,49 @@ def _template_packet(
     if audio_hint and not limitation_prompt:
         ltx_clauses.append(audio_hint)
     ltx = _single_paragraph(" ".join(ltx_clauses))
+    h3_clauses = [_ensure_terminal_period(base)]
+    if media_hint and not limitation_prompt:
+        h3_clauses.append(media_hint)
+    if art_direction and not limitation_prompt:
+        h3_clauses.append(art_direction)
+    if negative_prompt_guidance and negative_prompt_mode != "empty" and not limitation_prompt:
+        h3_clauses.append(f"Exclude these unwanted elements or behaviors: {negative_prompt_guidance}.")
+    h3_integrated = _single_paragraph(" ".join(h3_clauses))
+    if audio_mode == "visual_only":
+        h3_soundscape = "N/A"
+        h3_music = "N/A"
+    else:
+        music_terms = r"\b(music|score|soundtrack|orchestra|orchestral|percussion|piano|guitar|strings|synth)\b"
+        guidance_is_musical = bool(re.search(music_terms, audio_guidance, flags=re.IGNORECASE))
+        h3_soundscape = (
+            "Natural ambience and physical sounds follow the visible setting and actions."
+            if guidance_is_musical
+            else audio_guidance or "Natural ambience and physical sounds follow the visible setting and actions."
+        )
+        score_requested = bool(
+            re.search(
+                music_terms,
+                f"{user_prompt} {audio_guidance}",
+                flags=re.IGNORECASE,
+            )
+        )
+        if guidance_is_musical:
+            h3_music = audio_guidance
+        elif score_requested:
+            h3_music = "A restrained instrumental score uses measured percussion at a moderate tempo and builds dynamically with the action."
+        else:
+            h3_music = "N/A"
+    minimax_h3 = _sanitize_structured_prompt_text(
+        "\n\n".join(
+            (
+                f"integrated_multimodal_description: [Shot 1] {h3_integrated}",
+                f"overall_soundscape: {h3_soundscape}",
+                f"non_diegetic_music: {h3_music}",
+            )
+        ),
+        "",
+        8000,
+    )
     synthesis_metadata: dict[str, Any] = {}
     _copy_synthesis_metadata(synthesis_metadata, media_metadata)
 
@@ -2582,13 +4611,21 @@ def _template_packet(
         )
     if target_profile == "ltx":
         ideogram = ""
+        minimax_h3 = ""
     elif target_profile == "ideogram4":
         ltx = ""
-    segments = _split_segments(ltx, effective_duration) if ltx else []
+        minimax_h3 = ""
+    else:
+        ltx = ""
+        ideogram = ""
+    segments = _split_segments(ltx, effective_duration) if ltx and target_profile == "ltx" else []
     return {
         "ltx_prompt": ltx,
         "ideogram_prompt": ideogram,
-        "negative_prompt": _apply_negative_prompt_policy("", negative_prompt_mode, negative_prompt_guidance),
+        "minimax_h3_prompt": minimax_h3,
+        "negative_prompt": ""
+        if target_profile == "minimax_h3"
+        else _apply_negative_prompt_policy("", negative_prompt_mode, negative_prompt_guidance),
         "scene_segments": segments,
         "metadata": {
             "backend": "template",
@@ -2608,8 +4645,8 @@ def _template_packet(
     }
 
 
-def _extract_labeled_prompt(text: str, label: str) -> str:
-    labels = r"(?:ltx_prompt|ideogram_prompt|negative_prompt|scene_segments|metadata)"
+def _extract_labeled_prompt(text: str, label: str, structured: bool = False) -> str:
+    labels = r"(?:ltx_prompt|ideogram_prompt|minimax_h3_prompt|negative_prompt|scene_segments|metadata)"
     pattern = rf"['\"]?\b{re.escape(label)}\b['\"]?\s*[:=]\s*(.*?)(?=['\"]?\b{labels}\b['\"]?\s*[:=]|$)"
     match = re.search(pattern, text, flags=re.IGNORECASE | re.DOTALL)
     if not match:
@@ -2622,6 +4659,8 @@ def _extract_labeled_prompt(text: str, label: str) -> str:
                 value = decoded
         except Exception:
             value = value.strip('"')
+    if structured:
+        return _sanitize_structured_prompt_text(value, "", 8000)
     return _sanitize_prompt_text(value, "", 8000)
 
 
@@ -2656,14 +4695,20 @@ def _packet_from_plain_text_output(
     negative_prompt_guidance: str,
     max_output_chars: int,
 ) -> dict[str, Any] | None:
-    text = _sanitize_prompt_text(raw_output, "", int(max_output_chars))
+    profile = _normalize_target_profile(target_profile)
+    text = (
+        _sanitize_structured_prompt_text(raw_output, "", int(max_output_chars))
+        if profile == "minimax_h3"
+        else _sanitize_prompt_text(raw_output, "", int(max_output_chars))
+    )
     if not text:
         return None
 
     ltx_labeled = _extract_labeled_prompt(text, "ltx_prompt")
     ideogram_labeled = _extract_labeled_prompt(text, "ideogram_prompt")
+    minimax_h3_labeled = _extract_labeled_prompt(text, "minimax_h3_prompt", structured=True)
     negative_labeled = _extract_labeled_prompt(text, "negative_prompt")
-    if not (ltx_labeled or ideogram_labeled or _has_salvageable_plain_text(text)):
+    if not (ltx_labeled or ideogram_labeled or minimax_h3_labeled or _has_salvageable_plain_text(text)):
         return None
     template = _template_packet(
         user_prompt,
@@ -2683,23 +4728,31 @@ def _packet_from_plain_text_output(
         negative_prompt_mode,
         negative_prompt_guidance,
     )
-    profile = _normalize_target_profile(target_profile)
     plain = _clean_prompt_request_text(ltx_labeled or ideogram_labeled or text)
     if profile == "ideogram4":
         template["ideogram_prompt"] = ideogram_labeled or plain
+    elif profile == "minimax_h3":
+        template["minimax_h3_prompt"] = _sanitize_structured_prompt_text(
+            minimax_h3_labeled or text,
+            "",
+            int(max_output_chars),
+        )
+        template["negative_prompt"] = ""
+        template["scene_segments"] = []
     else:
         template["ltx_prompt"] = ltx_labeled or plain
         if ideogram_labeled:
             template["ideogram_prompt"] = ideogram_labeled
-    template["negative_prompt"] = _apply_negative_prompt_policy(
-        negative_labeled or template.get("negative_prompt", ""),
-        negative_prompt_mode,
-        negative_prompt_guidance,
-    )
-    template["scene_segments"] = _split_segments(
-        str(template.get("ltx_prompt", "")),
-        _effective_duration_seconds(media_metadata, target_duration_seconds) or duration_seconds,
-    )
+    if profile != "minimax_h3":
+        template["negative_prompt"] = _apply_negative_prompt_policy(
+            negative_labeled or template.get("negative_prompt", ""),
+            negative_prompt_mode,
+            negative_prompt_guidance,
+        )
+        template["scene_segments"] = _split_segments(
+            str(template.get("ltx_prompt", "")),
+            _effective_duration_seconds(media_metadata, target_duration_seconds) or duration_seconds,
+        )
     metadata = template.setdefault("metadata", {})
     metadata.pop("fallback_reason", None)
     metadata.update(
@@ -2726,6 +4779,58 @@ def repair_or_salvage_prompt_packet(
     duration = _safe_float(media_metadata.get("duration_seconds"), 0.0)
     parsed = _extract_json_object(raw_output)
     if parsed is not None:
+        if target["target_profile"] == "minimax_h3":
+            h3_mode = target["minimax_h3_mode"]
+            reference_manifest = _normalize_minimax_h3_reference_manifest(
+                str(media_metadata.get("minimax_h3_reference_manifest", ""))
+            )
+            normalized_h3_prompt = _normalize_minimax_h3_prompt(
+                str(parsed.get("minimax_h3_prompt", "") or ""),
+                h3_mode,
+            )
+            if h3_mode == "ref2va":
+                normalized_h3_prompt = _repair_minimax_h3_ref2va_structure(
+                    normalized_h3_prompt,
+                    context.user_prompt,
+                    reference_manifest,
+                    _minimax_h3_duration_seconds(
+                        media_metadata,
+                        target["target_duration_seconds"],
+                        context.user_prompt,
+                    ),
+                )
+            medium_locked_h3_prompt = _apply_minimax_h3_visual_medium_lock(
+                normalized_h3_prompt,
+                context.user_prompt,
+                h3_mode,
+            )
+            style_locked_h3_prompt = _apply_minimax_h3_style_locks(
+                medium_locked_h3_prompt,
+                context.user_prompt,
+                h3_mode,
+            )
+            parsed["minimax_h3_prompt"] = _apply_minimax_h3_sound_fidelity(
+                style_locked_h3_prompt,
+                context.user_prompt,
+                h3_mode,
+            )
+            if not isinstance(parsed.get("metadata"), dict):
+                parsed["metadata"] = {}
+            parsed["metadata"]["minimax_h3_mode"] = h3_mode
+            if h3_mode == "ref2va":
+                parsed["metadata"]["minimax_h3_reference_tags"] = _minimax_h3_reference_tags(reference_manifest)
+            if medium_locked_h3_prompt != normalized_h3_prompt:
+                if not isinstance(parsed.get("metadata"), dict):
+                    parsed["metadata"] = {}
+                parsed["metadata"]["minimax_h3_visual_medium_lock_injected"] = True
+            if style_locked_h3_prompt != medium_locked_h3_prompt:
+                if not isinstance(parsed.get("metadata"), dict):
+                    parsed["metadata"] = {}
+                parsed["metadata"]["minimax_h3_style_locks_injected"] = True
+            if parsed["minimax_h3_prompt"] != style_locked_h3_prompt:
+                if not isinstance(parsed.get("metadata"), dict):
+                    parsed["metadata"] = {}
+                parsed["metadata"]["minimax_h3_sound_fidelity_repaired"] = True
         return parsed, True, ""
     salvaged = _packet_from_plain_text_output(
         raw_output,
@@ -2777,7 +4882,7 @@ def _packet_to_prompt_outputs(
     max_output_chars: int,
     creativity_mode: str = "editorial",
     creative_strength: float = 0.6,
-) -> tuple[str, str, str, str, str, str, str, str]:
+) -> tuple[str, str, str, str, str, str, str, str, str]:
     target = _target_profile_config_to_dict(target_config)
     profile = target["target_profile"]
     prompt_fallback = _clean_prompt_request_text(context.user_prompt)
@@ -2886,29 +4991,95 @@ def _packet_to_prompt_outputs(
             packet.setdefault("metadata", {})["visual_grounding_warning"] = (
                 "Generic media-description output was replaced because the active backend cannot inspect pixels."
             )
-    negative_prompt = _apply_negative_prompt_policy(
-        packet.get("negative_prompt", ""),
-        target["negative_prompt_mode"],
-        target["negative_prompt_guidance"],
-    )
+    minimax_h3_prompt = ""
+    if profile == "minimax_h3":
+        h3_mode = target["minimax_h3_mode"]
+        reference_manifest = _normalize_minimax_h3_reference_manifest(
+            str(media_metadata.get("minimax_h3_reference_manifest", ""))
+        )
+        packet_metadata["minimax_h3_mode"] = h3_mode
+        if h3_mode == "ref2va":
+            packet_metadata["minimax_h3_reference_tags"] = _minimax_h3_reference_tags(reference_manifest)
+        requested_medium = _minimax_h3_requested_visual_medium(context.user_prompt)
+        if requested_medium:
+            packet_metadata["minimax_h3_requested_visual_medium"] = requested_medium
+        requested_style_locks = _minimax_h3_requested_style_locks(context.user_prompt)
+        if requested_style_locks:
+            packet_metadata["minimax_h3_requested_style_locks"] = requested_style_locks
+        normalized_h3_prompt = _normalize_minimax_h3_prompt(
+            str(packet.get("minimax_h3_prompt", "") or ""),
+            h3_mode,
+        )
+        if h3_mode == "ref2va":
+            normalized_h3_prompt = _repair_minimax_h3_ref2va_structure(
+                normalized_h3_prompt,
+                context.user_prompt,
+                reference_manifest,
+                _minimax_h3_duration_seconds(
+                    media_metadata,
+                    target["target_duration_seconds"],
+                    context.user_prompt,
+                ),
+            )
+        medium_locked_h3_prompt = _apply_minimax_h3_visual_medium_lock(
+            normalized_h3_prompt,
+            context.user_prompt,
+            h3_mode,
+        )
+        style_locked_h3_prompt = _apply_minimax_h3_style_locks(
+            medium_locked_h3_prompt,
+            context.user_prompt,
+            h3_mode,
+        )
+        packet["minimax_h3_prompt"] = _apply_minimax_h3_sound_fidelity(
+            style_locked_h3_prompt,
+            context.user_prompt,
+            h3_mode,
+        )
+        if medium_locked_h3_prompt != normalized_h3_prompt:
+            packet_metadata["minimax_h3_visual_medium_lock_injected"] = True
+        if style_locked_h3_prompt != medium_locked_h3_prompt:
+            packet_metadata["minimax_h3_style_locks_injected"] = True
+        if packet["minimax_h3_prompt"] != style_locked_h3_prompt:
+            packet_metadata["minimax_h3_sound_fidelity_repaired"] = True
+        minimax_h3_prompt = _sanitize_structured_prompt_text(
+            packet["minimax_h3_prompt"],
+            "",
+            int(max_output_chars),
+        )
+    negative_prompt = ""
+    if profile != "minimax_h3":
+        negative_prompt = _apply_negative_prompt_policy(
+            packet.get("negative_prompt", ""),
+            target["negative_prompt_mode"],
+            target["negative_prompt_guidance"],
+        )
     scene_segments = packet.get("scene_segments")
-    if profile == "ideogram4":
+    if profile in {"ideogram4", "minimax_h3"}:
         scene_segments = []
     elif not isinstance(scene_segments, list) or not scene_segments:
         scene_segments = _split_segments(ltx_prompt, effective_duration or duration)
     local_prompts, segment_lengths = _segments_to_director_strings(scene_segments)
-    prompt_to_verify = ideogram_prompt if profile == "ideogram4" else ltx_prompt
+    if profile == "ideogram4":
+        prompt_to_verify = ideogram_prompt
+    elif profile == "minimax_h3":
+        prompt_to_verify = minimax_h3_prompt
+    else:
+        prompt_to_verify = ltx_prompt
     ready_for_generation, blocked_reasons = _packet_generation_readiness(
         packet,
         context,
         target,
         prompt_to_verify,
         media_metadata,
+        int(max_output_chars),
     )
     packet_metadata["ready_for_generation"] = bool(ready_for_generation)
     packet_metadata["blocked_reasons"] = blocked_reasons
     packet_metadata["claim_verification_passed"] = not any(
-        reason.startswith("metadata_only_") or reason.startswith("visual_description_contradiction")
+        reason.startswith("metadata_only_")
+        or reason.startswith("visual_description_contradiction")
+        or reason.startswith("minimax_h3_")
         for reason in blocked_reasons
     )
     return (
@@ -2920,6 +5091,7 @@ def _packet_to_prompt_outputs(
         segment_lengths,
         target["ideogram_aspect_ratio"],
         _json_dumps(packet.get("metadata") if isinstance(packet.get("metadata"), dict) else {}),
+        minimax_h3_prompt,
     )
 
 
@@ -3316,6 +5488,25 @@ def _transformers_messages_and_processor_kwargs(
     media_metadata = media_context.metadata if isinstance(media_context, MediaContext) else {}
     source = media_metadata.get("source", "none") if isinstance(media_metadata, dict) else "none"
     processor_kwargs: dict[str, Any] = {}
+    if pil_images and _is_minimax_h3_reference_context(media_metadata):
+        picture_count = int(max(0.0, _safe_float(media_metadata.get("minimax_h3_reference_image_batch_count"), 0.0)))
+        picture_count = min(picture_count, len(pil_images))
+        for index, image in enumerate(pil_images[:picture_count], start=1):
+            content.append(
+                {
+                    "type": "text",
+                    "text": f"<Picture {index}> pixel evidence. Apply only the role assigned to <Picture {index}> in the H3 reference manifest.",
+                }
+            )
+            content.append({"type": "image", "image": image})
+        video_images = pil_images[picture_count:]
+        if video_images:
+            content.append({"type": "text", "text": _sampled_video_frame_sequence_intro(media_metadata, len(pil_images))})
+            for image in video_images:
+                content.append({"type": "image", "image": image})
+            content.append({"type": "text", "text": "End of sampled <Video 1> analysis frames."})
+        content.append({"type": "text", "text": prompt})
+        return [{"role": "user", "content": content}], processor_kwargs
     if pil_images and source in {"video", "image+video"}:
         frame_count = len(pil_images)
         transport = str(media_metadata.get("transformers_video_transport", _video_transport_mode()))
@@ -3575,6 +5766,12 @@ def _run_generation_packet(
     thinking_mode = _normalize_thinking_mode(thinking_mode)
     media_notes = _media_policy_notes(model_config, context)
     media_metadata = context.media_metadata.copy()
+    if target["target_profile"] == "minimax_h3":
+        media_metadata["minimax_h3_mode"] = target["minimax_h3_mode"]
+        if target["minimax_h3_mode"] == "ref2va":
+            media_metadata["minimax_h3_reference_manifest"] = _normalize_minimax_h3_reference_manifest(
+                str(media_metadata.get("minimax_h3_reference_manifest", ""))
+            )
     pixels_present = bool(context.images is not None)
     pixels_sent = bool(pixels_present and _runtime_supports_pixels(model_config))
     visual_description = _clean_visual_description(context.visual_description or str(media_metadata.get("visual_description", "")))
@@ -3582,7 +5779,9 @@ def _run_generation_packet(
     media_metadata["pixels_sent_to_backend"] = pixels_sent
     if pixels_sent and str(media_metadata.get("source", "none")).lower() in {"video", "image+video"}:
         media_metadata["transformers_video_transport"] = (
-            "sampled_frame_images" if _is_image_identity_video_control(media_metadata) else _video_transport_mode()
+            "sampled_frame_images"
+            if _is_image_identity_video_control(media_metadata) or _is_minimax_h3_reference_context(media_metadata)
+            else _video_transport_mode()
         )
     if visual_description:
         media_metadata["visual_description"] = visual_description
@@ -3623,6 +5822,8 @@ def _run_generation_packet(
         target["negative_prompt_mode"],
         target["negative_prompt_guidance"],
         thinking_mode,
+        target["minimax_h3_mode"],
+        str(media_metadata.get("minimax_h3_reference_manifest", "")),
     )
     raw_output = ""
     answer_text = ""
@@ -3633,7 +5834,15 @@ def _run_generation_packet(
     salvage_warning = ""
     backend_seconds = 0.0
     postprocess_seconds = 0.0
+    refinement_backend_seconds = 0.0
+    refinement_report: dict[str, Any] = {}
     try:
+        if target["target_profile"] == "minimax_h3" and target["minimax_h3_mode"] == "ref2va":
+            manifest_reasons = _minimax_h3_reference_manifest_validation_reasons(
+                str(media_metadata.get("minimax_h3_reference_manifest", ""))
+            )
+            if manifest_reasons:
+                raise ValueError("MiniMax H3 Ref2VA manifest is invalid: " + ", ".join(manifest_reasons))
         if model_config.backend == "template":
             raise RuntimeError("Template backend selected.")
         backend_media_context = _media_context_from_gemma_context(working_context) if _runtime_supports_pixels(model_config) else None
@@ -3652,7 +5861,135 @@ def _run_generation_packet(
             creativity_mode,
             creative_strength,
         )
-        postprocess_seconds = time.perf_counter() - postprocess_start
+        if target["target_profile"] == "minimax_h3" and parse_valid:
+            h3_duration = _minimax_h3_duration_seconds(
+                media_metadata,
+                target["target_duration_seconds"],
+                working_context.user_prompt,
+            )
+            initial_h3_prompt = str(packet.get("minimax_h3_prompt", "") or "")
+            initial_h3_reasons = _minimax_h3_prompt_validation_reasons(
+                initial_h3_prompt,
+                h3_duration,
+                working_context.user_prompt,
+                target["audio_mode"],
+                int(max_output_chars),
+                target["minimax_h3_mode"],
+                str(media_metadata.get("minimax_h3_reference_manifest", "")),
+            )
+            refinable_reasons = [reason for reason in initial_h3_reasons if reason in _MINIMAX_H3_REFINABLE_REASONS]
+            if refinable_reasons:
+                refinement_report = {
+                    "attempted": True,
+                    "accepted": False,
+                    "trigger_reasons": refinable_reasons,
+                    "initial_reasons": initial_h3_reasons,
+                    "attempts": [],
+                }
+                refinement_candidate_prompt = initial_h3_prompt
+                refinement_candidate_reasons = initial_h3_reasons
+                refinement_base = "initial"
+                max_refinement_attempts = 2 if target["minimax_h3_mode"] == "ref2va" else 1
+                for refinement_index in range(max_refinement_attempts):
+                    refinement_prompt = _build_minimax_h3_refinement_prompt(
+                        working_context.user_prompt,
+                        refinement_candidate_prompt,
+                        h3_duration,
+                        target["audio_mode"],
+                        refinement_candidate_reasons,
+                        target["minimax_h3_mode"],
+                        str(media_metadata.get("minimax_h3_reference_manifest", "")),
+                        strict_retry=refinement_index > 0,
+                    )
+                    if refinement_index == 0:
+                        refinement_report["prompt_chars"] = len(refinement_prompt)
+                    attempt_report = {
+                        "attempt": refinement_index + 1,
+                        "base": refinement_base,
+                        "prompt_chars": len(refinement_prompt),
+                        "trigger_reasons": refinement_candidate_reasons,
+                    }
+                    refinement_report["attempts"].append(attempt_report)
+                    try:
+                        refinement_start = time.perf_counter()
+                        try:
+                            refined_raw_output = _run_backend(
+                                model_config,
+                                refinement_prompt,
+                                backend_media_context,
+                                int(max_new_tokens),
+                                node_id=node_id,
+                            )
+                        finally:
+                            attempt_backend_seconds = time.perf_counter() - refinement_start
+                            refinement_backend_seconds += attempt_backend_seconds
+                            backend_seconds += attempt_backend_seconds
+                            attempt_report["backend_seconds"] = round(attempt_backend_seconds, 3)
+                        refined_reasoning, refined_answer = _split_reasoning_and_answer(refined_raw_output)
+                        refined_packet, refined_parse_valid, refined_salvage_warning = repair_or_salvage_prompt_packet(
+                            refined_answer or refined_raw_output,
+                            working_context,
+                            target,
+                            int(max_output_chars),
+                            creativity_mode,
+                            creative_strength,
+                        )
+                        refined_h3_prompt = str(refined_packet.get("minimax_h3_prompt", "") or "")
+                        refined_h3_reasons = _minimax_h3_prompt_validation_reasons(
+                            refined_h3_prompt,
+                            h3_duration,
+                            working_context.user_prompt,
+                            target["audio_mode"],
+                            int(max_output_chars),
+                            target["minimax_h3_mode"],
+                            str(media_metadata.get("minimax_h3_reference_manifest", "")),
+                        )
+                        attempt_report["parse_valid"] = bool(refined_parse_valid)
+                        attempt_report["candidate_reasons"] = refined_h3_reasons
+                        refinement_report["candidate_reasons"] = refined_h3_reasons
+                        if refined_parse_valid and not refined_h3_reasons:
+                            initial_metadata = packet.get("metadata") if isinstance(packet.get("metadata"), dict) else {}
+                            refined_metadata = refined_packet.get("metadata") if isinstance(refined_packet.get("metadata"), dict) else {}
+                            if "observations" in initial_metadata and "observations" not in refined_metadata:
+                                refined_metadata["observations"] = initial_metadata["observations"]
+                            refined_packet["metadata"] = refined_metadata
+                            packet = refined_packet
+                            raw_output = refined_raw_output
+                            reasoning_text = refined_reasoning
+                            answer_text = refined_answer
+                            parse_valid = refined_parse_valid
+                            salvage_warning = refined_salvage_warning
+                            attempt_report["accepted"] = True
+                            refinement_report["accepted"] = True
+                            refinement_report["accepted_attempt"] = refinement_index + 1
+                            break
+
+                        rejected_prompt = _sanitize_structured_prompt_text(refined_h3_prompt, "", 6000)
+                        attempt_report["accepted"] = False
+                        attempt_report["rejected_candidate_prompt"] = rejected_prompt
+                        refinement_report["rejected_candidate_prompt"] = rejected_prompt
+                        if refinement_index + 1 >= max_refinement_attempts:
+                            break
+                        refinement_candidate_prompt = initial_h3_prompt
+                        refinement_candidate_reasons = list(dict.fromkeys([*initial_h3_reasons, *refined_h3_reasons]))
+                        refinement_base = "initial"
+                        refinement_report["retry_attempted"] = True
+                        refinement_report["retry_base"] = refinement_base
+                    except Exception as refinement_exc:
+                        attempt_report["error"] = str(refinement_exc)
+                        refinement_report["error"] = str(refinement_exc)
+                        _dg_log(
+                            "MiniMax H3 refinement pass %s failed; keeping the first candidate. error=%s",
+                            refinement_index + 1,
+                            refinement_exc,
+                        )
+                        break
+                refinement_report["backend_seconds"] = round(refinement_backend_seconds, 3)
+        if refinement_report:
+            if not isinstance(packet.get("metadata"), dict):
+                packet["metadata"] = {}
+            packet["metadata"]["minimax_h3_refinement"] = refinement_report
+        postprocess_seconds = max(0.0, time.perf_counter() - postprocess_start - refinement_backend_seconds)
     except Exception as exc:
         fallback_reason = str(exc)
         if model_config.backend == "transformers_inprocess":
@@ -3698,7 +6035,11 @@ def _run_generation_packet(
             "media": media_metadata,
             "audio_mode": target["audio_mode"],
             "audio_guidance": target["audio_guidance"],
-            "target_duration_seconds": _effective_duration_seconds(media_metadata, target["target_duration_seconds"]),
+            "target_duration_seconds": (
+                _minimax_h3_duration_seconds(media_metadata, target["target_duration_seconds"], working_context.user_prompt)
+                if target["target_profile"] == "minimax_h3"
+                else _effective_duration_seconds(media_metadata, target["target_duration_seconds"])
+            ),
             "ltx_style": target["ltx_style"],
             "ideogram_aspect_ratio": target["ideogram_aspect_ratio"],
             "ideogram_render_style": target["ideogram_render_style"],
@@ -3725,6 +6066,25 @@ def _run_generation_packet(
             },
         }
     )
+    if target["target_profile"] == "minimax_h3":
+        h3_duration = _minimax_h3_duration_seconds(
+            media_metadata,
+            target["target_duration_seconds"],
+            working_context.user_prompt,
+        )
+        metadata["minimax_h3_planning_duration_seconds"] = round(h3_duration, 3)
+        metadata["minimax_h3_duration_advisory"] = _minimax_h3_duration_advisory(h3_duration)
+        metadata["minimax_h3_mode"] = target["minimax_h3_mode"]
+        if target["minimax_h3_mode"] == "ref2va":
+            metadata["minimax_h3_reference_tags"] = _minimax_h3_reference_tags(
+                str(media_metadata.get("minimax_h3_reference_manifest", ""))
+            )
+        requested_medium = _minimax_h3_requested_visual_medium(working_context.user_prompt)
+        if requested_medium:
+            metadata["minimax_h3_requested_visual_medium"] = requested_medium
+        requested_style_locks = _minimax_h3_requested_style_locks(working_context.user_prompt)
+        if requested_style_locks:
+            metadata["minimax_h3_requested_style_locks"] = requested_style_locks
     _copy_synthesis_metadata(metadata, media_metadata)
     readiness_prompt = ""
     if target["target_profile"] == "ideogram4":
@@ -3739,6 +6099,12 @@ def _run_generation_packet(
             creativity_mode,
             creative_strength,
         )
+    elif target["target_profile"] == "minimax_h3":
+        readiness_prompt = _sanitize_structured_prompt_text(
+            str(packet.get("minimax_h3_prompt", "") or ""),
+            "",
+            int(max_output_chars),
+        )
     else:
         readiness_prompt = _prompt_field_to_text(
             packet.get("ltx_prompt", ""),
@@ -3751,11 +6117,14 @@ def _run_generation_packet(
         target,
         readiness_prompt,
         media_metadata,
+        int(max_output_chars),
     )
     metadata["ready_for_generation"] = bool(ready_for_generation)
     metadata["blocked_reasons"] = blocked_reasons
     metadata["claim_verification_passed"] = not any(
-        reason.startswith("metadata_only_") or reason.startswith("visual_description_contradiction")
+        reason.startswith("metadata_only_")
+        or reason.startswith("visual_description_contradiction")
+        or reason.startswith("minimax_h3_")
         for reason in blocked_reasons
     )
     packet["metadata"] = metadata
@@ -4158,6 +6527,156 @@ class DiffusionGemmaContextHub:
         return (context, _json_dumps(_gemma_context_to_jsonable(context)), _gemma_context_preview(context))
 
 
+class DiffusionGemmaH3ReferenceContext:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "user_prompt": (
+                    "STRING",
+                    {"default": "", "multiline": True, "forceInput": True, "dynamicPrompts": True},
+                ),
+                "reference_manifest": (
+                    "STRING",
+                    {
+                        "default": "<Picture 1>: protagonist identity, face, hair, body, wardrobe, and distinguishing traits; preserve fully.",
+                        "multiline": True,
+                        "dynamicPrompts": True,
+                        "tooltip": "One role per connected Ref2VA asset. Use the exact one-based native tags <Picture N>, <Video N>, and <Audio N>; number each type independently and describe precisely what must or must not transfer. Paired video soundtracks consume Audio labels before standalone audio references.",
+                    },
+                ),
+                "sample_fps": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 24.0, "step": 0.1}),
+                "max_duration_seconds": ("FLOAT", {"default": 60.0, "min": 1.0, "max": 600.0, "step": 1.0}),
+                "max_frames": ("INT", {"default": 60, "min": 1, "max": 512}),
+                "overlong_policy": (["trim", "reject"], {"default": "trim"}),
+            },
+            "optional": {
+                "reference_images": (
+                    "IMAGE",
+                    {
+                        "tooltip": "Optional ordered image batch for DiffusionGemma to inspect. Batch item 1 maps to <Picture 1>, and so on. Connect the same images separately to MiniMaxH3ReferenceToVideo.",
+                    },
+                ),
+                "reference_video": (
+                    "VIDEO",
+                    {
+                        "tooltip": "Optional reference video for DiffusionGemma to inspect as <Video 1>. The native H3 node still needs its 24-fps IMAGE frame batch connected separately.",
+                    },
+                ),
+                "visual_description": ("STRING", {"default": "", "multiline": True, "dynamicPrompts": True}),
+            },
+        }
+
+    RETURN_TYPES = (CONTEXT_TYPE, "STRING", "STRING")
+    RETURN_NAMES = ("gemma_context", "context_json", "context_preview")
+    FUNCTION = "build"
+    CATEGORY = CATEGORY
+
+    def build(
+        self,
+        user_prompt: str,
+        reference_manifest: str,
+        sample_fps: float = 1.0,
+        max_duration_seconds: float = 60.0,
+        max_frames: int = 60,
+        overlong_policy: str = "trim",
+        reference_images=None,
+        reference_video=None,
+        visual_description: str = "",
+    ):
+        normalized_manifest = _normalize_minimax_h3_reference_manifest(reference_manifest)
+        definitions = _minimax_h3_reference_definitions(normalized_manifest)
+        manifest_reasons = _minimax_h3_reference_manifest_validation_reasons(normalized_manifest)
+        reference_image_count = (
+            int(reference_images.shape[0])
+            if getattr(reference_images, "shape", None) is not None and len(reference_images.shape) >= 4
+            else (1 if reference_images is not None else 0)
+        )
+        media_context, _metadata_json = DiffusionGemmaMediaSampler().sample(
+            sample_fps,
+            max_duration_seconds,
+            max_frames,
+            overlong_policy,
+            True,
+            reference_images if reference_video is None else None,
+            reference_video,
+            visual_description,
+            "video_recreation",
+        )
+        if reference_images is not None and reference_video is not None and media_context.images is not None:
+            try:
+                import torch
+
+                prepared_references = []
+                resized = False
+                for index in range(reference_image_count):
+                    prepared, report = _fit_reference_image_to_video_frames(
+                        reference_images[index : index + 1],
+                        media_context.images,
+                    )
+                    prepared_references.append(prepared)
+                    resized = resized or bool(report.get("reference_image_resized_for_video_control"))
+                media_context.images = torch.cat([*prepared_references, media_context.images], dim=0)
+                media_context.source = "image+video"
+                media_context.metadata.update(
+                    {
+                        "source": "image+video",
+                        "reference_image_count": reference_image_count,
+                        "reference_image_backend_attached": True,
+                        "sampled_frame_count": int(media_context.images.shape[0]),
+                        "minimax_h3_reference_images_resized_for_video_analysis": resized,
+                    }
+                )
+            except Exception as exc:
+                media_context.metadata.setdefault("warnings", []).append(
+                    f"Could not attach all H3 reference pictures beside the sampled reference video: {exc}"
+                )
+        else:
+            media_context.metadata["reference_image_count"] = reference_image_count
+        context = _gemma_context_from_media(user_prompt, media_context, visual_description)
+        context.media_metadata.update(
+            {
+                "minimax_h3_mode": "ref2va",
+                "minimax_h3_reference_manifest": normalized_manifest,
+                "minimax_h3_reference_tags": [tag for tag, _description in definitions],
+                "minimax_h3_reference_manifest_reasons": manifest_reasons,
+                "minimax_h3_reference_image_batch_count": reference_image_count,
+                "minimax_h3_reference_video_attached_for_analysis": bool(reference_video is not None),
+            }
+        )
+        if reference_video is not None:
+            context.media_metadata["transformers_video_transport"] = "sampled_frame_images"
+        context.source = "minimax_h3_ref2va"
+        if manifest_reasons:
+            warning = "H3 reference manifest needs correction: " + ", ".join(manifest_reasons)
+            if warning not in context.warnings:
+                context.warnings.append(warning)
+        manifest_picture_count = sum(tag.startswith("<Picture ") for tag, _description in definitions)
+        manifest_video_count = sum(tag.startswith("<Video ") for tag, _description in definitions)
+        manifest_audio_count = sum(tag.startswith("<Audio ") for tag, _description in definitions)
+        if reference_image_count and reference_image_count != manifest_picture_count:
+            context.warnings.append(
+                f"DiffusionGemma received {reference_image_count} reference picture(s) for analysis, while the manifest declares {manifest_picture_count}; the native H3 connections remain authoritative."
+            )
+        if reference_video is not None and manifest_video_count != 1:
+            context.warnings.append(
+                f"The analysis socket inspects one reference video as <Video 1>, while the manifest declares {manifest_video_count}; describe additional videos in the manifest or visual_description and connect them directly to native H3."
+            )
+        elif manifest_video_count > 1:
+            context.warnings.append(
+                "DiffusionGemma can inspect only <Video 1> through this context node; additional declared videos still connect directly to native H3 and rely on their manifest roles."
+            )
+        if manifest_audio_count:
+            context.warnings.append(
+                "Reference audio is not inspected by DiffusionGemma; its declared role is compiled from the manifest, while the audio signal connects directly to native H3."
+            )
+        context.media_metadata["warnings"] = context.warnings
+        preview = _gemma_context_preview(context)
+        if normalized_manifest:
+            preview += "\nH3 reference manifest:\n" + normalized_manifest
+        return (context, _json_dumps(_gemma_context_to_jsonable(context)), preview)
+
+
 class DiffusionGemmaReferencePrep:
     @classmethod
     def INPUT_TYPES(cls):
@@ -4274,16 +6793,24 @@ class DiffusionGemmaTargetProfile:
     def INPUT_TYPES(cls):
         return {
             "required": {
-                "target_profile": (["ltx", "ideogram4"], {"default": "ltx"}),
+                "target_profile": (["ltx", "ideogram4", "minimax_h3"], {"default": "ltx"}),
                 "audio_mode": (
                     ["auto_scene_audio", "explicit_sound_design", "visual_only"],
                     {"default": "auto_scene_audio"},
                 ),
                 "audio_guidance": ("STRING", {"default": "", "multiline": True}),
-                "target_duration_seconds": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 600.0, "step": 0.1}),
+                "target_duration_seconds": (
+                    "FLOAT",
+                    {
+                        "default": 0.0,
+                        "min": 0.0,
+                        "step": 0.1,
+                        "tooltip": "User-selected target duration with no H3-specific upper clamp. At 0, H3 uses media duration, then an explicit duration in the text brief, then a 5-second planning default.",
+                    },
+                ),
                 "ltx_style": ("STRING", {"default": "", "multiline": False}),
                 "ideogram_aspect_ratio": (
-                    ["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
+                    ["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "21:9"],
                     {"default": "1:1"},
                 ),
                 "ideogram_render_style": ("STRING", {"default": "", "multiline": False}),
@@ -4292,6 +6819,15 @@ class DiffusionGemmaTargetProfile:
                 "negative_prompt_mode": (["auto", "empty", "custom"], {"default": "auto"}),
                 "negative_prompt_guidance": ("STRING", {"default": "", "multiline": True}),
             },
+            "optional": {
+                "minimax_h3_mode": (
+                    ["t2va", "ref2va"],
+                    {
+                        "default": "t2va",
+                        "tooltip": "t2va writes the three-field text-to-video prompt. ref2va writes MiniMax's six-section full-reference prompt for MiniMaxH3ReferenceToVideo.",
+                    },
+                ),
+            },
         }
 
     RETURN_TYPES = (TARGET_PROFILE_TYPE, "STRING")
@@ -4312,6 +6848,7 @@ class DiffusionGemmaTargetProfile:
         ideogram_json_output: bool,
         negative_prompt_mode: str,
         negative_prompt_guidance: str,
+        minimax_h3_mode: str = "t2va",
     ):
         config = _make_target_profile_config(
             target_profile,
@@ -4325,6 +6862,7 @@ class DiffusionGemmaTargetProfile:
             ideogram_json_output,
             negative_prompt_mode,
             negative_prompt_guidance,
+            minimax_h3_mode,
         )
         return (config, _json_dumps(_target_profile_config_to_dict(config)))
 
@@ -4356,11 +6894,11 @@ class DiffusionGemmaCoTGenerator:
                 "max_new_tokens": (
                     "INT",
                     {
-                        "default": 640,
+                        "default": 1024,
                         "min": 128,
                         "max": 2048,
                         "step": 64,
-                        "tooltip": "Output budget. Lower values are faster; higher values help Ideogram JSON or dual-output prompts finish cleanly.",
+                        "tooltip": "Output budget. Detailed MiniMax H3 storyboards and Ideogram JSON benefit from 1024 or more tokens.",
                     },
                 ),
             },
@@ -4383,7 +6921,7 @@ class DiffusionGemmaCoTGenerator:
         creativity_mode: str,
         creative_strength: float,
         thinking_mode: str,
-        max_new_tokens: int = 640,
+        max_new_tokens: int = 1024,
         unique_id: str | None = None,
     ):
         runtime_config = _runtime_config_with_temperature(model_config, temperature)
@@ -4403,6 +6941,834 @@ class DiffusionGemmaCoTGenerator:
         return (_json_dumps(packet), reasoning_text, raw_output, _json_dumps(metadata))
 
 
+def _splatstage_blueprint_schema_path() -> Path:
+    return Path(__file__).resolve().parent / "schemas" / "splatstage_show_blueprint.schema.json"
+
+
+def _splatstage_blueprint_errors(value: Any, schema: dict[str, Any]) -> list[str]:
+    try:
+        from jsonschema import Draft202012Validator
+    except Exception as exc:
+        raise RuntimeError("DiffusionGemma SplatStage Planner requires jsonschema.") from exc
+
+    def json_path(error: Any) -> str:
+        parts = [str(part) for part in error.absolute_path]
+        return "$" + "".join(f"[{part}]" if part.isdigit() else f".{part}" for part in parts)
+
+    errors = [
+        f"{json_path(error)}: {error.message}"
+        for error in sorted(
+            Draft202012Validator(schema).iter_errors(value),
+            key=lambda item: tuple(str(part) for part in item.absolute_path),
+        )
+    ]
+    if not isinstance(value, dict):
+        return errors
+    music = value.get("music")
+    if isinstance(music, dict):
+        mode = str(music.get("vocal_mode", ""))
+        lyrics = str(music.get("lyrics", "")).strip()
+        if mode == "instrumental" and lyrics.casefold() not in {"[instrumental]", "instrumental"}:
+            errors.append("$.music.lyrics: instrumental blueprints must use [Instrumental]")
+        if mode == "vocal_hook" and lyrics.casefold() in {"[instrumental]", "instrumental"}:
+            errors.append("$.music.lyrics: vocal_hook requires singable lyrics")
+        if len([line for line in lyrics.splitlines() if line.strip()]) > 48:
+            errors.append("$.music.lyrics: at most 48 non-empty lyric lines are allowed")
+    performer = value.get("performer")
+    if isinstance(performer, dict):
+        segmentation = str(performer.get("segmentation_prompt", "")).strip()
+        categories = [
+            part.strip()
+            for part in re.split(r"[,;]", segmentation)
+            if part.strip()
+        ]
+        if len(categories) > 6:
+            errors.append(
+                "$.performer.segmentation_prompt: use at most 6 comma-separated SAM3 object categories"
+            )
+        if len(categories) == 1 and re.search(
+            r"(?i)\b(?:and|holding|carrying|playing|wearing|wielding)\b",
+            segmentation,
+        ):
+            errors.append(
+                "$.performer.segmentation_prompt: separate the performer, worn silhouette items, "
+                "instruments, and props with commas instead of describing them as one prose subject"
+            )
+    technical = re.compile(
+        r"(?i)(?:\b(?:node(?:\s+id)?|sampler|checkpoint|model[_ -]?path|filename[_ -]?prefix|"
+        r"workflow|output[_ -]?directory|denoise|cuda|comfyui)\b|"
+        r"\b\d+\s+(?:sampling\s+)?steps\b|"
+        r"[A-Za-z]:[\\/]|(?:^|[\s\"'])/(?:home|mnt|tmp|var)/)"
+    )
+    positive_negation = re.compile(
+        r"(?i)\b(?:no|not|without|avoid|exclude|excluding|free\s+of|absent)\b|"
+        r"\b(?:non[- ]?text(?:ual)?|non[- ]?symbolic|text[- ]free|letter[- ]free|"
+        r"word[- ]free|logo[- ]free|watermark[- ]free|people[- ]free|person[- ]free)\b"
+    )
+    for section in ("performer", "background", "fx"):
+        parent = value.get(section)
+        positive = parent.get("positive_prompt") if isinstance(parent, dict) else None
+        if isinstance(positive, str) and positive_negation.search(positive):
+            errors.append(
+                f"$.{section}.positive_prompt: use affirmative visual language only; "
+                f"put exclusions in $.{section}.negative_prompt"
+            )
+    for section, field_name in (
+        ("", "creative_summary"),
+        ("music", "ace_tags"),
+        ("performer", "positive_prompt"),
+        ("performer", "negative_prompt"),
+        ("background", "positive_prompt"),
+        ("background", "negative_prompt"),
+        ("fx", "positive_prompt"),
+        ("fx", "negative_prompt"),
+    ):
+        parent = value if not section else value.get(section)
+        text = parent.get(field_name) if isinstance(parent, dict) else None
+        if isinstance(text, str) and technical.search(text):
+            dotted = f"{section}.{field_name}".strip(".")
+            errors.append(f"$.{dotted}: technical workflow instructions are not allowed")
+    return errors
+
+
+def _normalize_splatstage_lane_cardinality(
+    value: Any,
+    *,
+    selection_index: int,
+) -> tuple[Any, dict[str, Any]]:
+    """Canonicalize an accidental one-or-two-option lane list to one object.
+
+    The node workflow generates two clips from one creative lane contract. Some
+    model responses mirror that physical clip count as arrays despite the JSON
+    schema. Preserve fail-closed validation for malformed data, but safely
+    select the same seed-addressed option across coherent lane alternatives.
+    """
+
+    if not isinstance(value, dict):
+        return value, {}
+    normalized = json.loads(json.dumps(value, ensure_ascii=False))
+    report: dict[str, Any] = {}
+    for lane in ("performer", "background", "fx"):
+        options = normalized.get(lane)
+        if not (
+            isinstance(options, list)
+            and 1 <= len(options) <= 2
+            and all(isinstance(option, dict) for option in options)
+        ):
+            continue
+        selected = int(selection_index) % len(options)
+        normalized[lane] = options[selected]
+        report[lane] = {
+            "input_count": len(options),
+            "selected_index": selected,
+            "policy": "seed_addressed_single_lane_selection",
+        }
+    return normalized, report
+
+
+def _normalize_splatstage_blueprint_identity(
+    value: Any,
+    schema: dict[str, Any],
+) -> tuple[Any, dict[str, Any]]:
+    """Canonicalize only the schema document ID alias to its payload identity.
+
+    The JSON Schema document is identified by ``splatstage.show_blueprint@1``,
+    while blueprint payloads deliberately carry separate ``schema`` and
+    ``version`` fields. DiffusionGemma can reasonably copy the document ID into
+    the payload. Accept that one exact, version-matched alias without relaxing
+    validation for any other schema name or version.
+    """
+
+    if not isinstance(value, dict):
+        return value, {}
+    properties = schema.get("properties")
+    if not isinstance(properties, dict):
+        return value, {}
+    schema_contract = properties.get("schema")
+    version_contract = properties.get("version")
+    if not isinstance(schema_contract, dict) or not isinstance(version_contract, dict):
+        return value, {}
+    expected_schema = schema_contract.get("const")
+    expected_version = version_contract.get("const")
+    document_id = schema.get("$id")
+    if not (
+        isinstance(expected_schema, str)
+        and isinstance(expected_version, int)
+        and isinstance(document_id, str)
+        and document_id == f"{expected_schema}@{expected_version}"
+    ):
+        return value, {}
+    if value.get("schema") != document_id:
+        return value, {}
+    if value.get("version") != expected_version:
+        return value, {}
+
+    normalized = json.loads(json.dumps(value, ensure_ascii=False))
+    normalized["schema"] = expected_schema
+    return normalized, {
+        "input_schema": document_id,
+        "canonical_schema": expected_schema,
+        "version": expected_version,
+        "policy": "exact_version_matched_document_id_alias",
+    }
+
+
+def _splatstage_blueprint_prompt(
+    user_prompt: str,
+    root_seed: int,
+    schema: dict[str, Any],
+    *,
+    duration_seconds: float,
+    aspect_ratio: str,
+) -> str:
+    compact_schema = json.dumps(schema, ensure_ascii=False, separators=(",", ":"))
+    excerpt_label = f"{float(duration_seconds):g}-second"
+    return f"""You are the creative director for a fully automatic SplatStage music-video run.
+Return exactly one JSON object and nothing else. It must validate against the JSON Schema below.
+The two payload identity fields are exact and separate: use "schema":"splatstage.show_blueprint" and
+"version":1. Never put the schema document identifier "splatstage.show_blueprint@1" in the schema field.
+
+Creative brief:
+{user_prompt}
+
+Production facts are fixed and are context only: duration {duration_seconds:.3f} seconds, aspect ratio {aspect_ratio},
+two separately generated performer clips, two people-free background clips, and two people-free FX plates. The JSON still
+contains exactly one performer object, one background object, and one FX object—never arrays or clip-specific lists. The
+workflow derives both clip variants from each single creative lane definition. Do not put technical production
+settings, model names, workflow instructions, file paths, node IDs, samplers, steps, or output prefixes in the JSON.
+The root seed is {int(root_seed)}; use it only as creative entropy and do not include it in the JSON.
+
+Every positive_prompt must use affirmative visual language only. Never put exclusions such as "no words",
+"without people", "text-free", or "non-symbolic" in a positive_prompt. Put every unwanted element as a plain
+comma-separated concept in that lane's negative_prompt instead; for example, positive_prompt "abstract neon
+particles on black" and negative_prompt "text, words, letters, logos".
+
+Interpret a generic pop-music request as a compact English 90-second vocal song with original lyrics. An explicit
+instrumental request overrides that default and must use vocal_mode "instrumental" with lyrics "[Instrumental]".
+An explicit language request overrides English. Make the performer prompt describe one visible primary person with
+useful full-body or medium-full motion. The performer segmentation_prompt is a SAM3 object-category bundle, not a
+sentence: write 1 to 6 short comma-separated noun phrases. Put the primary performer first, followed by every visibly
+present silhouette-extending costume item, worn headwear, handheld instrument, and attached prop described in the
+performer prompt. For example: "cyberpunk mariachi musician, black charro suit, wide-brimmed mariachi hat, glowing
+electric guitar". Omit categories that are not visibly present. The background prompt must contain no person, crowd, performer, dancer,
+singer, face, body, or human subject. The FX positive prompt must affirmatively describe isolated abstract light or
+particle effects on a pure black backing. Put people, faces, logos, typography, letters, words, and symbols in the FX
+negative prompt. The final video uses a {excerpt_label} excerpt, but ACE-Step
+generates a complete 90-second source song first. For vocal music, provide enough lyrics for that full source song,
+normally 16 to 36 non-empty lines organized with section labels such as [Verse], [Chorus], and [Bridge].
+
+JSON Schema:
+{compact_schema}"""
+
+
+class DiffusionGemmaSplatStagePlanner:
+    """Strict creative planner for SplatStage Auto Director."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model_config": (MODEL_TYPE,),
+                "gemma_context": (CONTEXT_TYPE,),
+                "root_seed": (
+                    "INT",
+                    {
+                        "default": 0,
+                        "min": 0,
+                        "max": 9223372036854775807,
+                        "control_after_generate": False,
+                    },
+                ),
+                "duration_seconds": (
+                    "FLOAT",
+                    {"default": 20.0, "min": 10.0, "max": 20.0, "step": 10.0},
+                ),
+                "aspect_ratio": (["16:9"], {"default": "16:9"}),
+                "temperature": (
+                    "FLOAT",
+                    {"default": 0.35, "min": 0.0, "max": 1.0, "step": 0.05},
+                ),
+                "max_new_tokens": (
+                    "INT",
+                    {"default": 2048, "min": 768, "max": 3072, "step": 128},
+                ),
+            },
+            "hidden": {
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("STRING", "BOOLEAN", "STRING", "STRING")
+    RETURN_NAMES = ("blueprint_json", "is_valid", "validation_errors_json", "raw_response")
+    FUNCTION = "plan"
+    OUTPUT_NODE = True
+    CATEGORY = EXPERIMENTAL_MOTION_CATEGORY
+
+    def plan(
+        self,
+        model_config: RuntimeConfig,
+        gemma_context: GemmaContext,
+        root_seed: int,
+        duration_seconds: float,
+        aspect_ratio: str,
+        temperature: float,
+        max_new_tokens: int,
+        unique_id: str | None = None,
+    ):
+        if not isinstance(model_config, RuntimeConfig):
+            raise TypeError("model_config must come from DiffusionGemma Model Loader.")
+        if not isinstance(gemma_context, GemmaContext):
+            raise TypeError("gemma_context must come from DiffusionGemma Context Hub.")
+        supported_durations = {10.0, 20.0}
+        if float(duration_seconds) not in supported_durations or str(aspect_ratio) != "16:9":
+            raise ValueError(
+                "SplatStage planning supports fixed 10- or 20-second 16:9 productions."
+            )
+        user_prompt = " ".join(str(gemma_context.user_prompt or "").split())
+        if not user_prompt:
+            raise ValueError("SplatStage Planner requires a non-empty creative prompt.")
+        if len(user_prompt) > 4000:
+            raise ValueError("SplatStage Planner prompt must be at most 4000 characters.")
+
+        schema_path = _splatstage_blueprint_schema_path()
+        if not schema_path.is_file():
+            raise RuntimeError(f"SplatStage blueprint schema is missing: {schema_path}")
+        schema = json.loads(schema_path.read_text(encoding="utf-8"))
+        runtime_config = _runtime_config_with_temperature(model_config, float(temperature))
+        node_id = str(unique_id) if unique_id is not None else None
+
+        try:
+            try:
+                import torch
+
+                torch.manual_seed(int(root_seed) & 0x7FFF_FFFF_FFFF_FFFF)
+                if torch.cuda.is_available():
+                    torch.cuda.manual_seed_all(int(root_seed) & 0x7FFF_FFFF_FFFF_FFFF)
+            except Exception:
+                pass
+            initial_raw = _run_backend(
+                runtime_config,
+                _splatstage_blueprint_prompt(
+                    user_prompt,
+                    int(root_seed),
+                    schema,
+                    duration_seconds=float(duration_seconds),
+                    aspect_ratio=str(aspect_ratio),
+                ),
+                None,
+                int(max_new_tokens),
+                node_id=node_id,
+            )
+            parsed = _extract_json_object(initial_raw)
+            parsed, initial_identity_normalization = (
+                _normalize_splatstage_blueprint_identity(parsed, schema)
+            )
+            parsed, initial_normalization = _normalize_splatstage_lane_cardinality(
+                parsed,
+                selection_index=int(root_seed),
+            )
+            errors = _splatstage_blueprint_errors(parsed, schema)
+            raw_record: dict[str, Any] = {"initial": initial_raw}
+            if initial_identity_normalization:
+                raw_record["initial_identity_normalization"] = (
+                    initial_identity_normalization
+                )
+            if initial_normalization:
+                raw_record["initial_lane_normalization"] = initial_normalization
+            if errors:
+                repair_prompt = (
+                    "Repair the following candidate so it validates. Return only the corrected JSON object. "
+                    "Do not add fields and do not explain. performer, background, and fx must each be "
+                    "one JSON object, never arrays; choose one coherent option if the candidate contains alternatives.\n\n"
+                    "Set the identity fields exactly to schema=\"splatstage.show_blueprint\" and version=1; "
+                    "do not use the schema document ID splatstage.show_blueprint@1 as the schema value.\n\n"
+                    "Validation errors:\n- "
+                    + "\n- ".join(errors)
+                    + "\n\nCandidate:\n"
+                    + (json.dumps(parsed, ensure_ascii=False) if parsed is not None else initial_raw)
+                )
+                repair_seed = (int(root_seed) + 1) & 0x7FFF_FFFF_FFFF_FFFF
+                try:
+                    import torch
+
+                    torch.manual_seed(repair_seed)
+                    if torch.cuda.is_available():
+                        torch.cuda.manual_seed_all(repair_seed)
+                except Exception:
+                    pass
+                repaired_raw = _run_backend(
+                    runtime_config,
+                    repair_prompt,
+                    None,
+                    int(max_new_tokens),
+                    node_id=node_id,
+                )
+                raw_record["repair"] = repaired_raw
+                parsed = _extract_json_object(repaired_raw)
+                parsed, repair_identity_normalization = (
+                    _normalize_splatstage_blueprint_identity(parsed, schema)
+                )
+                if repair_identity_normalization:
+                    raw_record["repair_identity_normalization"] = (
+                        repair_identity_normalization
+                    )
+                parsed, repair_normalization = _normalize_splatstage_lane_cardinality(
+                    parsed,
+                    selection_index=int(root_seed),
+                )
+                if repair_normalization:
+                    raw_record["repair_lane_normalization"] = repair_normalization
+                errors = _splatstage_blueprint_errors(parsed, schema)
+            if errors or parsed is None:
+                raise RuntimeError(
+                    "DiffusionGemma could not produce a valid SplatStage blueprint after one repair: "
+                    + "; ".join(errors or ["response did not contain a JSON object"])
+                )
+            blueprint_json = json.dumps(parsed, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
+            raw_response = json.dumps(raw_record, ensure_ascii=False)
+            result = (blueprint_json, True, "[]", raw_response)
+            return {
+                "ui": {
+                    "splatstage_show_blueprint": [blueprint_json],
+                    "splatstage_raw_response": [raw_response],
+                },
+                "result": result,
+            }
+        finally:
+            _maybe_unload(runtime_config)
+
+
+def _splatstage_efficient_blueprint_schema_path() -> Path:
+    return Path(__file__).resolve().parent / "schemas" / "splatstage_efficient_show_blueprint.schema.json"
+
+
+def _splatstage_efficient_blueprint_errors(
+    value: Any,
+    schema: dict[str, Any],
+) -> list[str]:
+    try:
+        from jsonschema import Draft202012Validator
+    except Exception as exc:
+        raise RuntimeError(
+            "DiffusionGemma Efficient SplatStage Planner requires jsonschema."
+        ) from exc
+
+    def json_path(error: Any) -> str:
+        parts = [str(part) for part in error.absolute_path]
+        return "$" + "".join(
+            f"[{part}]" if part.isdigit() else f".{part}" for part in parts
+        )
+
+    errors = [
+        f"{json_path(error)}: {error.message}"
+        for error in sorted(
+            Draft202012Validator(schema).iter_errors(value),
+            key=lambda item: tuple(str(part) for part in item.absolute_path),
+        )
+    ]
+    if not isinstance(value, dict):
+        return errors
+
+    positive_negation = re.compile(
+        r"(?i)\b(?:no|not|without|avoid|exclude|excluding|free\s+of|absent)\b|"
+        r"\b(?:non[- ]?text(?:ual)?|non[- ]?symbolic|text[- ]free|letter[- ]free|"
+        r"word[- ]free|logo[- ]free|watermark[- ]free|people[- ]free|person[- ]free)\b"
+    )
+    technical = re.compile(
+        r"(?i)(?:\b(?:node(?:\s+id)?|sampler|checkpoint|model[_ -]?path|filename[_ -]?prefix|"
+        r"workflow|output[_ -]?directory|denoise|cuda|comfyui)\b|"
+        r"\b\d+\s+(?:sampling\s+)?steps\b|[A-Za-z]:[\\/])"
+    )
+    positive_fields: list[tuple[str, Any]] = []
+    for lane in ("performer", "background", "fx"):
+        section = value.get(lane)
+        if not isinstance(section, dict):
+            continue
+        positive_fields.append((f"$.{lane}.positive_prompt", section.get("positive_prompt")))
+        if lane in {"performer", "background"}:
+            variations = section.get("variations")
+            if isinstance(variations, list):
+                positive_fields.extend(
+                    (f"$.{lane}.variations[{index}]", variation)
+                    for index, variation in enumerate(variations)
+                )
+    for path, text in positive_fields:
+        if isinstance(text, str) and positive_negation.search(text):
+            errors.append(
+                f"{path}: use affirmative visual language only; put exclusions in the lane negative_prompt"
+            )
+        if isinstance(text, str) and technical.search(text):
+            errors.append(f"{path}: technical workflow instructions are not allowed")
+
+    fx = value.get("fx")
+    if isinstance(fx, dict):
+        fx_positive = str(fx.get("positive_prompt", ""))
+        if re.search(
+            r"(?i)\b(?:touchdesigner|software|application|app|editor|viewport|"
+            r"interface|node\s+graph|control\s+panel|menu|screenshot)\b",
+            fx_positive,
+        ):
+            errors.append(
+                "$.fx.positive_prompt: describe only the show-specific artwork; "
+                "SplatStage supplies the locked TouchDesigner-inspired finished-artwork style"
+            )
+
+    performer = value.get("performer")
+    if isinstance(performer, dict):
+        segmentation = str(performer.get("segmentation_prompt", "")).strip()
+        categories = [
+            part.strip() for part in re.split(r"[,;]", segmentation) if part.strip()
+        ]
+        if not 1 <= len(categories) <= 6:
+            errors.append(
+                "$.performer.segmentation_prompt: use 1 to 6 comma-separated object categories"
+            )
+        if len(categories) == 1 and re.search(
+            r"(?i)\b(?:and|holding|carrying|playing|wearing|wielding)\b",
+            segmentation,
+        ):
+            errors.append(
+                "$.performer.segmentation_prompt: separate the performer, headwear, costume, instrument, and props with commas"
+            )
+        generic_component = re.compile(
+            r"(?i)^(?:face and head|head and face|hair and worn headwear|"
+            r"costume and worn accessories|held instrument and attached props|"
+            r"full body person|torso|arms|hands|legs)$"
+        )
+        if any(generic_component.fullmatch(category) for category in categories[1:]):
+            errors.append(
+                "$.performer.segmentation_prompt: list only identity-specific items actually present; "
+                "generic anatomy and placeholder accessory categories are added deterministically"
+            )
+
+    music = value.get("music")
+    if isinstance(music, dict):
+        mode = str(music.get("vocal_mode", ""))
+        lyrics = str(music.get("lyrics", "")).strip()
+        if mode == "instrumental" and lyrics.casefold() not in {
+            "[instrumental]",
+            "instrumental",
+        }:
+            errors.append(
+                "$.music.lyrics: instrumental blueprints must use [Instrumental]"
+            )
+        if mode == "vocal_hook" and lyrics.casefold() in {
+            "[instrumental]",
+            "instrumental",
+        }:
+            errors.append("$.music.lyrics: vocal_hook requires singable lyrics")
+        lyric_lines = [line for line in lyrics.splitlines() if line.strip()]
+        if len(lyric_lines) > 48:
+            errors.append("$.music.lyrics: at most 48 non-empty lyric lines are allowed")
+    return errors
+
+
+def _normalize_splatstage_efficient_segmentation(value: Any) -> Any:
+    """Make the canonical identity authoritative for the SAM3 prompt bundle.
+
+    Generic anatomy queries caused SAM3 to union unrelated people from crowded
+    source plates.  The efficient workflow now asks for one connected character
+    identity and adds completeness language downstream.  This normalizer keeps
+    only concrete, actually authored costume/headwear/instrument/prop hints.
+    """
+
+    if not isinstance(value, dict):
+        return value
+    normalized = copy.deepcopy(value)
+    performer = normalized.get("performer")
+    if not isinstance(performer, dict):
+        return normalized
+    identity = " ".join(str(performer.get("identity_description", "")).strip().split())
+    identity = re.sub(r"(?i)^(?:a|an|the)\s+", "", identity).strip(" ,.;")
+    authored = [
+        " ".join(part.strip().split()).strip(" ,.;")
+        for part in re.split(r"[,;]", str(performer.get("segmentation_prompt", "")))
+        if part.strip()
+    ]
+    if not identity and authored:
+        identity = authored[0]
+    if not identity:
+        return normalized
+
+    generic = re.compile(
+        r"(?i)^(?:face and head|head and face|hair and worn headwear|"
+        r"costume and worn accessories|held instrument and attached props|"
+        r"full body person|torso|arms|hands|legs)$"
+    )
+    specific = re.compile(
+        r"(?i)\b(?:hat|helmet|headdress|crown|cap|hood|turban|goggles|glasses|"
+        r"jacket|coat|robe|dress|suit|armor|charro|backpack|cape|wings?|"
+        r"guitar|bass|violin|cello|trumpet|trombone|saxophone|drum|keyboard|"
+        r"piano|microphone|sword|staff|cane|shield|instrument|prop)\w*\b"
+    )
+    candidates = authored[1:] if authored else []
+    candidates.extend(
+        " ".join(part.strip().split()).strip(" ,.;")
+        for part in re.split(r"[,;]", str(performer.get("positive_prompt", "")))
+        if part.strip()
+    )
+    selected = [identity]
+    for category in candidates:
+        if generic.fullmatch(category) or not specific.search(category):
+            continue
+        if category.casefold() in {item.casefold() for item in selected}:
+            continue
+        selected.append(category)
+        if len(selected) >= 6:
+            break
+    performer["segmentation_prompt"] = ", ".join(selected)
+    return normalized
+
+
+def _splatstage_efficient_blueprint_prompt(
+    user_prompt: str,
+    root_seed: int,
+    schema: dict[str, Any],
+) -> str:
+    compact_schema = json.dumps(schema, ensure_ascii=False, separators=(",", ":"))
+    return f"""You are the creative director for an efficient, fully automatic SplatStage music video.
+Return exactly one JSON object and nothing else. It must validate against the JSON Schema below.
+Use the exact payload identity \"schema\":\"splatstage.efficient_show_blueprint\" and \"version\":1.
+
+Creative brief:
+{user_prompt}
+
+The production creates one complete 90-second song, three separate five-second performer videos,
+two separate five-second people-free background videos, and one still-image abstract FX source.
+Describe one canonical performer identity, then provide exactly three affirmative shot/action/camera
+variations that preserve that identity. Describe one background art direction, then provide exactly
+two affirmative environment/camera variations. The segmentation_prompt is a comma-separated SAM3
+identity hint bundle with at most six entries. Restate the canonical performer identity first, then list
+only concrete silhouette-extending items actually present, such as a specific hat, costume, held
+instrument, or attached prop. Do not invent generic placeholders for absent items. Example:
+\"cyberpunk mariachi musician, wide-brimmed mariachi hat, black charro suit, glowing electric guitar\".
+
+Every positive_prompt and every variation must use affirmative visual language only. Never write
+phrases such as \"no words\", \"without people\", \"text-free\", or \"non-symbolic\" in positive
+fields. Put exclusions only in negative_prompt. The background positive prompt must describe scenery
+without introducing a person, crowd, performer, dancer, singer, face, body, or human silhouette. The
+FX positive_prompt is the show-specific creative slot for one richly art-directed luminous motif.
+Coordinate it with the same theme, music, performer, and background, and explicitly specify: the
+abstract motif or shape language; a limited color palette; material and texture; rhythmic energy or
+particle behavior; and a clear spatial composition with controlled negative space. Use affirmative
+visual language and prefer organic light phenomena or sculptural forms over recognizable icons.
+Do not name TouchDesigner or any software, application, interface, editor, node graph, control panel,
+viewport, or screenshot in the positive prompt. SplatStage adds a locked TouchDesigner-inspired
+finished-artwork style spine after planning. Put people, faces, typography, letters, words, logos,
+watermarks, symbols, UI, HUD, labels, and signage in the FX negative prompt.
+
+Interpret a generic pop request as an English vocal song with an original hook and 16 to 36 non-empty
+lyric lines organized with labels such as [Verse], [Chorus], and [Bridge]. Explicit instrumental or
+language requests override that default. Do not include model names, node IDs, dimensions, frame
+counts, samplers, steps, paths, output prefixes, or other technical workflow instructions. The root
+seed {int(root_seed)} is creative entropy only and must not appear in the JSON.
+
+JSON Schema:
+{compact_schema}"""
+
+
+def _splatstage_efficient_cache_path(user_prompt: str, root_seed: int) -> Path:
+    import hashlib
+    import folder_paths
+
+    identity = hashlib.sha256(
+        json.dumps(
+            {
+                "schema": "splatstage.efficient_planner_cache_input",
+                "version": 1,
+                "user_prompt": " ".join(str(user_prompt).split()),
+                "root_seed": int(root_seed),
+                "planner_prompt_version": 2,
+            },
+            sort_keys=True,
+            separators=(",", ":"),
+        ).encode("utf-8")
+    ).hexdigest()
+    root = (
+        Path(folder_paths.get_output_directory()).resolve()
+        / "4d_splat"
+        / "splatstage_efficient_node_workflow"
+        / "planner"
+        / identity
+    )
+    return root / "blueprint.json"
+
+
+class DiffusionGemmaSplatStageEfficientPlanner:
+    """One-call creative planner for the efficient full-song node workflow."""
+
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "model_config": (MODEL_TYPE,),
+                "gemma_context": (CONTEXT_TYPE,),
+                "root_seed": (
+                    "INT",
+                    {
+                        "default": 26073001,
+                        "min": 0,
+                        "max": 9223372036854775807,
+                        "control_after_generate": False,
+                    },
+                ),
+                "temperature": (
+                    "FLOAT",
+                    {"default": 0.35, "min": 0.0, "max": 1.0, "step": 0.05},
+                ),
+                "max_new_tokens": (
+                    "INT",
+                    {"default": 3072, "min": 1024, "max": 4096, "step": 128},
+                ),
+                "reuse_cached_blueprint": (
+                    "BOOLEAN",
+                    {"default": True},
+                ),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("STRING", "BOOLEAN", "STRING", "STRING")
+    RETURN_NAMES = (
+        "blueprint_json",
+        "is_valid",
+        "validation_errors_json",
+        "raw_response",
+    )
+    FUNCTION = "plan"
+    OUTPUT_NODE = True
+    CATEGORY = EXPERIMENTAL_MOTION_CATEGORY
+
+    def plan(
+        self,
+        model_config: RuntimeConfig,
+        gemma_context: GemmaContext,
+        root_seed: int,
+        temperature: float,
+        max_new_tokens: int,
+        reuse_cached_blueprint: bool,
+        unique_id: str | None = None,
+    ):
+        if not isinstance(model_config, RuntimeConfig):
+            raise TypeError("model_config must come from DiffusionGemma Model Loader.")
+        if not isinstance(gemma_context, GemmaContext):
+            raise TypeError("gemma_context must come from DiffusionGemma Context Hub.")
+        user_prompt = " ".join(str(gemma_context.user_prompt or "").split())
+        if not user_prompt:
+            raise ValueError("Efficient SplatStage Planner requires a non-empty prompt.")
+        if len(user_prompt) > 4000:
+            raise ValueError("Efficient SplatStage Planner prompt must be at most 4000 characters.")
+
+        schema_path = _splatstage_efficient_blueprint_schema_path()
+        if not schema_path.is_file():
+            raise RuntimeError(f"Efficient SplatStage blueprint schema is missing: {schema_path}")
+        schema = json.loads(schema_path.read_text(encoding="utf-8"))
+        cache_path = _splatstage_efficient_cache_path(user_prompt, int(root_seed))
+        if bool(reuse_cached_blueprint) and cache_path.is_file():
+            try:
+                cached = _normalize_splatstage_efficient_segmentation(
+                    json.loads(cache_path.read_text(encoding="utf-8"))
+                )
+                errors = _splatstage_efficient_blueprint_errors(cached, schema)
+                if not errors:
+                    atomic_cache = cache_path.with_name(
+                        f".{cache_path.name}.{os.getpid()}.partial"
+                    )
+                    atomic_cache.write_text(
+                        json.dumps(cached, ensure_ascii=False, sort_keys=True, indent=2),
+                        encoding="utf-8",
+                    )
+                    os.replace(atomic_cache, cache_path)
+                    blueprint_json = json.dumps(
+                        cached, ensure_ascii=False, sort_keys=True, separators=(",", ":")
+                    )
+                    raw_response = json.dumps(
+                        {"cache": "reused", "path": str(cache_path)}, ensure_ascii=False
+                    )
+                    return {
+                        "ui": {
+                            "splatstage_efficient_show_blueprint": [blueprint_json],
+                            "splatstage_raw_response": [raw_response],
+                        },
+                        "result": (blueprint_json, True, "[]", raw_response),
+                    }
+            except (OSError, json.JSONDecodeError):
+                pass
+
+        runtime_config = _runtime_config_with_temperature(model_config, float(temperature))
+        node_id = str(unique_id) if unique_id is not None else None
+        try:
+            initial_raw = _run_backend(
+                runtime_config,
+                _splatstage_efficient_blueprint_prompt(
+                    user_prompt, int(root_seed), schema
+                ),
+                None,
+                int(max_new_tokens),
+                node_id=node_id,
+            )
+            parsed = _normalize_splatstage_efficient_segmentation(
+                _extract_json_object(initial_raw)
+            )
+            errors = _splatstage_efficient_blueprint_errors(parsed, schema)
+            raw_record: dict[str, Any] = {"initial": initial_raw}
+            if errors:
+                repair_prompt = (
+                    "Repair this candidate to satisfy the validation errors. Return only the corrected JSON object, "
+                    "with no explanation and no extra fields. Keep every positive prompt and variation affirmative; "
+                    "move exclusions into negative_prompt.\n\nValidation errors:\n- "
+                    + "\n- ".join(errors)
+                    + "\n\nCandidate:\n"
+                    + (
+                        json.dumps(parsed, ensure_ascii=False)
+                        if parsed is not None
+                        else initial_raw
+                    )
+                )
+                repaired_raw = _run_backend(
+                    runtime_config,
+                    repair_prompt,
+                    None,
+                    int(max_new_tokens),
+                    node_id=node_id,
+                )
+                raw_record["repair"] = repaired_raw
+                parsed = _normalize_splatstage_efficient_segmentation(
+                    _extract_json_object(repaired_raw)
+                )
+                errors = _splatstage_efficient_blueprint_errors(parsed, schema)
+            if errors or parsed is None:
+                raise RuntimeError(
+                    "DiffusionGemma could not produce a valid efficient SplatStage blueprint after one repair: "
+                    + "; ".join(errors or ["response did not contain a JSON object"])
+                )
+
+            cache_path.parent.mkdir(parents=True, exist_ok=True)
+            partial = cache_path.with_name(f".{cache_path.name}.{os.getpid()}.partial")
+            partial.write_text(
+                json.dumps(parsed, ensure_ascii=False, sort_keys=True, indent=2),
+                encoding="utf-8",
+            )
+            os.replace(partial, cache_path)
+            diagnostic = cache_path.parent / "raw_response.json"
+            diagnostic.write_text(
+                json.dumps(raw_record, ensure_ascii=False, indent=2), encoding="utf-8"
+            )
+            blueprint_json = json.dumps(
+                parsed, ensure_ascii=False, sort_keys=True, separators=(",", ":")
+            )
+            raw_response = json.dumps(raw_record, ensure_ascii=False)
+            return {
+                "ui": {
+                    "splatstage_efficient_show_blueprint": [blueprint_json],
+                    "splatstage_raw_response": [raw_response],
+                },
+                "result": (blueprint_json, True, "[]", raw_response),
+            }
+        finally:
+            _maybe_unload(runtime_config)
+
+
 class DiffusionGemmaJSONSplitter:
     @classmethod
     def INPUT_TYPES(cls):
@@ -4451,6 +7817,7 @@ class DiffusionGemmaJSONSplitter:
         "INT",
         "INT",
         "BOOLEAN",
+        "STRING",
     )
     RETURN_NAMES = (
         "ltx_prompt",
@@ -4466,6 +7833,7 @@ class DiffusionGemmaJSONSplitter:
         "resolution_width",
         "resolution_height",
         "ready_for_generation",
+        "minimax_h3_prompt",
     )
     FUNCTION = "split"
     CATEGORY = CATEGORY
@@ -4499,7 +7867,7 @@ class DiffusionGemmaJSONSplitter:
             packet_metadata_for_split["json_parse_valid"] = bool(strict_packet is not None)
             if warning:
                 packet_metadata_for_split["salvage_warning"] = warning
-        ltx_prompt, ideogram_prompt, negative_prompt, scene_segments_json, local_prompts, segment_lengths, aspect_ratio, packet_metadata = _packet_to_prompt_outputs(
+        ltx_prompt, ideogram_prompt, negative_prompt, scene_segments_json, local_prompts, segment_lengths, aspect_ratio, packet_metadata, minimax_h3_prompt = _packet_to_prompt_outputs(
             packet,
             context,
             target,
@@ -4521,6 +7889,7 @@ class DiffusionGemmaJSONSplitter:
             scene_segments_json = "[]"
             local_prompts = ""
             segment_lengths = ""
+            minimax_h3_prompt = ""
         metadata.update(
             {
                 "splitter_json_valid": bool(is_valid and ready_for_generation),
@@ -4554,9 +7923,42 @@ class DiffusionGemmaJSONSplitter:
             resolution_width,
             resolution_height,
             ready_for_generation,
+            minimax_h3_prompt,
         )
 
 
+class DiffusionGemmaGenerationGate:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "prompt": ("STRING", {"default": "", "multiline": True, "forceInput": True}),
+                "ready_for_generation": ("BOOLEAN", {"default": False, "forceInput": True}),
+            },
+            "optional": {
+                "metadata_json": ("STRING", {"default": "", "multiline": True, "forceInput": True}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING",)
+    RETURN_NAMES = ("prompt",)
+    FUNCTION = "gate"
+    CATEGORY = CATEGORY
+
+    def gate(self, prompt: str, ready_for_generation: bool, metadata_json: str = ""):
+        if not bool(ready_for_generation):
+            metadata = _extract_json_object(metadata_json) or {}
+            blocked_reasons = metadata.get("blocked_reasons") if isinstance(metadata.get("blocked_reasons"), list) else []
+            details = ", ".join(str(reason) for reason in blocked_reasons if str(reason).strip())
+            message = "DiffusionGemma blocked generation because the prompt did not pass validation."
+            if details:
+                message += f" Fix these issues and queue again: {details}."
+            raise ValueError(message)
+        if not str(prompt or "").strip():
+            raise ValueError("DiffusionGemma blocked generation because the validated prompt is empty.")
+        return (str(prompt),)
+
+
 class DiffusionGemmaPromptBuilder:
     @classmethod
     def INPUT_TYPES(cls):
@@ -4747,6 +8149,7 @@ class DiffusionGemmaPromptBuilder:
             segment_lengths,
             _aspect_ratio,
             _packet_metadata,
+            _minimax_h3_prompt,
         ) = _packet_to_prompt_outputs(
             packet,
             context,
@@ -5347,15 +8750,23 @@ class TextOutput:
 NODE_CLASS_MAPPINGS = {
     "DiffusionGemmaModelLoader": DiffusionGemmaModelLoader,
     "DiffusionGemmaContextHub": DiffusionGemmaContextHub,
+    "DiffusionGemmaH3ReferenceContext": DiffusionGemmaH3ReferenceContext,
     "DiffusionGemmaTargetProfile": DiffusionGemmaTargetProfile,
     "DiffusionGemmaCoTGenerator": DiffusionGemmaCoTGenerator,
+    "DiffusionGemmaSplatStagePlanner": DiffusionGemmaSplatStagePlanner,
+    "DiffusionGemmaSplatStageEfficientPlanner": DiffusionGemmaSplatStageEfficientPlanner,
     "DiffusionGemmaJSONSplitter": DiffusionGemmaJSONSplitter,
+    "DiffusionGemmaGenerationGate": DiffusionGemmaGenerationGate,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "DiffusionGemmaModelLoader": "DiffusionGemma Model Loader (Advanced)",
     "DiffusionGemmaContextHub": "DiffusionGemma Context Hub",
+    "DiffusionGemmaH3ReferenceContext": "DiffusionGemma H3 Reference Context",
     "DiffusionGemmaTargetProfile": "DiffusionGemma Target Profile",
     "DiffusionGemmaCoTGenerator": "DiffusionGemma CoT Generator",
+    "DiffusionGemmaSplatStagePlanner": "DiffusionGemma SplatStage Planner (Experimental)",
+    "DiffusionGemmaSplatStageEfficientPlanner": "DiffusionGemma SplatStage Efficient Planner (Experimental)",
     "DiffusionGemmaJSONSplitter": "DiffusionGemma JSON Splitter",
+    "DiffusionGemmaGenerationGate": "DiffusionGemma Generation Gate",
 }
diff --git a/pyproject.toml b/pyproject.toml
index 14bcddf..b93ec67 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,7 @@
 [project]
 name = "diffusiongemma-prompt-builder"
-version = "0.1.0"
-description = "Five-node DiffusionGemma Director Assistant bridge for ComfyUI LTX workflows."
+version = "0.2.0"
+description = "Stable DiffusionGemma Director prompt authoring for LTX, Ideogram4, and MiniMax H3 in ComfyUI, plus experimental SplatStage planners."
 license = { file = "LICENSE" }
 requires-python = ">=3.11"
 dependencies = [
@@ -25,4 +25,4 @@ Repository = "https://github.com/exportAnything/ComfyUI-DiffusionGemmaPromptBuil
 [tool.comfy]
 PublisherId = "exportanything"
 DisplayName = "DiffusionGemma Prompt Builder"
-requires-comfyui = ">=1.0.0"
+requires-comfyui = ">=0.30.1"
diff --git a/schemas/prompt_packet.schema.json b/schemas/prompt_packet.schema.json
index b43d2aa..58cfc2d 100644
--- a/schemas/prompt_packet.schema.json
+++ b/schemas/prompt_packet.schema.json
@@ -13,6 +13,7 @@
         {"type": "object"}
       ]
     },
+    "minimax_h3_prompt": {"type": "string"},
     "negative_prompt": {"type": "string"},
     "scene_segments": {
       "type": "array",
diff --git a/schemas/splatstage_efficient_show_blueprint.schema.json b/schemas/splatstage_efficient_show_blueprint.schema.json
new file mode 100644
index 0000000..09a5659
--- /dev/null
+++ b/schemas/splatstage_efficient_show_blueprint.schema.json
@@ -0,0 +1,109 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "splatstage.efficient_show_blueprint@1",
+  "title": "SplatStage Efficient Show Blueprint",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema",
+    "version",
+    "project_title",
+    "creative_summary",
+    "music",
+    "performer",
+    "background",
+    "fx",
+    "editing",
+    "finish"
+  ],
+  "properties": {
+    "schema": {"const": "splatstage.efficient_show_blueprint"},
+    "version": {"const": 1},
+    "project_title": {"type": "string", "minLength": 1, "maxLength": 120},
+    "creative_summary": {"type": "string", "minLength": 1, "maxLength": 800},
+    "music": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["ace_tags", "vocal_mode", "lyrics", "bpm", "key", "time_signature", "language"],
+      "properties": {
+        "ace_tags": {"type": "string", "minLength": 1, "maxLength": 800},
+        "vocal_mode": {"enum": ["vocal_hook", "instrumental"]},
+        "lyrics": {"type": "string", "minLength": 1, "maxLength": 5000},
+        "bpm": {"type": "integer", "minimum": 60, "maximum": 220},
+        "key": {
+          "enum": [
+            "C major", "C minor", "C# major", "C# minor", "D major", "D minor",
+            "D# major", "D# minor", "E major", "E minor", "F major", "F minor",
+            "F# major", "F# minor", "G major", "G minor", "G# major", "G# minor",
+            "A major", "A minor", "A# major", "A# minor", "B major", "B minor"
+          ]
+        },
+        "time_signature": {"enum": ["2", "3", "4", "6"]},
+        "language": {
+          "enum": ["en", "es", "fr", "de", "it", "pt", "ja", "ko", "zh", "hi", "ar"]
+        }
+      }
+    },
+    "performer": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["identity_description", "positive_prompt", "negative_prompt", "segmentation_prompt", "variations"],
+      "properties": {
+        "identity_description": {"type": "string", "minLength": 1, "maxLength": 800},
+        "positive_prompt": {"type": "string", "minLength": 1, "maxLength": 2200},
+        "negative_prompt": {"type": "string", "minLength": 1, "maxLength": 1600},
+        "segmentation_prompt": {"type": "string", "minLength": 1, "maxLength": 500},
+        "variations": {
+          "type": "array",
+          "minItems": 3,
+          "maxItems": 3,
+          "items": {"type": "string", "minLength": 1, "maxLength": 700}
+        }
+      }
+    },
+    "background": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["positive_prompt", "negative_prompt", "variations"],
+      "properties": {
+        "positive_prompt": {"type": "string", "minLength": 1, "maxLength": 2200},
+        "negative_prompt": {"type": "string", "minLength": 1, "maxLength": 1600},
+        "variations": {
+          "type": "array",
+          "minItems": 2,
+          "maxItems": 2,
+          "items": {"type": "string", "minLength": 1, "maxLength": 700}
+        }
+      }
+    },
+    "fx": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["positive_prompt", "negative_prompt", "blend_mode", "key_mode", "opacity"],
+      "properties": {
+        "positive_prompt": {"type": "string", "minLength": 1, "maxLength": 1800},
+        "negative_prompt": {"type": "string", "minLength": 1, "maxLength": 1600},
+        "blend_mode": {"enum": ["screen", "add", "lighten"]},
+        "key_mode": {"enum": ["luminance", "black"]},
+        "opacity": {"type": "number", "minimum": 0.2, "maximum": 0.65}
+      }
+    },
+    "editing": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["density", "music_reactivity"],
+      "properties": {
+        "density": {"enum": ["sparse", "musical", "energetic"]},
+        "music_reactivity": {"enum": ["low", "medium", "high"]}
+      }
+    },
+    "finish": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": ["style_preset"],
+      "properties": {
+        "style_preset": {"enum": ["broadcast_clean", "cable_access", "late_night_vhs", "damaged_tape"]}
+      }
+    }
+  }
+}
diff --git a/schemas/splatstage_show_blueprint.schema.json b/schemas/splatstage_show_blueprint.schema.json
new file mode 100644
index 0000000..f106b96
--- /dev/null
+++ b/schemas/splatstage_show_blueprint.schema.json
@@ -0,0 +1,311 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "splatstage.show_blueprint@1",
+  "title": "SplatStage Show Blueprint",
+  "type": "object",
+  "additionalProperties": false,
+  "required": [
+    "schema",
+    "version",
+    "project_title",
+    "creative_summary",
+    "music",
+    "performer",
+    "background",
+    "fx",
+    "editing",
+    "finish"
+  ],
+  "properties": {
+    "schema": {
+      "const": "splatstage.show_blueprint"
+    },
+    "version": {
+      "const": 1
+    },
+    "project_title": {
+      "type": "string",
+      "minLength": 1,
+      "maxLength": 120
+    },
+    "creative_summary": {
+      "type": "string",
+      "minLength": 1,
+      "maxLength": 1000
+    },
+    "music": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "ace_tags",
+        "vocal_mode",
+        "lyrics",
+        "bpm",
+        "key",
+        "time_signature",
+        "language"
+      ],
+      "properties": {
+        "ace_tags": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 1200
+        },
+        "vocal_mode": {
+          "enum": [
+            "vocal_hook",
+            "instrumental"
+          ]
+        },
+        "lyrics": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 2000
+        },
+        "bpm": {
+          "type": "integer",
+          "minimum": 10,
+          "maximum": 300
+        },
+        "key": {
+          "enum": [
+            "C major",
+            "C# major",
+            "Db major",
+            "D major",
+            "D# major",
+            "Eb major",
+            "E major",
+            "F major",
+            "F# major",
+            "Gb major",
+            "G major",
+            "G# major",
+            "Ab major",
+            "A major",
+            "A# major",
+            "Bb major",
+            "B major",
+            "C minor",
+            "C# minor",
+            "Db minor",
+            "D minor",
+            "D# minor",
+            "Eb minor",
+            "E minor",
+            "F minor",
+            "F# minor",
+            "Gb minor",
+            "G minor",
+            "G# minor",
+            "Ab minor",
+            "A minor",
+            "A# minor",
+            "Bb minor",
+            "B minor"
+          ]
+        },
+        "time_signature": {
+          "enum": [
+            "2",
+            "3",
+            "4",
+            "6"
+          ]
+        },
+        "language": {
+          "enum": [
+            "ar",
+            "az",
+            "bg",
+            "bn",
+            "ca",
+            "cs",
+            "da",
+            "de",
+            "el",
+            "en",
+            "es",
+            "fa",
+            "fi",
+            "fr",
+            "he",
+            "hi",
+            "hr",
+            "ht",
+            "hu",
+            "id",
+            "is",
+            "it",
+            "ja",
+            "ko",
+            "la",
+            "lt",
+            "ms",
+            "ne",
+            "nl",
+            "no",
+            "pa",
+            "pl",
+            "pt",
+            "ro",
+            "ru",
+            "sa",
+            "sk",
+            "sr",
+            "sv",
+            "sw",
+            "ta",
+            "te",
+            "th",
+            "tl",
+            "tr",
+            "uk",
+            "ur",
+            "vi",
+            "yue",
+            "zh",
+            "unknown"
+          ]
+        }
+      }
+    },
+    "performer": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "label",
+        "positive_prompt",
+        "negative_prompt",
+        "segmentation_prompt"
+      ],
+      "properties": {
+        "label": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 120
+        },
+        "positive_prompt": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 4000
+        },
+        "negative_prompt": {
+          "type": "string",
+          "maxLength": 2000
+        },
+        "segmentation_prompt": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 320
+        }
+      }
+    },
+    "background": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "positive_prompt",
+        "negative_prompt"
+      ],
+      "properties": {
+        "positive_prompt": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 4000
+        },
+        "negative_prompt": {
+          "type": "string",
+          "maxLength": 2000
+        }
+      }
+    },
+    "fx": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "positive_prompt",
+        "negative_prompt",
+        "blend_mode",
+        "key_mode",
+        "opacity"
+      ],
+      "properties": {
+        "positive_prompt": {
+          "type": "string",
+          "minLength": 1,
+          "maxLength": 4000
+        },
+        "negative_prompt": {
+          "type": "string",
+          "maxLength": 2000
+        },
+        "blend_mode": {
+          "enum": [
+            "screen",
+            "add",
+            "overlay"
+          ]
+        },
+        "key_mode": {
+          "enum": [
+            "luminance",
+            "none"
+          ]
+        },
+        "opacity": {
+          "type": "number",
+          "minimum": 0.05,
+          "maximum": 1.0
+        }
+      }
+    },
+    "editing": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "background_density",
+        "fx_density",
+        "music_reactivity"
+      ],
+      "properties": {
+        "background_density": {
+          "enum": [
+            "calm",
+            "musical",
+            "energetic"
+          ]
+        },
+        "fx_density": {
+          "enum": [
+            "calm",
+            "musical",
+            "energetic"
+          ]
+        },
+        "music_reactivity": {
+          "type": "number",
+          "minimum": 0.0,
+          "maximum": 1.0
+        }
+      }
+    },
+    "finish": {
+      "type": "object",
+      "additionalProperties": false,
+      "required": [
+        "style_preset"
+      ],
+      "properties": {
+        "style_preset": {
+          "enum": [
+            "broadcast_clean",
+            "cable_access",
+            "late_night_vhs",
+            "damaged_tape"
+          ]
+        }
+      }
+    }
+  }
+}
diff --git a/scripts/build_ltx_reframe_outpaint_workflow.py b/scripts/build_ltx_reframe_outpaint_workflow.py
new file mode 100644
index 0000000..9f19a41
--- /dev/null
+++ b/scripts/build_ltx_reframe_outpaint_workflow.py
@@ -0,0 +1,301 @@
+from __future__ import annotations
+
+import json
+import shutil
+import uuid
+from pathlib import Path
+from typing import Any
+
+
+ROOT = Path(__file__).resolve().parents[1]
+SOURCE = (
+    ROOT.parent
+    / "ComfyUI-LTXVideo"
+    / "example_workflows"
+    / "2.3"
+    / "LTX-2.3_ICLoRA_Outpaint_Two_Stage_Distilled.json"
+)
+OUTPUT = ROOT / "examples" / "08_ltx23_reframe_iclora_outpaint_two_stage.json"
+INSTALLED_OUTPUT = (
+    ROOT.parents[1]
+    / "user"
+    / "default"
+    / "workflows"
+    / "LTX 2.3 Reframe IC-LoRA Outpaint - Verified.json"
+)
+
+REMOVED_NODE_IDS = {5168, 5363, 5371, 5393}
+REMOVED_LINK_IDS = {14383, 14395, 14445, 14446}
+
+
+def node_by_id(workflow: dict[str, Any], node_id: int) -> dict[str, Any]:
+    return next(node for node in workflow["nodes"] if node["id"] == node_id)
+
+
+def build_reframe_node() -> dict[str, Any]:
+    return {
+        "id": 5368,
+        "type": "LTXReframeLayout",
+        "pos": [-4225.854240441995, 3547.938320763212],
+        "size": [600, 760],
+        "flags": {},
+        "order": 12,
+        "mode": 0,
+        "inputs": [],
+        "outputs": [
+            {"name": "padded_frames", "type": "IMAGE", "links": [14364]},
+            {"name": "outpaint_mask", "type": "MASK", "links": [14369]},
+            {"name": "source_mask", "type": "MASK", "links": []},
+            {"name": "source_frames", "type": "IMAGE", "links": []},
+            {"name": "audio", "type": "AUDIO", "links": [14422]},
+            {"name": "fps", "type": "FLOAT", "links": [13889, 13968]},
+            {"name": "source_video", "type": "VIDEO", "links": []},
+            {"name": "target_width", "type": "INT", "links": []},
+            {"name": "target_height", "type": "INT", "links": []},
+            {"name": "placement_json", "type": "STRING", "links": []},
+        ],
+        "title": "1. LTX Reframe Layout · Source & Framing",
+        "properties": {"Node name for S&R": "LTXReframeLayout"},
+        "widgets_values": [
+            "LTX-2_00180_.mp4",
+            "9:16",
+            576,
+            1024,
+            0.5,
+            0.5,
+            0.7,
+            32,
+        ],
+        "color": "#1f4d52",
+        "bgcolor": "#16383c",
+    }
+
+
+def build_mask_batch_node(
+    node_id: int,
+    *,
+    pos: list[float],
+    mask_link: int,
+    size_link: int,
+    output_link: int,
+    title: str,
+) -> dict[str, Any]:
+    return {
+        "id": node_id,
+        "type": "MaskExpandBatch+",
+        "pos": pos,
+        "size": [270, 106],
+        "flags": {},
+        "order": 63 if node_id == 5400 else 64,
+        "mode": 0,
+        "inputs": [
+            {"name": "mask", "type": "MASK", "link": mask_link},
+            {
+                "name": "size",
+                "type": "INT",
+                "widget": {"name": "size"},
+                "link": size_link,
+            },
+        ],
+        "outputs": [{"name": "MASK", "type": "MASK", "links": [output_link]}],
+        "title": title,
+        "properties": {
+            "cnr_id": "comfyui_essentials",
+            "Node name for S&R": "MaskExpandBatch+",
+        },
+        "widgets_values": [121, "expand"],
+        "color": "#5b4636",
+        "bgcolor": "#3c3027",
+    }
+
+
+def clean_graph_references(workflow: dict[str, Any]) -> None:
+    valid_node_ids = {node["id"] for node in workflow["nodes"]}
+    workflow["links"] = [
+        link
+        for link in workflow["links"]
+        if link[0] not in REMOVED_LINK_IDS
+        and link[1] in valid_node_ids
+        and link[3] in valid_node_ids
+    ]
+    valid_link_ids = {link[0] for link in workflow["links"]}
+
+    for node in workflow["nodes"]:
+        for item in node.get("inputs", []):
+            if item.get("link") not in valid_link_ids:
+                item["link"] = None
+        for item in node.get("outputs", []):
+            item["links"] = [
+                link_id for link_id in item.get("links") or [] if link_id in valid_link_ids
+            ]
+
+    extra = workflow.get("extra", {})
+    reroutes = extra.get("reroutes", [])
+    for reroute in reroutes:
+        reroute["linkIds"] = [
+            link_id for link_id in reroute.get("linkIds", []) if link_id in valid_link_ids
+        ]
+    extra["reroutes"] = [reroute for reroute in reroutes if reroute.get("linkIds")]
+    extra["linkExtensions"] = [
+        extension
+        for extension in extra.get("linkExtensions", [])
+        if extension.get("id") in valid_link_ids
+    ]
+
+    for group in workflow.get("groups", []):
+        group["nodes"] = [node_id for node_id in group.get("nodes", []) if node_id in valid_node_ids]
+
+    workflow["last_node_id"] = max(valid_node_ids)
+    workflow["last_link_id"] = max(valid_link_ids)
+
+
+def validate_graph(workflow: dict[str, Any]) -> None:
+    nodes = {node["id"]: node for node in workflow["nodes"]}
+    links = {link[0]: link for link in workflow["links"]}
+    if len(nodes) != len(workflow["nodes"]):
+        raise ValueError("Duplicate node IDs remain in the generated workflow.")
+    if len(links) != len(workflow["links"]):
+        raise ValueError("Duplicate link IDs remain in the generated workflow.")
+
+    for link_id, link in links.items():
+        if link[1] not in nodes or link[3] not in nodes:
+            raise ValueError(f"Link {link_id} has a missing endpoint: {link}")
+        source_output = nodes[link[1]]["outputs"][link[2]]
+        destination_input = nodes[link[3]]["inputs"][link[4]]
+        if link_id not in (source_output.get("links") or []):
+            raise ValueError(f"Link {link_id} is absent from its source output socket.")
+        if destination_input.get("link") != link_id:
+            raise ValueError(f"Link {link_id} is absent from its destination input socket.")
+
+    reframe = nodes.get(5368)
+    if reframe is None or reframe["type"] != "LTXReframeLayout":
+        raise ValueError("The generated workflow is missing LTXReframeLayout node 5368.")
+
+    expected = {
+        14364: (5368, 0, 5360, 0, "IMAGE"),
+        14369: (5368, 1, 5365, 0, "MASK"),
+        13889: (5368, 5, 1241, 2, "FLOAT"),
+        13968: (5368, 5, 5227, 2, "FLOAT"),
+        14422: (5368, 4, 5382, 0, "AUDIO"),
+        14376: (5365, 0, 5400, 0, "MASK"),
+        14377: (5364, 0, 5401, 0, "MASK"),
+        14447: (5054, 2, 5400, 1, "INT"),
+        14448: (5054, 2, 5401, 1, "INT"),
+        14449: (5400, 0, 5226, 2, "MASK"),
+        14450: (5401, 0, 5266, 2, "MASK"),
+    }
+    for link_id, contract in expected.items():
+        actual = tuple(links[link_id][1:])
+        if actual != contract:
+            raise ValueError(f"Link {link_id} is {actual}, expected {contract}.")
+
+
+def build() -> dict[str, Any]:
+    workflow = json.loads(SOURCE.read_text(encoding="utf-8"))
+    workflow["id"] = str(
+        uuid.uuid5(uuid.NAMESPACE_URL, "ltx-2.3-reframe-iclora-outpaint-two-stage")
+    )
+    workflow["nodes"] = [
+        node for node in workflow["nodes"] if node["id"] not in REMOVED_NODE_IDS | {5368}
+    ]
+    workflow["nodes"].append(build_reframe_node())
+    workflow["nodes"].extend(
+        [
+            build_mask_batch_node(
+                5400,
+                pos=[1050.0, 3261.0],
+                mask_link=14376,
+                size_link=14447,
+                output_link=14449,
+                title="Full-Res Mask · Match Video Frame Count",
+            ),
+            build_mask_batch_node(
+                5401,
+                pos=[-660.0, 3432.0],
+                mask_link=14377,
+                size_link=14448,
+                output_link=14450,
+                title="Half-Res Mask · Match Video Frame Count",
+            ),
+        ]
+    )
+
+    links = {link[0]: link for link in workflow["links"]}
+    links[14364] = [14364, 5368, 0, 5360, 0, "IMAGE"]
+    links[14369] = [14369, 5368, 1, 5365, 0, "MASK"]
+    links[13889] = [13889, 5368, 5, 1241, 2, "FLOAT"]
+    links[13968] = [13968, 5368, 5, 5227, 2, "FLOAT"]
+    links[14422] = [14422, 5368, 4, 5382, 0, "AUDIO"]
+    links[14376] = [14376, 5365, 0, 5400, 0, "MASK"]
+    links[14377] = [14377, 5364, 0, 5401, 0, "MASK"]
+    links[14447] = [14447, 5054, 2, 5400, 1, "INT"]
+    links[14448] = [14448, 5054, 2, 5401, 1, "INT"]
+    links[14449] = [14449, 5400, 0, 5226, 2, "MASK"]
+    links[14450] = [14450, 5401, 0, 5266, 2, "MASK"]
+    workflow["links"] = list(links.values())
+
+    node_by_id(workflow, 5226)["inputs"][2]["link"] = 14449
+    node_by_id(workflow, 5266)["inputs"][2]["link"] = 14450
+    node_by_id(workflow, 5054)["outputs"][2]["links"].extend([14447, 14448])
+
+    node_by_id(workflow, 5023)["widgets_values"][0] = (
+        "gemma_3_12B_it_fp4_mixed.safetensors"
+    )
+    node_by_id(workflow, 4922)["widgets_values"][0] = (
+        "ltx-2.3-22b-distilled-lora-384-1.1.safetensors"
+    )
+    node_by_id(workflow, 5011)["widgets_values"][0] = (
+        "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors"
+    )
+    node_by_id(workflow, 2483)["widgets_values"][0] = (
+        "A cinematic live-action scene continuing naturally beyond the source frame, "
+        "with a coherent environment, realistic lighting, consistent subjects and camera "
+        "motion, seamless edges, and stable fine detail."
+    )
+    node_by_id(workflow, 5228)["widgets_values"][0] = "video/LTX_Reframe_Outpaint"
+    node_by_id(workflow, 5356)["widgets_values"][0] = (
+        "### Reframe Setup\n\nUse **LTX Reframe Layout** to choose the final canvas and place "
+        "the source. `source_scale < 1` zooms out for outpainting; `source_scale > 1` "
+        "zooms in and crops. The example starts at **576 × 1024**, 121 frames, with "
+        "audio, so it is suitable for a first end-to-end test."
+    )
+
+    for group in workflow.get("groups", []):
+        if group["id"] == 164:
+            group["title"] = "1. Reframe Source Video"
+            group["bounding"] = [-4235.854240441995, 3507.938320763212, 620, 800]
+        elif group["id"] == 168:
+            group["title"] = "2. Prepare Two-Stage Outpaint Conditions"
+        elif group["id"] == 178:
+            group["title"] = "Stage 2 Full-Resolution Mask"
+        elif group["id"] == 179:
+            group["title"] = "Stage 1 Half-Resolution Mask"
+
+    workflow.setdefault("extra", {})["ltx_reframe_workflow"] = {
+        "schema": "ltx_reframe_iclora_outpaint/v1",
+        "based_on": SOURCE.name,
+        "source_node": "LTXReframeLayout",
+        "default_source": "LTX-2_00180_.mp4",
+        "default_target": [576, 1024],
+        "required_lora": "ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors",
+        "mask_batch_safety": "MaskExpandBatch+ before both Laplacian blends",
+    }
+
+    clean_graph_references(workflow)
+    validate_graph(workflow)
+    return workflow
+
+
+def main() -> None:
+    workflow = build()
+    OUTPUT.parent.mkdir(parents=True, exist_ok=True)
+    OUTPUT.write_text(json.dumps(workflow, indent=2) + "\n", encoding="utf-8")
+    INSTALLED_OUTPUT.parent.mkdir(parents=True, exist_ok=True)
+    shutil.copy2(OUTPUT, INSTALLED_OUTPUT)
+    print(f"Wrote {OUTPUT}")
+    print(f"Installed {INSTALLED_OUTPUT}")
+    print(f"Nodes: {len(workflow['nodes'])}; links: {len(workflow['links'])}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/dg_to_ltx_prompt_injector.py b/scripts/dg_to_ltx_prompt_injector.py
index 1443c9a..123beb5 100644
--- a/scripts/dg_to_ltx_prompt_injector.py
+++ b/scripts/dg_to_ltx_prompt_injector.py
@@ -9,9 +9,9 @@ from pathlib import Path
 from typing import Any
 
 
-DEFAULT_WORKFLOW = Path(r"C:\Users\danrh\Desktop\diffusion_gemma_workflows\LTX-2.3-INPAINTING-EDIT-ANY-VIDEO.json")
-DEFAULT_OUTPUT_WORKFLOW = Path(r"C:\Users\danrh\Desktop\diffusion_gemma_workflows\LTX-2.3-INPAINTING-EDIT-ANY-VIDEO_DG_PROMPT_INJECTED_v0.json")
-DEFAULT_MANIFEST = Path(r"C:\Users\danrh\Desktop\diffusion_gemma_workflows\LTX-2.3-INPAINTING-EDIT-ANY-VIDEO_DG_PROMPT_INJECTED_v0.manifest.json")
+DEFAULT_WORKFLOW = Path("LTX-2.3-INPAINTING-EDIT-ANY-VIDEO.json")
+DEFAULT_OUTPUT_WORKFLOW = Path("LTX-2.3-INPAINTING-EDIT-ANY-VIDEO_DG_PROMPT_INJECTED_v0.json")
+DEFAULT_MANIFEST = Path("LTX-2.3-INPAINTING-EDIT-ANY-VIDEO_DG_PROMPT_INJECTED_v0.manifest.json")
 REPAIR_SCOPE = "localized_masked_inpaint"
 
 
diff --git a/scripts/smoke_prompt_builder.py b/scripts/smoke_prompt_builder.py
index 661d0b2..9d1f362 100644
--- a/scripts/smoke_prompt_builder.py
+++ b/scripts/smoke_prompt_builder.py
@@ -27,9 +27,13 @@ def _packet(prompt: str, metadata: dict | None = None, segments: list[dict] | No
     }
 
 
-def _split(packet_or_text, context=None):
+def _split(packet_or_text, context=None, target=None):
     text = packet_or_text if isinstance(packet_or_text, str) else json.dumps(packet_or_text)
-    return dg.DiffusionGemmaJSONSplitter().split(text, gemma_context=context)
+    return dg.DiffusionGemmaJSONSplitter().split(
+        text,
+        gemma_context=context,
+        target_profile_config=target,
+    )
 
 
 def _metadata(result):
@@ -50,6 +54,23 @@ def _media_context(user_prompt: str, media: dict, visual_description: str = ""):
     )
 
 
+def _h3_ref_context(user_prompt: str, manifest: str):
+    normalized_manifest = dg._normalize_minimax_h3_reference_manifest(manifest)
+    definitions = dg._minimax_h3_reference_definitions(normalized_manifest)
+    return dg.GemmaContext(
+        user_prompt=user_prompt,
+        source="minimax_h3_ref2va",
+        media_metadata={
+            "minimax_h3_mode": "ref2va",
+            "minimax_h3_reference_manifest": normalized_manifest,
+            "minimax_h3_reference_tags": [tag for tag, _description in definitions],
+            "minimax_h3_reference_manifest_reasons": dg._minimax_h3_reference_manifest_validation_reasons(
+                normalized_manifest
+            ),
+        },
+    )
+
+
 def _assert_ready(result):
     assert result[8] is True, _metadata(result)
     assert result[12] is True, _metadata(result)
@@ -128,13 +149,1344 @@ def test_readiness_gates():
 def test_prompt_packet_schema():
     schema = json.loads((ROOT / "schemas" / "prompt_packet.schema.json").read_text(encoding="utf-8"))
     validator = Draft202012Validator(schema)
+
+    # Packets created before MiniMax H3 support remain valid.
     valid_packet = _packet("A clean prompt.", {"ready_for_generation": True, "blocked_reasons": []})
     validator.validate(valid_packet)
+
+    h3_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    h3_packet["minimax_h3_prompt"] = (
+        "integrated_multimodal_description: [Shot 1] A locked-off wide shot holds on a rain-darkened city roof.\n\n"
+        "overall_soundscape: Wind moves across the roof while distant traffic hums below.\n\n"
+        "non_diegetic_music: N/A"
+    )
+    validator.validate(h3_packet)
+
+    invalid_h3_packet = dict(h3_packet, minimax_h3_prompt=["not", "a", "string"])
+    assert list(validator.iter_errors(invalid_h3_packet)), "schema should reject a non-string H3 prompt"
+
     invalid_packet = {"ltx_prompt": "missing required fields"}
     errors = sorted(validator.iter_errors(invalid_packet), key=lambda item: item.path)
     assert errors, "schema should reject incomplete prompt packets"
 
 
+def test_minimax_h3_contract_and_splitter_compatibility():
+    profile_choices = dg.DiffusionGemmaTargetProfile.INPUT_TYPES()["required"]["target_profile"][0]
+    assert "minimax_h3" in profile_choices, profile_choices
+
+    target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=20.0,
+        audio_mode="auto_scene_audio",
+    )
+    model_prompt = dg._build_model_prompt(
+        "A runner crosses rain-darkened rooftops in four hard-cut shots.",
+        dg.DEFAULT_MASTER_PROMPT,
+        "minimax_h3",
+        {},
+        target_duration_seconds=20.0,
+        audio_mode="auto_scene_audio",
+        thinking_mode="on",
+    )
+    lowered = model_prompt.lower()
+    assert dg._target_profile_config_to_dict(target)["target_duration_seconds"] == 20.0
+    assert "minimax_h3_prompt" in model_prompt
+    assert "integrated_multimodal_description" in model_prompt
+    assert "overall_soundscape" in model_prompt
+    assert "non_diegetic_music" in model_prompt
+    assert "20" in model_prompt, "20-second H3 requests must not be clamped to the published 15-second guidance"
+    assert "target profile: minimax_h3" in lowered
+    assert "composed for 1:1" not in lowered, "the Ideogram-only aspect control must not silently frame H3"
+    for phrase in (
+        "binding directing specification",
+        "continuity ledger",
+        "generic real animals",
+        "prior shot's resulting position and state",
+        "shot size, angle, or the word handheld alone is not a camera path",
+        "event-specific sound immediately beside its visible cause",
+        "the final frame holds on",
+    ):
+        assert phrase in lowered, phrase
+
+    inferred_duration_prompt = dg._build_model_prompt(
+        "Make a 20-second rooftop chase.",
+        dg.DEFAULT_MASTER_PROMPT,
+        "minimax_h3",
+        {},
+        target_duration_seconds=0.0,
+        audio_mode="auto_scene_audio",
+        thinking_mode="on",
+    )
+    assert "planned for exactly 20.000 seconds" in inferred_duration_prompt, inferred_duration_prompt
+
+    h3_prompt = (
+        "integrated_multimodal_description: [Shot 1] A high side angle tracks a runner toward a wet rooftop edge.\n\n"
+        "[Shot 2] At 00:10.000, the camera cuts to a wide profile as the runner clears the gap.\n\n"
+        "overall_soundscape: Strong wind, rapid footsteps, and distant city traffic surround the action.\n\n"
+        "non_diegetic_music: A low percussion pulse builds under the chase."
+    )
+    packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    packet["minimax_h3_prompt"] = h3_prompt
+    result = _split(packet, context=None, target=target)
+
+    assert dg.DiffusionGemmaJSONSplitter.RETURN_NAMES[:13] == (
+        "ltx_prompt",
+        "ideogram_prompt",
+        "negative_prompt",
+        "aspect_ratio",
+        "metadata_json",
+        "scene_segments_json",
+        "local_prompts",
+        "segment_lengths",
+        "is_valid",
+        "resolution_selector_preset",
+        "resolution_width",
+        "resolution_height",
+        "ready_for_generation",
+    )
+    assert dg.DiffusionGemmaJSONSplitter.RETURN_NAMES[13] == "minimax_h3_prompt"
+    assert len(result) == 14, result
+    assert result[13] == h3_prompt, result[13]
+    assert "\n\noverall_soundscape:" in result[13], result[13]
+    assert "\n\nnon_diegetic_music:" in result[13], result[13]
+    assert result[0] == "", result[0]
+    assert result[1] == "", result[1]
+    assert result[2] == "", result[2]
+    assert result[8] is True, _metadata(result)
+    assert result[12] is True, _metadata(result)
+
+    model_style_h3_prompt = (
+        "[Shot 1] A low-angle tracking shot follows a runner across a rain-darkened rooftop. "
+        "[Shot 2] At 00:04.500 the camera cuts to a wide profile as the runner clears the gap.\n\n"
+        "Strong wind, rapid footsteps, splashing puddles, and distant city traffic surround the action.\n\n"
+        "A restrained electronic score uses deep cello pulses and measured percussion, building with each leap."
+    )
+    model_style_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    model_style_packet["minimax_h3_prompt"] = model_style_h3_prompt
+    model_style_result = _split(model_style_packet, context=None, target=target)
+    normalized_h3_prompt = model_style_result[13]
+    assert model_style_result[12] is True, _metadata(model_style_result)
+    assert normalized_h3_prompt.startswith("integrated_multimodal_description: [Shot 1]"), normalized_h3_prompt
+    assert "[Shot 2] At 00:04.500, the camera cuts to" in normalized_h3_prompt, normalized_h3_prompt
+    assert "\n\noverall_soundscape: Strong wind" in normalized_h3_prompt, normalized_h3_prompt
+    assert "\n\nnon_diegetic_music: A restrained electronic score" in normalized_h3_prompt, normalized_h3_prompt
+    assert dg._normalize_minimax_h3_prompt(normalized_h3_prompt) == normalized_h3_prompt
+
+    unlabeled_cut_prompt = (
+        "integrated_multimodal_description: [Shot 1] A locked-off wide shot holds on a rain-darkened roof. "
+        "At 00:04.500, the camera cuts to a low-angle tracking shot beside the runner. "
+        "At 00:09.000, the camera cuts to an overhead static hold as the runner lands.\n\n"
+        "overall_soundscape: Wind, footsteps, and splashing water follow the action.\n\n"
+        "non_diegetic_music: Low percussion builds under the cuts."
+    )
+    normalized_unlabeled_cut_prompt = dg._normalize_minimax_h3_prompt(unlabeled_cut_prompt)
+    assert "[Shot 2] At 00:04.500," in normalized_unlabeled_cut_prompt, normalized_unlabeled_cut_prompt
+    assert "[Shot 3] At 00:09.000," in normalized_unlabeled_cut_prompt, normalized_unlabeled_cut_prompt
+    assert "minimax_h3_unlabeled_cut" not in dg._minimax_h3_prompt_validation_reasons(
+        normalized_unlabeled_cut_prompt,
+        20.0,
+    )
+
+    ambiguous_unlabeled_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    ambiguous_unlabeled_packet["minimax_h3_prompt"] = (
+        "[Shot 1] A locked-off view holds on a rooftop.\n\nA second unlabeled visual paragraph follows."
+    )
+    ambiguous_unlabeled_result = _split(ambiguous_unlabeled_packet, context=None, target=target)
+    assert ambiguous_unlabeled_result[13] == "", ambiguous_unlabeled_result[13]
+    assert "minimax_h3_missing_integrated_multimodal_description" in _metadata(ambiguous_unlabeled_result)["blocked_reasons"]
+
+    invalid_cases = (
+        (
+            h3_prompt.replace("[Shot 2] At 00:10.000,", "[10s-20s] Shot 2:"),
+            "minimax_h3_noncanonical_time_range",
+        ),
+        (
+            h3_prompt.replace("At 00:10.000,", "At 00:20.000,"),
+            "minimax_h3_cut_timestamp_out_of_range",
+        ),
+        (
+            h3_prompt.replace("[Shot 1]", "[Shot 1] At 00:00.000,"),
+            "minimax_h3_shot1_timestamp_invalid",
+        ),
+        (
+            h3_prompt.rsplit("non_diegetic_music:", 1)[0] + "non_diegetic_music:",
+            "minimax_h3_empty_non_diegetic_music",
+        ),
+        (
+            h3_prompt.replace(
+                "[Shot 1] A high side angle tracks a runner toward a wet rooftop edge.",
+                "[Shot 1]",
+            ),
+            "minimax_h3_empty_visual_timeline",
+        ),
+        (
+            h3_prompt.replace("[Shot 2]", "[shot 2]"),
+            "minimax_h3_shot_syntax_invalid",
+        ),
+        (
+            h3_prompt.replace("\n\noverall_soundscape:", "\n\nCamera: handheld.\n\noverall_soundscape:"),
+            "minimax_h3_extra_top_level_field",
+        ),
+        (
+            h3_prompt.replace(
+                "A high side angle tracks a runner toward a wet rooftop edge.",
+                "A high side angle tracks a runner who says <d>[English] Jump now.</d> near a wet rooftop edge.",
+            ),
+            "minimax_h3_dialogue_speaker_invalid",
+        ),
+    )
+    for invalid_prompt, expected_reason in invalid_cases:
+        invalid_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+        invalid_packet["minimax_h3_prompt"] = invalid_prompt
+        invalid_result = _split(invalid_packet, context=None, target=target)
+        invalid_metadata = _metadata(invalid_result)
+        assert invalid_result[13] == "", invalid_result[13]
+        assert expected_reason in invalid_metadata["blocked_reasons"], invalid_metadata
+
+    visual_only_target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=20.0,
+        audio_mode="visual_only",
+    )
+    visual_only_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    visual_only_packet["minimax_h3_prompt"] = h3_prompt
+    visual_only_result = _split(visual_only_packet, context=None, target=visual_only_target)
+    assert visual_only_result[13] == "", visual_only_result[13]
+    assert "minimax_h3_visual_only_audio_fields_invalid" in _metadata(visual_only_result)["blocked_reasons"]
+
+    requested_transition_context = dg.GemmaContext(user_prompt="Use a cross-dissolve for the second shot.", media_metadata={})
+    requested_transition_prompt = h3_prompt.replace("the camera cuts to", "the camera cross-dissolves to")
+    requested_transition_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    requested_transition_packet["minimax_h3_prompt"] = requested_transition_prompt
+    requested_transition_result = _split(requested_transition_packet, requested_transition_context, target)
+    assert requested_transition_result[13] == requested_transition_prompt, _metadata(requested_transition_result)
+
+    oversized_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    oversized_packet["minimax_h3_prompt"] = h3_prompt.replace(
+        "A high side angle tracks a runner toward a wet rooftop edge.",
+        "A high side angle tracks a runner toward a wet rooftop edge. " + ("Visible rain streaks across the frame. " * 260),
+    )
+    oversized_result = _split(oversized_packet, context=None, target=target)
+    assert oversized_result[13] == "", oversized_result[13]
+    assert "minimax_h3_prompt_truncated" in _metadata(oversized_result)["blocked_reasons"]
+
+
+def test_minimax_h3_ref2va_contract():
+    mode_input = dg.DiffusionGemmaTargetProfile.INPUT_TYPES()["optional"]["minimax_h3_mode"]
+    assert mode_input[0] == ["t2va", "ref2va"], mode_input
+    assert mode_input[1]["default"] == "t2va", mode_input
+    default_target = dg._make_target_profile_config(target_profile="minimax_h3", target_duration_seconds=8.0)
+    assert dg._target_profile_config_to_dict(default_target)["minimax_h3_mode"] == "t2va"
+
+    default_t2va_prompt = (
+        "integrated_multimodal_description: [Shot 1] A static wide shot holds on a quiet harbor at dawn. "
+        "The final frame holds on the anchored boats beneath the pale sky.\n\n"
+        "overall_soundscape: Small waves touch the dock while ropes creak in the breeze.\n\n"
+        "non_diegetic_music: N/A"
+    )
+    default_t2va_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    default_t2va_packet["minimax_h3_prompt"] = default_t2va_prompt
+    default_t2va_result = _split(default_t2va_packet, target=default_target)
+    assert default_t2va_result[13] == default_t2va_prompt, _metadata(default_t2va_result)
+    assert default_t2va_result[12] is True, _metadata(default_t2va_result)
+
+    manifest = (
+        "<Picture 1>: first-frame composition and primary identity reference for the rooftop courier; fully preserve the face, "
+        "short black hair, red waxed-canvas coat, silver satchel, body proportions, and opening screen position.\n"
+        "<Audio 1>: synchronized soundtrack reference for <Video 1>; reference its pulse, footstep timing, wind texture, and impact "
+        "accents without copying speech.\n"
+        "<Video 1>: motion and camera reference; transfer the running cadence, lateral tracking path, hard-cut rhythm, and landing "
+        "mechanics, but not the source performer's identity or wardrobe."
+    )
+    assert dg._minimax_h3_reference_manifest_validation_reasons(manifest) == []
+    assert [tag for tag, _description in dg._minimax_h3_reference_definitions(manifest)] == [
+        "<Picture 1>",
+        "<Audio 1>",
+        "<Video 1>",
+    ]
+    context = _h3_ref_context(
+        "Create an eight-second live-action rooftop escape. Use Picture 1 for the courier and opening composition, Video 1 for "
+        "motion and camera rhythm, and Audio 1 for timing and sound texture.",
+        manifest,
+    )
+    target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=8.0,
+        audio_mode="auto_scene_audio",
+        minimax_h3_mode="ref2va",
+    )
+    model_prompt = dg._build_model_prompt(
+        context.user_prompt,
+        dg.DEFAULT_MASTER_PROMPT,
+        "minimax_h3",
+        context.media_metadata,
+        target_duration_seconds=8.0,
+        audio_mode="auto_scene_audio",
+        minimax_h3_mode="ref2va",
+        minimax_h3_reference_manifest=manifest,
+    )
+    assert "Final MiniMax H3 Ref2VA full-reference prompt" in model_prompt
+    assert "subject_definitions, summary, retention_analysis, detailed_description" in model_prompt
+    assert "a semantic reference, not a guaranteed first-frame or last-frame pixel lock" in model_prompt
+    assert manifest in model_prompt
+    ref_prompt = (
+        "subject_definitions:\n"
+        "<Subject 1> is the rooftop courier from <Picture 1>, with an angular face, short black hair, a red waxed-canvas coat, "
+        "a silver cross-body satchel, lean body proportions, and black running boots.\n"
+        "<Picture 1> is the first frame of [Shot 1], fixing the courier on the left third of a rain-darkened roof with a low "
+        "parapet ahead and the blue dusk skyline behind.\n"
+        "<Video 1> is the whole-video motion and camera reference for running cadence, lateral tracking, hard-cut rhythm, and the "
+        "physical mechanics of the final landing; its performer identity and wardrobe are not transferred.\n"
+        "<Audio 1> is the synchronized sound reference for the running pulse, timed footsteps, wind texture, and landing accent; "
+        "no source speech is copied.\n\n"
+        "summary:\n"
+        "[keyframe completion + reference generation + audio reference] The target video begins from <Picture 1> and follows "
+        "<Subject 1> through a three-shot rooftop escape, transferring the motion, camera path, and cut rhythm of <Video 1> while "
+        "using <Audio 1> only as timing and sound-texture guidance.\n\n"
+        "retention_analysis:\n"
+        "<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the angular face, short black hair, red coat, "
+        "silver satchel, lean proportions, and black boots remain unchanged.\n"
+        "<Picture 1> ([Shot 1] first frame): fully_preserved - the opening subject placement, wet parapet, and blue dusk skyline "
+        "are retained as the exact opening composition.\n"
+        "<Video 1> (running cadence, camera path, cut rhythm, and landing mechanics): attribute_transfer - its temporal behavior "
+        "is transferred to <Subject 1> without copying the source performer or wardrobe.\n"
+        "<Audio 1>: reference - its pulse, footstep timing, wind texture, and landing accent guide newly generated audio without "
+        "copying speech or the original signal.\n\n"
+        "detailed_description:\n"
+        "The target video uses realistic live-action practical film photography with an anamorphic lens, restrained cool-blue dusk "
+        "grading, shallow depth of field, fine film grain, and rain-slick textures. The courier identity, wardrobe, weather, and "
+        "skyline geography remain continuous through all three hard-cut shots.\n"
+        "[Shot 1] The shot begins from <Picture 1> exactly: a medium-wide side view places <Subject 1> on the left third of the "
+        "rain-darkened roof, with the angular face in profile, short black hair damp, red coat darkened at the shoulders, and silver "
+        "satchel against the right hip. A low parapet crosses the middle distance and the blue skyline recedes through mist. After a "
+        "readable hold, <Subject 1> drives forward into the cadence referenced from <Video 1>, pushing off with the left boot. The "
+        "camera trucks right smoothly at the runner's speed, keeping the torso centered while foreground vents slide left with clear "
+        "parallax. Each newly generated boot impact follows <Audio 1>; wet rubber hits the roof, puddles splash, coat fabric snaps, "
+        "and the runner exhales without speaking. The courier reaches the parapet with the right foot planted and eyes on the next roof.\n"
+        "[Shot 2] At 00:02.750, the camera hard cuts to a wide profile view that follows the cut rhythm of <Video 1>. <Subject 1> "
+        "springs from the planted right foot, clears the parapet, and crosses the narrow alley in one continuous jump. The red coat "
+        "opens behind the lean silhouette, the satchel lifts while its strap remains across the chest, and both boots stay visible "
+        "against the skyline. The camera tracks right at fast speed with medium amplitude, matching the body without orbiting, zooming, "
+        "or changing screen direction. Warm windows and wet reflections streak through shallow focus below. Wind rises around the "
+        "airborne body and the pulse referenced from <Audio 1> tightens without copied speech. The courier extends both legs toward "
+        "the far roof and braces the left arm, making the landing preparation visible before contact.\n"
+        "[Shot 3] At 00:05.500, the camera hard cuts to a low three-quarter tracking view beside the destination roof. <Subject 1> "
+        "lands with the mechanics referenced from <Video 1>: both boots contact in sequence, knees compress, the left hand touches "
+        "the wet surface, and momentum becomes one controlled shoulder roll. The new landing accent follows <Audio 1>, combining a "
+        "heavy impact, satchel-buckle clink, fabric scrape, and water spray. The camera shakes slightly on impact, then steadies and "
+        "trucks right as the same courier rises without changing face, hair, coat, satchel, proportions, or screen direction. The "
+        "courier takes two strides, stops behind a ventilation housing, and looks back across the completed gap while breathing with "
+        "closed lips. The final frame holds on <Subject 1> crouched safely on the right third, the satchel at the right hip, with the "
+        "empty gap and blue skyline behind.\n\n"
+        "overall_soundscape:\n"
+        "Cross-roof wind, distant traffic, wet boot impacts, puddle splashes, coat movement, breath, and the final buckle clink are "
+        "newly generated, with their cadence and texture guided by <Audio 1>. No dialogue or copied source signal is audible.\n\n"
+        "non_diegetic_music:\n"
+        "A restrained low electronic pulse follows the tempo reference of <Audio 1>, tightening during the jump and dropping out "
+        "on the landing accent before one sustained bass note closes the final hold."
+    )
+    sections = dg._minimax_h3_ref_sections(ref_prompt)
+    assert sections is not None
+    assert 350 <= len(sections["detailed_description"].split()) <= 500
+    assert dg._normalize_minimax_h3_prompt(ref_prompt, "ref2va") == ref_prompt
+
+    packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    packet["minimax_h3_prompt"] = ref_prompt
+    result = _split(packet, context=context, target=target)
+    assert result[13] == ref_prompt, _metadata(result)
+    assert result[12] is True, _metadata(result)
+
+    invalid_manifests = (
+        ("", "minimax_h3_ref_manifest_missing"),
+        (
+            "<Picture 1>: identity and wardrobe reference for the courier.\n"
+            "<Picture 1>: duplicate composition reference for the same courier.",
+            "minimax_h3_ref_manifest_duplicate_tag",
+        ),
+        (
+            "<Picture 2>: identity and wardrobe reference for the courier.",
+            "minimax_h3_ref_manifest_nonconsecutive_tags",
+        ),
+        (
+            "<Picture 10>: identity and wardrobe reference for the courier.",
+            "minimax_h3_ref_manifest_limit_exceeded",
+        ),
+        (
+            "<Audio 1>: voice timing and soundtrack rhythm reference only.",
+            "minimax_h3_ref_manifest_audio_only",
+        ),
+        (
+            "<Picture 1>: identity. Color.",
+            "minimax_h3_ref_manifest_role_missing",
+        ),
+        (
+            "<Picture 1>: identity and wardrobe reference taken from <Video 1>.",
+            "minimax_h3_ref_manifest_undefined_tag",
+        ),
+    )
+    for invalid_manifest, expected_reason in invalid_manifests:
+        reasons = dg._minimax_h3_reference_manifest_validation_reasons(invalid_manifest)
+        assert expected_reason in reasons, (expected_reason, reasons)
+        invalid_context = _h3_ref_context(context.user_prompt, invalid_manifest)
+        invalid_result = _split(packet, context=invalid_context, target=target)
+        assert invalid_result[13] == "", _metadata(invalid_result)
+        assert expected_reason in _metadata(invalid_result)["blocked_reasons"], _metadata(invalid_result)
+
+    invalid_subject_usage_sections = sections.copy()
+    invalid_subject_usage_sections["detailed_description"] = sections["detailed_description"].replace(
+        "<Subject 1>",
+        "the rooftop courier",
+    )
+    invalid_subject_usage_prompt = dg._rebuild_minimax_h3_ref_prompt(invalid_subject_usage_sections)
+    invalid_prompts = (
+        (
+            default_t2va_prompt,
+            {"minimax_h3_ref_sections_invalid"},
+        ),
+        (
+            ref_prompt.replace("<Audio 1>", "the soundtrack"),
+            {"minimax_h3_ref_missing_asset_tag"},
+        ),
+        (
+            ref_prompt.replace("<Video 1>", "<Video 2>"),
+            {"minimax_h3_ref_missing_asset_tag", "minimax_h3_ref_undefined_tag"},
+        ),
+        (
+            ref_prompt.replace(
+                "<Subject 1> is the rooftop courier from <Picture 1>, with an angular face, short black hair, a red waxed-canvas "
+                "coat, a silver cross-body satchel, lean body proportions, and black running boots.",
+                "<Subject 1> is <Picture 1>.",
+            ),
+            {"minimax_h3_ref_subject_definition_invalid"},
+        ),
+        (
+            invalid_subject_usage_prompt,
+            {"minimax_h3_ref_subject_usage_invalid"},
+        ),
+        (
+            ref_prompt.replace("fully_preserved", "preserved", 1),
+            {"minimax_h3_ref_retention_invalid"},
+        ),
+    )
+    for invalid_prompt, expected_reasons in invalid_prompts:
+        invalid_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+        invalid_packet["minimax_h3_prompt"] = invalid_prompt
+        invalid_result = _split(invalid_packet, context=context, target=target)
+        invalid_metadata = _metadata(invalid_result)
+        assert invalid_result[13] == "", invalid_metadata
+        assert expected_reasons.issubset(set(invalid_metadata["blocked_reasons"])), invalid_metadata
+
+
+def test_minimax_h3_ref2va_structure_repair_and_generation_gate():
+    manifest = (
+        "<Picture 1>: courier identity and wardrobe reference; preserve the face, short black hair, red coat, silver satchel, "
+        "lean proportions, and black boots."
+    )
+    user_prompt = "Create an eight-second 2D hand-drawn cel-animated rooftop run using Picture 1 as the courier reference."
+    full_sequence_manifest = manifest + " This is a full-sequence reference, not a keyframe."
+    assert dg._minimax_h3_ref_summary_task_types(user_prompt, full_sequence_manifest) == ["reference generation"]
+    colon_subject_prompt = (
+        "subject_definitions:\n"
+        "<Subject 1>: The rooftop courier from <Picture 1>, with an angular face, short black hair, a red waxed-canvas coat, "
+        "a silver cross-body satchel, lean body proportions, and black running boots.\n\n"
+        "summary:\n"
+        "[reference generation] The target video follows <Subject 1> across a rain-darkened rooftop while preserving the identity "
+        "and wardrobe defined by <Picture 1>.\n\n"
+        "retention_analysis:\n"
+        "<Subject 1> (appears in [Shot 1]): fully_preserved - the angular face, short black hair, red coat, silver satchel, lean "
+        "proportions, and black boots remain unchanged.\n\n"
+        "detailed_description:\n"
+        "The target video uses immutable 2D hand-drawn cel animation with crisp inked outlines, flat painted colors, restrained "
+        "two-tone shadows, and no live action, photorealism, or 3D CGI.\n"
+        "[Shot 1] A static medium-wide side view frames <Subject 1> on the left third of a rain-darkened rooftop, with the angular "
+        "face in profile, short black hair damp against the forehead, red waxed-canvas coat darkened at the shoulders, silver satchel "
+        "against the right hip, and black boots planted beside a shallow puddle. The courier leans forward, pushes off with the left "
+        "boot, and accelerates toward a low parapet while the camera trucks right smoothly at the same speed, keeping the torso "
+        "centered as vents slide left with clear parallax. Each footfall throws a small painted splash and the satchel swings once "
+        "behind the hip without changing its strap position. Near the parapet, the courier shortens the final stride, plants the "
+        "right boot, and stops under control rather than beginning a jump. The camera eases to a static hold. The final frame holds "
+        "on <Subject 1> standing safely on the right third, looking across the gap with the same face, hair, coat, satchel, body "
+        "proportions, and boots clearly visible.\n\n"
+        "overall_soundscape:\n"
+        "Cross-roof wind, wet boot impacts, puddle splashes, coat movement, and one quiet breath accompany the run.\n\n"
+        "non_diegetic_music:\n"
+        "N/A"
+    )
+    colon_reasons = dg._minimax_h3_ref2va_validation_reasons(
+        colon_subject_prompt,
+        8.0,
+        user_prompt,
+        "auto_scene_audio",
+        8000,
+        manifest,
+    )
+    assert colon_reasons == [], colon_reasons
+
+    no_colon_headings = colon_subject_prompt
+    for field_name in dg._MINIMAX_H3_REF_FIELDS:
+        no_colon_headings = no_colon_headings.replace(f"{field_name}:\n", f"{field_name}\n", 1)
+    no_colon_repaired = dg._repair_minimax_h3_ref2va_structure(no_colon_headings, user_prompt, manifest)
+    assert dg._minimax_h3_ref_sections(no_colon_repaired) is not None, no_colon_repaired
+    assert dg._minimax_h3_ref2va_validation_reasons(
+        no_colon_repaired,
+        8.0,
+        user_prompt,
+        "auto_scene_audio",
+        8000,
+        manifest,
+    ) == []
+
+    mixed_headings = colon_subject_prompt
+    for field_name in dg._MINIMAX_H3_REF_FIELDS[::2]:
+        mixed_headings = mixed_headings.replace(f"{field_name}:\n", f"{field_name}\n", 1)
+    mixed_repaired = dg._repair_minimax_h3_ref2va_structure(mixed_headings, user_prompt, manifest)
+    assert dg._minimax_h3_ref_sections(mixed_repaired) is not None, mixed_repaired
+
+    colon_context = _h3_ref_context(user_prompt, manifest)
+    ref_target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=8.0,
+        audio_mode="auto_scene_audio",
+        minimax_h3_mode="ref2va",
+    )
+    colon_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    colon_packet["minimax_h3_prompt"] = colon_subject_prompt
+    colon_result = _split(colon_packet, context=colon_context, target=ref_target)
+    assert colon_result[12] is True, _metadata(colon_result)
+    assert colon_result[13], _metadata(colon_result)
+
+    canonical_sections = dg._minimax_h3_ref_sections(colon_subject_prompt)
+    assert canonical_sections is not None
+    malformed_prompt = (
+        f"{canonical_sections['subject_definitions']}\n"
+        "summary[animation, duel]: The target video follows <Subject 1> across a rain-darkened rooftop while preserving the "
+        "identity and wardrobe defined by <Picture 1>.\n"
+        "retention_analysis:\n"
+        "[<Subject 1>: fully_preserved - preserve the angular face, short black hair, red coat, silver satchel, lean proportions, "
+        "and black boots.]\n"
+        f"detailed_description:\n{canonical_sections['detailed_description']}\n"
+        f"overall_soundscape:\n{canonical_sections['overall_soundscape']}\n"
+        f"non_diegetic_music:\n{canonical_sections['non_diegetic_music']}"
+    )
+    repaired_prompt = dg._repair_minimax_h3_ref2va_structure(
+        malformed_prompt,
+        user_prompt,
+        manifest,
+    )
+    assert repaired_prompt != malformed_prompt
+    assert repaired_prompt.startswith("subject_definitions:\n")
+    repaired_sections = dg._minimax_h3_ref_sections(repaired_prompt)
+    assert repaired_sections is not None, repaired_prompt
+    task_prefix = repaired_sections["summary"].split("]", 1)[0].removeprefix("[")
+    task_types = [item.strip() for item in task_prefix.split("+")]
+    assert task_types and all(task_type in dg._MINIMAX_H3_REF_TASK_TYPES for task_type in task_types), task_types
+    assert "<Subject 1>: fully_preserved - " in repaired_sections["retention_analysis"]
+    assert "[<Subject 1>" not in repaired_sections["retention_analysis"]
+    repaired_reasons = dg._minimax_h3_ref2va_validation_reasons(
+        repaired_prompt,
+        8.0,
+        user_prompt,
+        "auto_scene_audio",
+        8000,
+        manifest,
+    )
+    structural_reasons = {
+        "minimax_h3_ref_sections_invalid",
+        "minimax_h3_ref_field_order_invalid",
+        "minimax_h3_ref_field_spacing_invalid",
+        *(f"minimax_h3_ref_missing_{field_name}" for field_name in dg._MINIMAX_H3_REF_FIELDS),
+    }
+    assert structural_reasons.isdisjoint(repaired_reasons), repaired_reasons
+    assert "minimax_h3_ref_summary_invalid" not in repaired_reasons, repaired_reasons
+    assert "minimax_h3_ref_retention_invalid" not in repaired_reasons, repaired_reasons
+    assert dg._repair_minimax_h3_ref2va_structure(repaired_prompt, user_prompt, manifest) == repaired_prompt
+
+    canonical_repairable = colon_subject_prompt.replace(
+        "[reference generation] ",
+        "[keyframe completion + video editing] ",
+        1,
+    ).replace(
+        "[Shot 1] A static medium-wide",
+        "[Shot 1] At 00:00.000 A static medium-wide",
+        1,
+    ).replace(
+        "The camera eases to a static hold. The final frame holds ",
+        "The camera eases to a static hold. [Shot 2] At 00:06.000 The camera quick cuts to a locked wide view of the courier "
+        "beside the parapet. The final frame holds ",
+        1,
+    )
+    canonical_repaired = dg._repair_minimax_h3_ref2va_structure(canonical_repairable, user_prompt, manifest)
+    canonical_repaired_sections = dg._minimax_h3_ref_sections(canonical_repaired)
+    assert canonical_repaired_sections is not None, canonical_repaired
+    assert canonical_repaired_sections["summary"].startswith("[reference generation] ")
+    assert "[Shot 1] At " not in canonical_repaired_sections["detailed_description"]
+    assert "[Shot 2] At 00:06.000, the camera cuts to a locked wide view" in canonical_repaired_sections["detailed_description"]
+    assert "the camera cuts to a static wide view as The camera quick cuts" not in canonical_repaired_sections["detailed_description"]
+
+    valid_cut_malformed = malformed_prompt.replace(
+        "The camera eases to a static hold. The final frame holds ",
+        "The camera eases to a static hold. [Shot 2] At 00:06.000, the camera cuts to a locked wide view as the courier stops "
+        "beside the parapet. The final frame holds ",
+    )
+    valid_cut_repaired = dg._repair_minimax_h3_ref2va_structure(valid_cut_malformed, user_prompt, manifest)
+    assert "the camera cuts to a static wide view as the camera cuts to" not in valid_cut_repaired.lower()
+    assert valid_cut_repaired.count("[Shot 2] At 00:06.000, the camera cuts to") == 1, valid_cut_repaired
+
+    unlabeled_manifest = (
+        "<Picture 1>: protagonist identity and appearance; preserve exactly across every shot; this is a full-sequence reference, "
+        "not a first or last frame.\n"
+        "<Picture 2>: western town environment, palette, lighting, and 2D animation style; do not copy pictured subject identity; "
+        "this is a full-sequence reference, not a keyframe."
+    )
+    unlabeled_user_prompt = (
+        "Create a 20-second 2D hand-drawn animated monster duel in a late-1800s spaghetti-western town square. Use Picture 1 "
+        "for the protagonist and Picture 2 for the environment and style. Show clear action and reaction and a resolved final frame."
+    )
+    unlabeled_prefix_prompt = (
+        "<Subject 1>: A yellow trumpet-shaped monster with two oval eyes and three green leaves at its base, from <Picture 1>.\n"
+        "<Subject 2>: A late-1800s western town square with adobe buildings, cacti, and dusty ground, from <Picture 2>.\n\n"
+        "[keyframe completion + reference generation + video editing] <Subject 1> engages in a monster duel within <Subject 2> "
+        "using a 2D animated style.\n\n"
+        "<Subject 1>: fully_preserved - The yellow trumpet body, oval eyes, and green leaf base remain consistent.\n"
+        "<Subject 2>: fully_preserved - The adobe town, desert lighting, and western atmosphere remain constant.\n\n"
+        "The scene uses 2D hand-drawn cel animation with clean linework, soft cel shading, and stylized proportions; no live action, "
+        "photorealism, or 3D CGI. Bright warm light casts soft shadows across the dusty town floor.\n\n"
+        "[Shot 1] At 00:00.000 A low-angle wide shot reveals <Subject 1> in the center of <Subject 2>. The creature narrows its "
+        "eyes and its leaves twitch in the wind while the camera slowly pushes toward its face. Adobe storefronts and a cactus hold "
+        "steady behind it, establishing the opponent across the street and the distance between them.\n\n"
+        "[Shot 2] At 00:06.000 A quick cut to a close-up of the shadowy desert opponent. <Subject 1> leans forward, its body "
+        "vibrating with tension as dust swirls around its base. The locked camera shakes once with a heavy warning stomp while both "
+        "characters maintain their screen positions.\n\n"
+        "[Shot 3] At 00:12.000 The action explodes as <Subject 1> lunges forward and emits a blast of air. The shockwave crosses "
+        "the street and blows dust backward while the camera tracks beside <Subject 1>, keeping the opponent visible beyond the "
+        "advancing wave and carrying both positions into the next view.\n\n"
+        "[Shot 4] At 00:17.000 The dust settles and <Subject 1> stands victorious in the center of the frame. The camera pulls back "
+        "to a wide establishing view as the warm sunset silhouettes the mountains and holds the resolved final composition.\n\n"
+        "overall_soundscape: Wind through cacti, dry footfalls, a resonant air blast, and shifting gravel synchronize with the action.\n\n"
+        "non_diegetic_music: A tense western whistle builds into a brief orchestral flourish."
+    )
+    unlabeled_repaired = dg._repair_minimax_h3_ref2va_structure(
+        unlabeled_prefix_prompt,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+    )
+    unlabeled_sections = dg._minimax_h3_ref_sections(unlabeled_repaired)
+    assert unlabeled_sections is not None, unlabeled_repaired
+    assert unlabeled_sections["summary"].startswith("[reference generation] "), unlabeled_sections["summary"]
+    assert "keyframe completion" not in unlabeled_sections["summary"]
+    assert "video editing" not in unlabeled_sections["summary"]
+    assert "2D hand-drawn cel animation" in unlabeled_sections["detailed_description"]
+    assert "[Shot 1] At " not in unlabeled_sections["detailed_description"]
+    assert "[Shot 2] At 00:06.000, the camera cuts to a close-up" in unlabeled_sections["detailed_description"]
+    assert "[Shot 3] At 00:12.000, the camera cuts to a static wide view as the action explodes" in unlabeled_sections["detailed_description"]
+    unlabeled_reasons = dg._minimax_h3_ref2va_validation_reasons(
+        unlabeled_repaired,
+        20.0,
+        unlabeled_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    )
+    false_structural_reasons = {
+        "minimax_h3_ref_missing_detailed_description",
+        "minimax_h3_ref_subject_usage_invalid",
+        "minimax_h3_ref_summary_invalid",
+        "minimax_h3_ref_style_opening_missing",
+        "minimax_h3_cut_timestamp_invalid",
+    }
+    assert false_structural_reasons.isdisjoint(unlabeled_reasons), unlabeled_reasons
+    assert "minimax_h3_battle_result_missing" in unlabeled_reasons, unlabeled_reasons
+
+    reported_user_prompt = (
+        "Create a 20-second 2D hand-drawn cel-animated monster duel using Picture 1 as the protagonist reference and Picture 2 "
+        "as the western town environment and animation-style reference. End in a resolved post-duel state."
+    )
+    reported_no_colon_prompt = (
+        "<Subject 1>: A protagonist with pale skin, large yellow rabbit-like ears with white tufts, bright blue eyes, and a simple "
+        "smile, wearing a sleeveless pink dress (derived from <Picture 1>).\n"
+        "<Subject 2>: A shadowy, monstrous antagonist in a rugged western-style duster coat, introduced for the duel.\n"
+        "<Subject 3>: A late-1800s spaghetti-western town square with wooden storefronts, dusty ground, and a sun-drenched "
+        "atmosphere (derived from <Picture 2>).\n\n"
+        "[keyframe completion + reference generation] The protagonist <Subject 1> engages in a monster-duel within the "
+        "<Subject 3> environment using 2D animation style.\n\n"
+        "retention_analysis\n"
+        "<Subject 1>: fully_preserved - identity, face, ears, and pink dress are maintained throughout the sequence.\n"
+        "<Subject 2>: weak_reference - the antagonist is a new addition following the style of <Picture 2>.\n"
+        "<Subject 3>: fully_preserved - the town square environment and color palette remain consistent.\n\n"
+        "detail_description\n"
+        "2D hand-drawn cel animation, preserve drawn linework, cel shading, stylized proportions, and design continuity across every "
+        "cut, no live action, photorealism, or 3D CGI. The scene utilizes the high-contrast lighting of a western with deep shadows "
+        "and vibrant highlights.\n\n"
+        "[Shot 1] At 00:00.000, a low-angle shot captures <Subject 1> standing in the center of the <Subject 3> town square. The wind "
+        "blows through <Subject 1>'s large yellow ears. The camera slowly zooms in on <Subject 1>'s intense blue eyes as a shadow "
+        "falls across the face.\n\n"
+        "[Shot 2] At 00:05.000, a quick cut to a wide shot shows <Subject 2> standing at the end of the street. <Subject 2> draws a "
+        "spectral, glowing energy-weapon. The camera pans rapidly between <Subject 1> and <Subject 2> to build tension.\n\n"
+        "[Shot 3] At 00:10.000, the duel begins. <Subject 1> lunges forward with superhuman speed, the pink dress fluttering. "
+        "<Subject 2> fires a blast of dark energy. <Subject 1> dodges with a roll, sending up clouds of dust from the <Subject 3> "
+        "ground.\n\n"
+        "[Shot 4] At 00:15.000, <Subject 1> leaps into the air, delivering a powerful strike to <Subject 2>. The impact creates a "
+        "shockwave that shakes the nearby wooden buildings. The camera shakes violently with the force of the hit-stop.\n\n"
+        "[Shot 5] At 00:18.000, <Subject 2> dissipates into smoke. <Subject 1> lands gracefully in a defiant pose, breathing heavily, "
+        "adjusting the pink dress as the sun sets over the <Subject 3>. The camera holds a medium shot.\n\n"
+        "overall_soundscape\n"
+        "The whistling wind, the creak of wooden signs, heavy boots sliding on dirt, the hum of energy blasts, and a sharp thud upon "
+        "impact.\n\n"
+        "non_diegetic_music\n"
+        "/A"
+    )
+    reported_repaired = dg._repair_minimax_h3_ref2va_structure(
+        reported_no_colon_prompt,
+        reported_user_prompt,
+        unlabeled_manifest,
+    )
+    reported_sections = dg._minimax_h3_ref_sections(reported_repaired)
+    assert reported_sections is not None, reported_repaired
+    assert reported_repaired.startswith("subject_definitions:\n")
+    assert reported_sections["summary"].startswith("[reference generation] "), reported_sections["summary"]
+    assert "keyframe completion" not in reported_sections["summary"]
+    assert "2D hand-drawn cel animation" in reported_sections["detailed_description"]
+    assert "[Shot 1] At " not in reported_sections["detailed_description"]
+    assert "[Shot 2] At 00:05.000, the camera cuts to a wide shot" in reported_sections["detailed_description"]
+    assert "[Shot 3] At 00:10.000, the camera cuts to a static wide view as the duel begins" in reported_sections["detailed_description"]
+    reported_reasons = dg._minimax_h3_ref2va_validation_reasons(
+        reported_repaired,
+        20.0,
+        reported_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    )
+    assert reported_reasons == [], reported_reasons
+
+    subject_alias_sections = dict(reported_sections)
+    subject_alias_sections["detailed_description"] = subject_alias_sections["detailed_description"].replace(
+        "<Subject 3>",
+        "<Picture 2>",
+    )
+    subject_alias_prompt = dg._rebuild_minimax_h3_ref_prompt(subject_alias_sections)
+    assert "minimax_h3_ref_subject_usage_invalid" in dg._minimax_h3_ref2va_validation_reasons(
+        subject_alias_prompt,
+        20.0,
+        reported_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    )
+    subject_alias_repaired = dg._repair_minimax_h3_ref2va_structure(
+        subject_alias_prompt,
+        reported_user_prompt,
+        unlabeled_manifest,
+    )
+    subject_alias_repaired_sections = dg._minimax_h3_ref_sections(subject_alias_repaired)
+    assert subject_alias_repaired_sections is not None, subject_alias_repaired
+    assert "<Subject 3> (derived from <Picture 2>)" in subject_alias_repaired_sections["detailed_description"]
+    assert dg._minimax_h3_ref2va_validation_reasons(
+        subject_alias_repaired,
+        20.0,
+        reported_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    ) == []
+    assert dg._repair_minimax_h3_ref2va_structure(
+        subject_alias_repaired,
+        reported_user_prompt,
+        unlabeled_manifest,
+    ) == subject_alias_repaired
+
+    ambiguous_alias_sections = dict(subject_alias_sections)
+    ambiguous_alias_sections["subject_definitions"] = ambiguous_alias_sections["subject_definitions"].replace(
+        "(derived from <Picture 2>)",
+        "(derived from <Picture 1> and <Picture 2>)",
+        1,
+    )
+    ambiguous_alias_prompt = dg._rebuild_minimax_h3_ref_prompt(ambiguous_alias_sections)
+    ambiguous_alias_repaired = dg._repair_minimax_h3_ref2va_structure(
+        ambiguous_alias_prompt,
+        reported_user_prompt,
+        unlabeled_manifest,
+    )
+    ambiguous_alias_repaired_sections = dg._minimax_h3_ref_sections(ambiguous_alias_repaired)
+    assert ambiguous_alias_repaired_sections is not None, ambiguous_alias_repaired
+    assert "<Subject 3>" not in ambiguous_alias_repaired_sections["detailed_description"]
+    assert "minimax_h3_ref_subject_usage_invalid" in dg._minimax_h3_ref2va_validation_reasons(
+        ambiguous_alias_repaired,
+        20.0,
+        reported_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    )
+
+    reported_ambiguous = reported_no_colon_prompt.replace(
+        "[keyframe completion + reference generation] The protagonist",
+        "UNLABELED OUTPUT\n\n[keyframe completion + reference generation] The protagonist",
+        1,
+    )
+    reported_ambiguous_repaired = dg._repair_minimax_h3_ref2va_structure(
+        reported_ambiguous,
+        reported_user_prompt,
+        unlabeled_manifest,
+    )
+    assert dg._minimax_h3_ref_sections(reported_ambiguous_repaired) is None, reported_ambiguous_repaired
+
+    ambiguous_unlabeled = unlabeled_prefix_prompt.replace("<Subject 1>: fully_preserved - ", "<Subject 1>: preserved - ", 1)
+    ambiguous_repaired = dg._repair_minimax_h3_ref2va_structure(ambiguous_unlabeled, unlabeled_user_prompt, unlabeled_manifest)
+    assert dg._minimax_h3_ref_sections(ambiguous_repaired) is None, ambiguous_repaired
+
+    duplicate_suffix = unlabeled_prefix_prompt.replace(
+        "overall_soundscape:",
+        "overall_soundscape: First ambience.\n\noverall_soundscape:",
+        1,
+    )
+    duplicate_suffix_repaired = dg._repair_minimax_h3_ref2va_structure(
+        duplicate_suffix,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+    )
+    assert dg._minimax_h3_ref_sections(duplicate_suffix_repaired) is None, duplicate_suffix_repaired
+    assert "[Shot 4]" in duplicate_suffix_repaired
+
+    split_retention = unlabeled_prefix_prompt.replace(
+        "remain consistent.\n<Subject 2>: fully_preserved",
+        "remain consistent.\n\n<Subject 2>: fully_preserved",
+        1,
+    )
+    split_retention_repaired = dg._repair_minimax_h3_ref2va_structure(
+        split_retention,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+    )
+    split_retention_sections = dg._minimax_h3_ref_sections(split_retention_repaired)
+    assert split_retention_sections is not None, split_retention_repaired
+    assert "<Subject 1>: fully_preserved" in split_retention_sections["retention_analysis"]
+    assert "<Subject 2>: fully_preserved" in split_retention_sections["retention_analysis"]
+
+    split_invalid_retention = split_retention.replace(
+        "<Subject 2>: fully_preserved - ",
+        "<Subject 2>: preserved - ",
+        1,
+    )
+    split_invalid_repaired = dg._repair_minimax_h3_ref2va_structure(
+        split_invalid_retention,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+    )
+    assert dg._minimax_h3_ref_sections(split_invalid_repaired) is None, split_invalid_repaired
+
+    malformed_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    malformed_packet["minimax_h3_prompt"] = malformed_prompt
+    malformed_result = _split(malformed_packet, context=colon_context, target=ref_target)
+    assert malformed_result[12] is True, _metadata(malformed_result)
+    assert malformed_result[13], _metadata(malformed_result)
+    assert malformed_result[13] == repaired_prompt, malformed_result[13]
+    splitter_sections = dg._minimax_h3_ref_sections(malformed_result[13])
+    assert splitter_sections is not None, malformed_result[13]
+    assert splitter_sections["summary"].startswith("[reference generation] "), splitter_sections["summary"]
+    assert "<Subject 1>: fully_preserved - " in splitter_sections["retention_analysis"]
+    assert "[<Subject 1>" not in splitter_sections["retention_analysis"]
+
+    gate = dg.DiffusionGemmaGenerationGate()
+    ready_prompt = "A nonempty generation-ready prompt."
+    assert gate.gate(ready_prompt, True) == (ready_prompt,)
+    blocked_cases = (
+        (ready_prompt, False, json.dumps({"blocked_reasons": ["minimax_h3_ref_sections_invalid"]})),
+        ("   ", True, ""),
+    )
+    for prompt, ready_for_generation, metadata_json in blocked_cases:
+        try:
+            gate.gate(prompt, ready_for_generation, metadata_json)
+        except ValueError as exc:
+            if metadata_json:
+                assert "minimax_h3_ref_sections_invalid" in str(exc), str(exc)
+        else:
+            raise AssertionError((prompt, ready_for_generation, metadata_json))
+
+    valid_retry_prompt = unlabeled_repaired.replace(
+        "The shockwave crosses the street and blows dust backward",
+        "The shockwave strikes the opponent, who is knocked backward two steps as dust blows across the street",
+    )
+    valid_retry_reasons = dg._minimax_h3_ref2va_validation_reasons(
+        valid_retry_prompt,
+        20.0,
+        unlabeled_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    )
+    assert valid_retry_reasons == [], valid_retry_reasons
+    range_retry_prompt = (
+        valid_retry_prompt.replace("[Shot 1] ", "[Shot 1] 00:00-00:06: ", 1)
+        .replace(
+            "[Shot 2] At 00:06.000, the camera cuts to ",
+            "[Shot 2] 00:06-00:12: ",
+            1,
+        )
+        .replace(
+            "[Shot 3] At 00:12.000, the camera cuts to ",
+            "[Shot 3] 00:12-00:17: ",
+            1,
+        )
+        .replace(
+            "[Shot 4] At 00:17.000, the camera cuts to ",
+            "[Shot 4] 00:17-00:20: ",
+            1,
+        )
+    )
+    range_retry_repaired = dg._repair_minimax_h3_ref2va_structure(
+        range_retry_prompt,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+        20.0,
+    )
+    assert "[Shot 1] 00:00-00:06" not in range_retry_repaired
+    assert "[Shot 2] At 00:06.000, the camera cuts to" in range_retry_repaired
+    assert "[Shot 3] At 00:12.000, the camera cuts to" in range_retry_repaired
+    assert dg._minimax_h3_ref2va_validation_reasons(
+        range_retry_repaired,
+        20.0,
+        unlabeled_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    ) == []
+
+    seconds_range_retry_prompt = (
+        valid_retry_prompt.replace("[Shot 1] ", "[Shot 1] (00:00-06:00): ", 1)
+        .replace(
+            "[Shot 2] At 00:06.000, the camera cuts to ",
+            "[Shot 2] (06:00-12:00): ",
+            1,
+        )
+        .replace(
+            "[Shot 3] At 00:12.000, the camera cuts to ",
+            "[Shot 3] (12:00-17:00): ",
+            1,
+        )
+        .replace(
+            "[Shot 4] At 00:17.000, the camera cuts to ",
+            "[Shot 4] (17:00-20:00): ",
+            1,
+        )
+    )
+    seconds_range_retry_repaired = dg._repair_minimax_h3_ref2va_structure(
+        seconds_range_retry_prompt,
+        unlabeled_user_prompt,
+        unlabeled_manifest,
+        20.0,
+    )
+    assert "[Shot 1] (00:00-06:00)" not in seconds_range_retry_repaired
+    assert "[Shot 2] At 00:06.000, the camera cuts to" in seconds_range_retry_repaired
+    assert "[Shot 4] At 00:17.000, the camera cuts to" in seconds_range_retry_repaired
+    assert dg._minimax_h3_ref2va_validation_reasons(
+        seconds_range_retry_repaired,
+        20.0,
+        unlabeled_user_prompt,
+        "auto_scene_audio",
+        8000,
+        unlabeled_manifest,
+    ) == []
+    malformed_retry_prompt = (
+        f"subject_definitions:\n{unlabeled_sections['subject_definitions']}\n\n"
+        f"summary:\n{unlabeled_sections['summary']}\n\n"
+        f"retention_analysis:\n{unlabeled_sections['retention_analysis']}\n\n"
+        "detail_description:\nThe animation remains 2D hand-drawn.\n\n[Shot 1] A wide shot of\n\n"
+        "detailed_description:\n\n"
+        f"overall_soundscape:\n{unlabeled_sections['overall_soundscape']}\n\n"
+        f"non_diegetic_music:\n{unlabeled_sections['non_diegetic_music']}"
+    )
+
+    def refinement_response(prompt: str) -> str:
+        return (
+            f"{dg.FINAL_JSON_OPEN}"
+            + json.dumps(
+                {
+                    "ltx_prompt": "",
+                    "ideogram_prompt": "",
+                    "minimax_h3_prompt": prompt,
+                    "negative_prompt": "",
+                    "scene_segments": [],
+                    "metadata": {},
+                }
+            )
+            + f"{dg.FINAL_JSON_CLOSE}"
+        )
+
+    runtime = dg.RuntimeConfig(
+        model_path="mock",
+        backend="transformers_inprocess",
+        dtype="auto",
+        quantization="none",
+        local_files_only=True,
+        unload_policy="keep_loaded",
+        max_memory_gb=1.0,
+        status={"ready": True, "supports_pixels": False},
+    )
+    retry_context = _h3_ref_context(unlabeled_user_prompt, unlabeled_manifest)
+    retry_target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=20.0,
+        audio_mode="auto_scene_audio",
+        minimax_h3_mode="ref2va",
+    )
+
+    def run_refinement(outputs: list, generation_context=retry_context, generation_target=retry_target):
+        backend_prompts = []
+        original_run_backend = dg._run_backend
+
+        def fake_run_backend(_config, prompt, _media_context, _max_new_tokens, node_id=None):
+            backend_prompts.append(prompt)
+            output = outputs[len(backend_prompts) - 1]
+            if isinstance(output, Exception):
+                raise output
+            return output
+
+        dg._run_backend = fake_run_backend
+        try:
+            result = dg._run_generation_packet(
+                runtime,
+                generation_context,
+                generation_target,
+                runtime_required=True,
+                max_new_tokens=2048,
+                max_output_chars=8000,
+            )
+        finally:
+            dg._run_backend = original_run_backend
+        return result, backend_prompts
+
+    subject_alias_response = refinement_response(subject_alias_prompt)
+    subject_alias_context = _h3_ref_context(reported_user_prompt, unlabeled_manifest)
+    subject_alias_generation, subject_alias_backend_prompts = run_refinement(
+        [subject_alias_response],
+        subject_alias_context,
+        retry_target,
+    )
+    assert len(subject_alias_backend_prompts) == 1, len(subject_alias_backend_prompts)
+    assert subject_alias_generation[0]["minimax_h3_prompt"] == subject_alias_repaired
+    assert subject_alias_generation[6]["ready_for_generation"] is True, subject_alias_generation[6]
+    assert "minimax_h3_refinement" not in subject_alias_generation[6], subject_alias_generation[6]
+
+    subject_usage_refinement_prompt = dg._build_minimax_h3_refinement_prompt(
+        reported_user_prompt,
+        ambiguous_alias_prompt,
+        20.0,
+        "auto_scene_audio",
+        ["minimax_h3_ref_subject_usage_invalid"],
+        "ref2va",
+        unlabeled_manifest,
+    )
+    assert "Every defined <Subject N> must have its own retention_analysis line" in subject_usage_refinement_prompt
+    assert "Never substitute its source <Picture N> or <Video N> tag" in subject_usage_refinement_prompt
+    assert "use its <Picture N> or <Video N> source only as provenance" in subject_usage_refinement_prompt
+
+    timestamp_refinement_prompt = dg._build_minimax_h3_refinement_prompt(
+        unlabeled_user_prompt,
+        range_retry_prompt,
+        20.0,
+        "auto_scene_audio",
+        ["minimax_h3_cut_timestamp_invalid"],
+        "ref2va",
+        unlabeled_manifest,
+    )
+    assert "Shot 1 must begin exactly with '[Shot 1]' and no timestamp" in timestamp_refinement_prompt
+    assert "Never write start-end ranges" in timestamp_refinement_prompt
+    assert "'[Shot N] At MM:SS.mmm, the camera cuts to ...'" in timestamp_refinement_prompt
+
+    initial_response = refinement_response(unlabeled_repaired)
+    malformed_response = refinement_response(malformed_retry_prompt)
+    valid_response = refinement_response(valid_retry_prompt)
+    range_response = refinement_response(range_retry_prompt)
+    range_first_pass, range_first_pass_prompts = run_refinement([initial_response, range_response])
+    assert len(range_first_pass_prompts) == 2, len(range_first_pass_prompts)
+    assert range_first_pass[0]["minimax_h3_prompt"] == range_retry_repaired
+    assert range_first_pass[6]["ready_for_generation"] is True, range_first_pass[6]
+    range_first_pass_refinement = range_first_pass[6]["minimax_h3_refinement"]
+    assert range_first_pass_refinement["accepted_attempt"] == 1, range_first_pass_refinement
+
+    first_pass_success, first_pass_prompts = run_refinement([initial_response, valid_response])
+    assert len(first_pass_prompts) == 2, len(first_pass_prompts)
+    first_pass_refinement = first_pass_success[6]["minimax_h3_refinement"]
+    assert first_pass_refinement["accepted_attempt"] == 1, first_pass_refinement
+    assert "retry_attempted" not in first_pass_refinement, first_pass_refinement
+
+    retry_success, retry_success_prompts = run_refinement([initial_response, malformed_response, valid_response])
+    assert len(retry_success_prompts) == 3, len(retry_success_prompts)
+    assert "This is the one bounded retry" in retry_success_prompts[2]
+    assert "The shockwave crosses the street and blows dust backward" in retry_success_prompts[2]
+    assert "minimax_h3_battle_result_missing" in retry_success_prompts[2]
+    assert "minimax_h3_ref_detailed_description_short" in retry_success_prompts[2]
+    assert retry_success[0]["minimax_h3_prompt"] == valid_retry_prompt
+    assert retry_success[6]["ready_for_generation"] is True, retry_success[6]
+    success_refinement = retry_success[6]["minimax_h3_refinement"]
+    assert success_refinement["accepted"] is True, success_refinement
+    assert success_refinement["accepted_attempt"] == 2, success_refinement
+    assert success_refinement["retry_base"] == "initial", success_refinement
+
+    retry_failure, retry_failure_prompts = run_refinement([initial_response, malformed_response, malformed_response])
+    assert len(retry_failure_prompts) == 3, len(retry_failure_prompts)
+    assert retry_failure[0]["minimax_h3_prompt"] == unlabeled_repaired
+    assert retry_failure[6]["ready_for_generation"] is False, retry_failure[6]
+    assert "minimax_h3_battle_result_missing" in retry_failure[6]["blocked_reasons"], retry_failure[6]
+    failure_refinement = retry_failure[6]["minimax_h3_refinement"]
+    assert failure_refinement["accepted"] is False, failure_refinement
+    assert len(failure_refinement["attempts"]) == 2, failure_refinement
+
+    exception_failure, exception_prompts = run_refinement([initial_response, RuntimeError("mock refinement failure")])
+    assert len(exception_prompts) == 2, len(exception_prompts)
+    assert exception_failure[0]["minimax_h3_prompt"] == unlabeled_repaired
+    assert exception_failure[6]["ready_for_generation"] is False, exception_failure[6]
+    exception_refinement = exception_failure[6]["minimax_h3_refinement"]
+    assert "mock refinement failure" in exception_refinement["error"], exception_refinement
+    assert "backend_seconds" in exception_refinement["attempts"][0], exception_refinement
+
+
+def test_minimax_h3_reference_context_preserves_picture_batch_with_video():
+    import torch
+
+    class FakeVideo:
+        def __init__(self, images, frame_rate=3.0):
+            self.images = images
+            self.frame_rate = frame_rate
+
+        def get_duration(self):
+            return 1.0
+
+        def get_frame_count(self):
+            return int(self.images.shape[0])
+
+        def get_dimensions(self):
+            return (int(self.images.shape[2]), int(self.images.shape[1]))
+
+        def get_components(self):
+            return SimpleNamespace(images=self.images, frame_rate=self.frame_rate)
+
+    manifest = (
+        "<Picture 1>: courier identity, face, hair, wardrobe, and distinguishing visual traits.\n"
+        "<Picture 2>: rainy rooftop environment, dusk palette, wet materials, and lighting reference.\n"
+        "<Video 1>: running motion, lateral camera path, cut rhythm, and landing mechanics reference."
+    )
+    reference_images = torch.ones((2, 6, 8, 3))
+    reference_video = FakeVideo(torch.zeros((3, 2, 4, 3)))
+    context, _context_json, _preview = dg.DiffusionGemmaH3ReferenceContext().build(
+        "Create an eight-second rooftop chase.",
+        manifest,
+        sample_fps=3.0,
+        max_duration_seconds=10.0,
+        max_frames=10,
+        overlong_policy="trim",
+        reference_images=reference_images,
+        reference_video=reference_video,
+    )
+    assert tuple(context.images.shape) == (5, 2, 4, 3), context.images.shape
+    assert context.media_metadata["minimax_h3_reference_image_batch_count"] == 2
+    assert context.media_metadata["reference_image_count"] == 2
+    assert context.media_metadata["video_sampled_frame_count"] == 3
+    messages, _processor_kwargs = dg._transformers_messages_and_processor_kwargs(
+        "compile the Ref2VA prompt",
+        dg._media_context_from_gemma_context(context),
+    )
+    content = messages[0]["content"]
+    labels = " ".join(item.get("text", "") for item in content if item.get("type") == "text")
+    assert "<Picture 1> pixel evidence" in labels, labels
+    assert "<Picture 2> pixel evidence" in labels, labels
+    assert "<Video 1>" in labels, labels
+    assert sum(item.get("type") == "image" for item in content) == 5, content
+
+
+def test_minimax_h3_visual_medium_fidelity():
+    pokemon_brief = (
+        "Create a 10-second pokemon-inspired animated short with animal-like creatures having a battle in a town square. "
+        "The style should be reimagined into a late-1800's spaghetti-western"
+    )
+    assert dg._minimax_h3_requested_visual_medium(pokemon_brief) == "2d_animation"
+    assert dg._minimax_h3_requested_visual_medium("Pokémon-inspired animated short, no 3D CGI") == "2d_animation"
+    assert dg._minimax_h3_requested_visual_medium("Pokémon-inspired stylized 3D computer-animated short") == "3d_animation"
+    assert dg._minimax_h3_requested_visual_medium("Not a cartoon, use live action") == "live_action"
+    assert (
+        dg._minimax_h3_requested_visual_medium(
+            "Realistic live-action film photography with no animation, cartoon rendering, or CGI."
+        )
+        == "live_action"
+    )
+
+    model_prompt = dg._build_model_prompt(
+        pokemon_brief,
+        dg.DEFAULT_MASTER_PROMPT,
+        "minimax_h3",
+        {},
+        target_duration_seconds=10.0,
+        audio_mode="auto_scene_audio",
+        creativity_mode="faithful",
+        creative_strength=0.7,
+        thinking_mode="on",
+    )
+    lowered = model_prompt.lower()
+    assert '"minimax_h3_requested_visual_medium"' in lowered and '"2d_animation"' in lowered, model_prompt
+    assert "requested rendering medium is immutable 2d animation" in lowered
+    assert "2d hand-drawn cel animation" in lowered
+    assert "no live action, photorealism, or 3d cgi" in lowered
+    assert "minimax_h3_requested_style_locks" in lowered
+    assert "pokemon-inspired" in lowered
+    assert "late-1800's spaghetti-western" in lowered
+
+    target = dg._make_target_profile_config(
+        target_profile="minimax_h3",
+        target_duration_seconds=10.0,
+        audio_mode="auto_scene_audio",
+    )
+    context = dg.GemmaContext(user_prompt=pokemon_brief, media_metadata={})
+    failed_prompt = (
+        "integrated_multimodal_description: [Shot 1] A low-angle, extreme close-up of a reptilian-like creature with weathered "
+        "scales and a miniature leather duster, standing in a dusty sun-drenched 1800s town square. The creature narrows its eyes. "
+        "[Shot 2] At 00:03.500, the camera cuts to a wide shot of a large, avian-like creature with tattered feathers across the street. "
+        "The avian creature sends a shockwave across the ground. [Shot 3] At 00:07.200, the camera cuts to a medium shot as the "
+        "reptilian creature lunges forward and strikes the ground with a whip-pan.\n\n"
+        "overall_soundscape: Wind, gravel, a piercing screech, and electrical crackle surround the action.\n\n"
+        "non_diegetic_music: A whistle melody, Spanish guitar, and orchestral swells build through the battle."
+    )
+    failed_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    failed_packet["minimax_h3_prompt"] = failed_prompt
+    raw_failed_reasons = dg._minimax_h3_prompt_validation_reasons(failed_prompt, 10.0, pokemon_brief)
+    assert "minimax_h3_requested_visual_medium_missing" in raw_failed_reasons, raw_failed_reasons
+    assert "minimax_h3_requested_visual_exclusion_missing" in raw_failed_reasons, raw_failed_reasons
+    assert "minimax_h3_requested_style_missing" in raw_failed_reasons, raw_failed_reasons
+    failed_result = _split(failed_packet, context=context, target=target)
+    failed_metadata = _metadata(failed_result)
+    assert failed_result[13] == "", failed_result[13]
+    assert failed_result[12] is False, failed_metadata
+    assert "minimax_h3_battle_result_missing" in failed_metadata["blocked_reasons"], failed_metadata
+    assert "minimax_h3_requested_visual_medium_missing" not in failed_metadata["blocked_reasons"], failed_metadata
+    assert "minimax_h3_requested_visual_exclusion_missing" not in failed_metadata["blocked_reasons"], failed_metadata
+    assert failed_metadata["minimax_h3_requested_visual_medium"] == "2d_animation", failed_metadata
+    assert failed_metadata["minimax_h3_requested_style_locks"] == [
+        "pokemon-inspired",
+        "late-1800's spaghetti-western",
+    ], failed_metadata
+    assert failed_metadata["minimax_h3_style_locks_injected"] is True, failed_metadata
+    assert failed_metadata["minimax_h3_visual_medium_lock_injected"] is True, failed_metadata
+    assert failed_metadata["claim_verification_passed"] is False, failed_metadata
+
+    corrected_prompt = (
+        "integrated_multimodal_description: [Shot 1] 2D hand-drawn cel animation, Pokémon-inspired creature designs reimagined "
+        "through a late-1800s spaghetti-western aesthetic, with crisp inked outlines, flat cel shading, and no live action, "
+        "photorealism, or 3D CGI. A locked-off wide two-shot establishes Embercrest, a small red-orange lizard creature in a tan "
+        "duster, facing Skyspur, a tall cobalt bird creature in a black neckerchief, across the dusty town square. "
+        "[Shot 2] At 00:03.500, the camera cuts to a low-angle tracking shot beside Skyspur as it beats its wings and fires a blue "
+        "wind arc at Embercrest; the arc whistles, hits Embercrest's raised tail, and throws the same lizard two steps backward. "
+        "[Shot 3] At 00:07.200, the camera cuts to an overhead static hold as Embercrest plants its boots, redirects the fading arc into "
+        "the dirt, and Skyspur lowers its wings across the square. The final frame holds on both unchanged creature designs facing "
+        "each other through the settling dust in the overhead composition.\n\n"
+        "overall_soundscape: Dry wind crosses the square, leather creaks, boots scrape gravel, and dust settles after the impact.\n\n"
+        "non_diegetic_music: A steady Spanish-guitar rhythm and lone whistle pause on the wind-arc impact, then resolve on a low chord."
+    )
+    corrected_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    corrected_packet["minimax_h3_prompt"] = corrected_prompt
+    corrected_result = _split(corrected_packet, context=context, target=target)
+    assert corrected_result[13] == corrected_prompt, _metadata(corrected_result)
+    assert corrected_result[12] is True, _metadata(corrected_result)
+
+    missing_style_prompt = corrected_prompt.replace(
+        "Pokémon-inspired creature designs reimagined through a late-1800s spaghetti-western aesthetic, ",
+        "",
+    )
+    assert "minimax_h3_requested_style_missing" in dg._minimax_h3_prompt_validation_reasons(
+        missing_style_prompt,
+        10.0,
+        pokemon_brief,
+    )
+    repaired_style_prompt = dg._apply_minimax_h3_style_locks(missing_style_prompt, pokemon_brief)
+    assert "pokemon-inspired" in repaired_style_prompt.lower(), repaired_style_prompt
+    assert "late-1800's spaghetti-western" in repaired_style_prompt.lower(), repaired_style_prompt
+    assert "minimax_h3_requested_style_missing" not in dg._minimax_h3_prompt_validation_reasons(
+        repaired_style_prompt,
+        10.0,
+        pokemon_brief,
+    )
+    missing_style_packet = _packet("", {"ready_for_generation": True, "blocked_reasons": []}, segments=[])
+    missing_style_packet["minimax_h3_prompt"] = missing_style_prompt
+    repaired_style_result = _split(missing_style_packet, context=context, target=target)
+    assert repaired_style_result[13] == repaired_style_prompt, _metadata(repaired_style_result)
+    assert repaired_style_result[12] is True, _metadata(repaired_style_result)
+    assert _metadata(repaired_style_result)["minimax_h3_style_locks_injected"] is True
+
+    unresolved_battle_prompt = (
+        "integrated_multimodal_description: [Shot 1] 2D hand-drawn cel animation, pokemon-inspired and late-1800's "
+        "spaghetti-western visual styling, no live action, photorealism, or 3D CGI. A slow zoom-in frames a fox-like creature "
+        "opposite a reptilian creature in the town square. [Shot 2] At 00:03.500, the camera cuts to an extreme close-up of the "
+        "reptilian creature lunging forward with glowing claws. [Shot 3] At 00:07.000, the camera cuts to a wide profile-view shot "
+        "as the fox creature leaps and summons a golden shockwave. The final frame holds on the fox creature mid-air.\n\n"
+        "overall_soundscape: Wind, gravel, growls, and energy crackle surround the action.\n\n"
+        "non_diegetic_music: A whistle and rhythmic acoustic guitar build through the final shot."
+    )
+    unresolved_reasons = dg._minimax_h3_prompt_validation_reasons(
+        unresolved_battle_prompt,
+        10.0,
+        pokemon_brief,
+    )
+    assert "minimax_h3_battle_result_missing" in unresolved_reasons, unresolved_reasons
+    assert "minimax_h3_unresolved_final_state" in unresolved_reasons, unresolved_reasons
+    refinement_prompt = dg._build_minimax_h3_refinement_prompt(
+        pokemon_brief,
+        unresolved_battle_prompt,
+        10.0,
+        "auto_scene_audio",
+        unresolved_reasons,
+    ).lower()
+    assert "every attack must name attacker, target" in refinement_prompt
+    assert "do not end mid-air" in refinement_prompt
+    assert "the final frame holds on" in refinement_prompt
+
+    refined_artifact_prompt = (
+        "integrated_multimodal_description: [Shot 1] 2D hand-drawn cel animation, [pokemon-inspired\", \"late-1800's "
+        "spaghetti-western], no live action, photorealism, or 3D CGI. A slow low-angle zoom frames a fox-creature on the left "
+        "and a lizard-creature on the right. [Shot 2] At 00:03.500, the camera cuts to an extreme close-up as the lizard-creature "
+        "grips the dirt. [Shot 3] At 00:06.000, the camera cuts to a wide tracking shot as the fox-creature strikes the "
+        "lizard-creature; the lizard-creature is knocked backward into a hitching post. The final frame holds on the fox-creature "
+        "standing in the center and the dazed lizard-creature against the post on the right.\n\n"
+        "overall_soundscape: Wind, gravel, a heavy impact thud, and the crackle of a fire blast.\n\n"
+        "non_diegetic_music: Harmonica and acoustic guitar resolve on the final impact."
+    )
+    assert "minimax_h3_orphaned_sound_event" in dg._minimax_h3_prompt_validation_reasons(
+        refined_artifact_prompt,
+        10.0,
+        pokemon_brief,
+    )
+    canonical_style_prompt = dg._apply_minimax_h3_style_locks(refined_artifact_prompt, pokemon_brief)
+    assert "faithfully using pokemon-inspired and late-1800's spaghetti-western visual styling" in canonical_style_prompt
+    assert '[pokemon-inspired"' not in canonical_style_prompt
+    sound_repaired_prompt = dg._apply_minimax_h3_sound_fidelity(canonical_style_prompt, pokemon_brief)
+    assert "fire blast" not in sound_repaired_prompt.lower(), sound_repaired_prompt
+    assert dg._minimax_h3_prompt_validation_reasons(sound_repaired_prompt, 10.0, pokemon_brief) == []
+
+
+def test_minimax_h3_explicit_storyboard_fidelity():
+    brief = (
+        "Realistic live-action action trailer in four hard-cut shots. Shot 1 establishes the runner. Shot 2 is the leap. "
+        "Shot 3 is the landing. Shot 4 is a final freeze-frame holding on the next launch."
+    )
+    five_shot_prompt = (
+        "integrated_multimodal_description: [Shot 1] Realistic live-action practical film photography. A high-angle tracking shot "
+        "follows the runner. [Shot 2] At 00:02.000, the camera cuts to a wide profile as he leaps. "
+        "[Shot 3] At 00:04.000, the camera cuts to a low-angle handheld view as he lands. "
+        "[Shot 4] At 00:06.000, the camera cuts to an overhead static hold as he runs. "
+        "[Shot 5] At 00:08.000, the camera cuts to a long tracking shot as he reaches the next edge.\n\n"
+        "overall_soundscape: Wind and footsteps cross the roofs.\n\n"
+        "non_diegetic_music: Low percussion accelerates beneath the hard cuts."
+    )
+    reasons = dg._minimax_h3_prompt_validation_reasons(five_shot_prompt, 10.0, brief)
+    assert "minimax_h3_requested_shot_count_mismatch" in reasons, reasons
+    assert "minimax_h3_requested_final_hold_missing" in reasons, reasons
+
+    no_camera_prompt = five_shot_prompt.replace("A high-angle tracking shot follows the runner.", "The runner sprints toward the edge.")
+    assert "minimax_h3_shot_camera_unspecified" in dg._minimax_h3_prompt_validation_reasons(
+        no_camera_prompt,
+        10.0,
+        brief,
+    )
+
+
 def test_ltx_camera_choreography_contract():
     prompt = dg._build_model_prompt(
         "Recreate the source video as an LTX prompt.",
@@ -530,14 +1882,52 @@ def test_dg_to_ltx_prompt_injector():
         assert len(inventory) == 7
 
 
+def test_minimax_h3_ref2va_example_workflow_serialization():
+    workflow_path = ROOT / "examples" / "10_minimax_h3_ref2va_director.json"
+    workflow = json.loads(workflow_path.read_text(encoding="utf-8"))
+    links = {link[0]: link for link in workflow["links"]}
+
+    assert len(workflow["nodes"]) == 30
+    assert len(links) == 42
+    gate_node = next(node for node in workflow["nodes"] if node["id"] == 151)
+    assert gate_node["type"] == "DiffusionGemmaGenerationGate"
+    assert [input_slot["link"] for input_slot in gate_node["inputs"]] == [12, 40, 41]
+    assert gate_node["outputs"][0]["links"] == [42]
+    assert links[12][1:5] == [146, 13, 151, 0]
+    assert links[40][1:5] == [146, 12, 151, 1]
+    assert links[41][1:5] == [146, 4, 151, 2]
+    assert links[42][1:5] == [151, 0, 136, 9]
+    for node in workflow["nodes"]:
+        for output_index, output in enumerate(node.get("outputs", [])):
+            output_links = output.get("links")
+            assert output_links is None or isinstance(output_links, list), (node["id"], output["name"], output_links)
+            for link_id in output_links or []:
+                link = links[link_id]
+                assert link[1:3] == [node["id"], output_index], (node["id"], output["name"], link)
+        for input_index, input_slot in enumerate(node.get("inputs", [])):
+            link_id = input_slot.get("link")
+            if link_id is None:
+                continue
+            assert isinstance(link_id, int), (node["id"], input_slot["name"], link_id)
+            link = links[link_id]
+            assert link[3:5] == [node["id"], input_index], (node["id"], input_slot["name"], link)
+
+
 def main():
     test_readiness_gates()
     test_prompt_packet_schema()
+    test_minimax_h3_contract_and_splitter_compatibility()
+    test_minimax_h3_ref2va_contract()
+    test_minimax_h3_ref2va_structure_repair_and_generation_gate()
+    test_minimax_h3_reference_context_preserves_picture_batch_with_video()
+    test_minimax_h3_visual_medium_fidelity()
+    test_minimax_h3_explicit_storyboard_fidelity()
     test_ltx_camera_choreography_contract()
     test_image_identity_video_control_contract()
     test_media_sampler_forces_identity_image_in_synthesis_mode()
     test_transformers_media_message_labels_identity_and_control_frames()
     test_dg_to_ltx_prompt_injector()
+    test_minimax_h3_ref2va_example_workflow_serialization()
     print("Prompt Builder readiness/schema smoke passed")
 
 
diff --git a/tests/comfy_test_stubs.py b/tests/comfy_test_stubs.py
new file mode 100644
index 0000000..8952634
--- /dev/null
+++ b/tests/comfy_test_stubs.py
@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import ModuleType, SimpleNamespace
+import wave
+
+import torch
+import torch.nn.functional as torch_functional
+
+
+def install_comfy_stubs() -> None:
+    """Install the narrow ComfyUI seams exercised by the reframe unit tests."""
+
+    folder_paths = ModuleType("folder_paths")
+    folder_paths.get_input_directory = lambda: str(Path.cwd())
+    folder_paths.annotated_filepath = lambda name: (name, None)
+    folder_paths.filter_files_content_types = lambda files, _types: list(files)
+
+    def load_audio(path: str):
+        with wave.open(str(path), "rb") as handle:
+            if handle.getsampwidth() != 2:
+                raise ValueError("test audio stub supports 16-bit PCM only")
+            channels = handle.getnchannels()
+            frame_count = handle.getnframes()
+            sample_rate = handle.getframerate()
+            payload = bytearray(handle.readframes(frame_count))
+        samples = torch.frombuffer(payload, dtype=torch.int16).reshape(-1, channels)
+        waveform = samples.transpose(0, 1).to(torch.float32).div_(32768.0).contiguous()
+        return waveform, sample_rate
+
+    comfy_extras = ModuleType("comfy_extras")
+    comfy_extras.__path__ = []
+    comfy_audio = ModuleType("comfy_extras.nodes_audio")
+    comfy_audio.load = load_audio
+    comfy_extras.nodes_audio = comfy_audio
+
+    comfy = ModuleType("comfy")
+    comfy.__path__ = []
+    comfy_utils = ModuleType("comfy.utils")
+    comfy_utils.common_upscale = lambda images, width, height, _method, _crop: (
+        torch_functional.interpolate(
+            images,
+            size=(int(height), int(width)),
+            mode="bilinear",
+            align_corners=False,
+        )
+    )
+    comfy.utils = comfy_utils
+
+    class UnsupportedVideo:
+        def __init__(self, *_args, **_kwargs):
+            raise RuntimeError("video decoding is outside this unit-test seam")
+
+    comfy_latest = ModuleType("comfy_api.latest")
+    comfy_latest.InputImpl = SimpleNamespace(
+        VideoFromFile=UnsupportedVideo,
+        VideoFromComponents=UnsupportedVideo,
+    )
+    comfy_latest.Types = SimpleNamespace(
+        VideoComponents=lambda **values: SimpleNamespace(**values)
+    )
+    comfy_api = ModuleType("comfy_api")
+    comfy_api.__path__ = []
+    comfy_api.latest = comfy_latest
+
+    aiohttp_web = ModuleType("aiohttp.web")
+    aiohttp_web.json_response = lambda data, status=200: SimpleNamespace(
+        data=data,
+        status=status,
+    )
+    aiohttp = ModuleType("aiohttp")
+    aiohttp.__path__ = []
+    aiohttp.web = aiohttp_web
+
+    sys.modules.update(
+        {
+            "aiohttp": aiohttp,
+            "aiohttp.web": aiohttp_web,
+            "comfy": comfy,
+            "comfy.utils": comfy_utils,
+            "comfy_api": comfy_api,
+            "comfy_api.latest": comfy_latest,
+            "comfy_extras": comfy_extras,
+            "comfy_extras.nodes_audio": comfy_audio,
+            "folder_paths": folder_paths,
+        }
+    )
diff --git a/tests/test_ltx_reframe.py b/tests/test_ltx_reframe.py
new file mode 100644
index 0000000..68322a8
--- /dev/null
+++ b/tests/test_ltx_reframe.py
@@ -0,0 +1,365 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+import tempfile
+import unittest
+from unittest.mock import patch
+import wave
+
+import torch
+
+from comfy_test_stubs import install_comfy_stubs
+
+install_comfy_stubs()
+
+from ltx_reframe import (
+    LTXReframeCustomAudioLoader,
+    LTXReframeRectilinearCorrection,
+    LTXReframeTargetFit,
+    NODE_CLASS_MAPPINGS,
+    build_reframe_masks,
+    compute_reframe_placement,
+    ensure_audio_track,
+    fit_audio_track,
+    prepare_reframe_canvas,
+    rectilinear_correct_images,
+    resolve_frame_range,
+    snap_dimension,
+)
+
+
+class LTXReframeTests(unittest.TestCase):
+    @staticmethod
+    def _write_test_wav(path: Path, *, sample_rate: int = 8000, samples: int = 80):
+        with wave.open(str(path), "wb") as handle:
+            handle.setnchannels(1)
+            handle.setsampwidth(2)
+            handle.setframerate(sample_rate)
+            handle.writeframes(b"\x00\x00" * samples)
+
+    def test_custom_audio_loader_is_registered_and_accepts_an_empty_lazy_branch(self):
+        self.assertIs(
+            NODE_CLASS_MAPPINGS["LTXReframeCustomAudioLoader"],
+            LTXReframeCustomAudioLoader,
+        )
+        self.assertTrue(LTXReframeCustomAudioLoader.VALIDATE_INPUTS(""))
+        with self.assertRaisesRegex(ValueError, "no custom audio file is selected"):
+            LTXReframeCustomAudioLoader().load_audio("")
+
+    def test_custom_audio_loader_lists_audio_uploads_with_a_blank_default(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with (
+                patch("ltx_reframe.folder_paths.get_input_directory", return_value=directory),
+                patch(
+                    "ltx_reframe.folder_paths.filter_files_content_types",
+                    return_value=["song.wav"],
+                ),
+            ):
+                spec = LTXReframeCustomAudioLoader.INPUT_TYPES()
+
+        options, config = spec["required"]["audio_file"]
+        self.assertEqual(options, ["", "song.wav"])
+        self.assertTrue(config["audio_upload"])
+
+    def test_custom_audio_loader_decodes_and_fingerprints_an_input_file(self):
+        with tempfile.TemporaryDirectory() as directory:
+            audio_path = Path(directory) / "clip.wav"
+            self._write_test_wav(audio_path)
+            with (
+                patch("ltx_reframe.folder_paths.get_input_directory", return_value=directory),
+                patch(
+                    "ltx_reframe.folder_paths.annotated_filepath",
+                    side_effect=lambda name: (name, None),
+                ),
+            ):
+                audio, = LTXReframeCustomAudioLoader().load_audio("clip.wav")
+                first_fingerprint = LTXReframeCustomAudioLoader.IS_CHANGED("clip.wav")
+                second_fingerprint = LTXReframeCustomAudioLoader.IS_CHANGED("clip.wav")
+                with audio_path.open("ab") as handle:
+                    handle.write(b"changed")
+                changed_fingerprint = LTXReframeCustomAudioLoader.IS_CHANGED("clip.wav")
+
+        self.assertEqual(audio["sample_rate"], 8000)
+        self.assertEqual(tuple(audio["waveform"].shape), (1, 1, 80))
+        self.assertEqual(first_fingerprint, second_fingerprint)
+        self.assertNotEqual(first_fingerprint, changed_fingerprint)
+        self.assertEqual(len(first_fingerprint), 64)
+
+    def test_custom_audio_loader_reports_a_missing_input_file(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with (
+                patch("ltx_reframe.folder_paths.get_input_directory", return_value=directory),
+                patch(
+                    "ltx_reframe.folder_paths.annotated_filepath",
+                    side_effect=lambda name: (name, None),
+                ),
+            ):
+                with self.assertRaisesRegex(ValueError, "file was not found"):
+                    LTXReframeCustomAudioLoader().load_audio("missing.wav")
+
+    def test_custom_audio_loader_rejects_paths_outside_comfyui_input(self):
+        with tempfile.TemporaryDirectory() as directory:
+            with (
+                patch("ltx_reframe.folder_paths.get_input_directory", return_value=directory),
+                patch(
+                    "ltx_reframe.folder_paths.annotated_filepath",
+                    return_value=("../outside.wav", None),
+                ),
+            ):
+                with self.assertRaisesRegex(ValueError, "input directory"):
+                    LTXReframeCustomAudioLoader().load_audio("../outside.wav")
+
+    def test_custom_audio_example_workflow_has_consistent_links_and_safe_defaults(self):
+        workflow_path = (
+            Path(__file__).resolve().parents[1]
+            / "examples"
+            / "gemmaREFRAME_Experimental_audio_on-off.json"
+        )
+        workflow = json.loads(workflow_path.read_text(encoding="utf-8"))
+        nodes = {node["id"]: node for node in workflow["nodes"]}
+        links = {link[0]: link for link in workflow["links"]}
+
+        self.assertEqual(len(nodes), len(workflow["nodes"]))
+        self.assertEqual(len(links), len(workflow["links"]))
+        for link_id, (_, source_id, source_slot, target_id, target_slot, _type) in links.items():
+            self.assertIn(source_id, nodes)
+            self.assertIn(target_id, nodes)
+            self.assertIn(link_id, nodes[source_id]["outputs"][source_slot].get("links") or [])
+            self.assertEqual(nodes[target_id]["inputs"][target_slot].get("link"), link_id)
+
+        custom_loader = nodes[9086]
+        selector = nodes[9112]
+        reframe = nodes[5553]
+        stage_one = nodes[9055]
+        create_video = nodes[5539]
+        preview = nodes[9088]
+
+        self.assertEqual(custom_loader["type"], "LTXReframeCustomAudioLoader")
+        self.assertEqual(custom_loader["widgets_values"], [""])
+        self.assertEqual(selector["type"], "LazySwitchKJ")
+        self.assertFalse(selector["widgets_values"][0])
+        self.assertEqual(set(selector["outputs"][0]["links"]), {14666, 14818, 14918})
+        self.assertEqual(reframe["outputs"][4]["links"], [14990, 14991])
+        self.assertEqual(stage_one["inputs"][10]["link"], 14818)
+        self.assertEqual(preview["inputs"][0]["link"], 14918)
+        self.assertEqual(create_video["inputs"][1]["link"], 14999)
+        self.assertEqual(create_video["inputs"][2]["link"], 14666)
+
+    def test_existing_audio_is_preserved(self):
+        audio = {
+            "waveform": torch.ones((1, 1, 100), dtype=torch.float32),
+            "sample_rate": 16000,
+        }
+
+        result, synthesized = ensure_audio_track(audio, frame_count=24, fps=24.0)
+
+        self.assertIs(result, audio)
+        self.assertFalse(synthesized)
+
+    def test_audio_less_video_gets_duration_matched_stereo_silence(self):
+        audio, synthesized = ensure_audio_track(None, frame_count=121, fps=24.0)
+
+        self.assertTrue(synthesized)
+        self.assertEqual(audio["sample_rate"], 44100)
+        self.assertEqual(tuple(audio["waveform"].shape), (1, 2, 222338))
+        self.assertEqual(torch.count_nonzero(audio["waveform"]).item(), 0)
+
+    def test_silent_audio_fallback_rejects_invalid_fps(self):
+        with self.assertRaisesRegex(ValueError, "valid video frame rate"):
+            ensure_audio_track(None, frame_count=121, fps=0.0)
+
+    def test_audio_is_cropped_and_padded_to_selected_duration(self):
+        cropped, synthesized = fit_audio_track(
+            {"waveform": torch.ones((1, 2, 2000)), "sample_rate": 1000},
+            frame_count=25,
+            fps=25.0,
+        )
+        padded, _ = fit_audio_track(
+            {"waveform": torch.ones((1, 2, 500)), "sample_rate": 1000},
+            frame_count=25,
+            fps=25.0,
+        )
+
+        self.assertFalse(synthesized)
+        self.assertEqual(cropped["waveform"].shape[-1], 1000)
+        self.assertEqual(padded["waveform"].shape[-1], 1000)
+        self.assertEqual(torch.count_nonzero(padded["waveform"][..., 500:]).item(), 0)
+
+    def test_full_video_range_is_not_aligned_or_trimmed(self):
+        selected = resolve_frame_range(70, 25.0, False, 8, 16)
+
+        self.assertFalse(selected.range_enabled)
+        self.assertEqual(selected.effective_first_frame, 0)
+        self.assertEqual(selected.effective_last_frame, 69)
+        self.assertEqual(selected.selected_frame_count, 70)
+
+    def test_valid_sixty_five_frame_range_is_preserved(self):
+        selected = resolve_frame_range(241, 25.0, True, 0, 64)
+
+        self.assertEqual(selected.effective_first_frame, 0)
+        self.assertEqual(selected.effective_last_frame, 64)
+        self.assertEqual(selected.selected_frame_count, 65)
+
+    def test_range_alignment_preserves_first_frame_and_trims_end(self):
+        zero_based = resolve_frame_range(241, 25.0, True, 0, 70)
+        offset = resolve_frame_range(241, 25.0, True, 100, 199)
+
+        self.assertEqual((zero_based.effective_first_frame, zero_based.effective_last_frame), (0, 64))
+        self.assertEqual(zero_based.selected_frame_count, 65)
+        self.assertEqual((offset.effective_first_frame, offset.effective_last_frame), (100, 196))
+        self.assertEqual(offset.selected_frame_count, 97)
+
+    def test_last_frame_minus_one_resolves_to_source_end_then_aligns(self):
+        selected = resolve_frame_range(100, 25.0, True, 10, -1)
+
+        self.assertEqual(selected.requested_last_frame, 99)
+        self.assertEqual(selected.effective_last_frame, 98)
+        self.assertEqual(selected.selected_frame_count, 89)
+
+    def test_invalid_frame_ranges_are_rejected(self):
+        with self.assertRaisesRegex(ValueError, "greater than or equal"):
+            resolve_frame_range(100, 25.0, True, 30, 20)
+        with self.assertRaisesRegex(ValueError, "outside"):
+            resolve_frame_range(100, 25.0, True, 0, 100)
+        with self.assertRaisesRegex(ValueError, "at least nine"):
+            resolve_frame_range(100, 25.0, True, 0, 7)
+
+    def test_single_frame_range_is_allowed(self):
+        selected = resolve_frame_range(100, 25.0, True, 42, 42)
+
+        self.assertEqual(selected.selected_frame_count, 1)
+        self.assertEqual(selected.effective_last_frame, 42)
+
+    def test_dimensions_snap_to_ltx_multiple(self):
+        self.assertEqual(snap_dimension(1919), 1920)
+        self.assertEqual(snap_dimension(1050), 1024)
+        self.assertEqual(snap_dimension(1), 256)
+
+    def test_portrait_canvas_centers_aspect_locked_landscape_source(self):
+        placement = compute_reframe_placement(1920, 1080, 1088, 1920, 0.5, 0.5, 0.7)
+        self.assertEqual(placement.target_width, 1088)
+        self.assertEqual(placement.target_height, 1920)
+        self.assertLess(abs(placement.box_width / placement.box_height - 16 / 9), 0.01)
+        self.assertLessEqual(abs(placement.box_x * 2 + placement.box_width - placement.target_width), 1)
+        self.assertLessEqual(abs(placement.box_y * 2 + placement.box_height - placement.target_height), 1)
+
+    def test_canvas_and_masks_preserve_source_region(self):
+        frames = torch.ones((2, 8, 16, 3), dtype=torch.float32)
+        placement = compute_reframe_placement(16, 8, 256, 256, 0.5, 0.5, 0.5)
+        canvas = prepare_reframe_canvas(frames, placement)
+        outpaint, source = build_reframe_masks(placement, feather=0)
+
+        self.assertEqual(tuple(canvas.shape), (2, 256, 256, 3))
+        self.assertEqual(tuple(outpaint.shape), (1, 256, 256))
+        self.assertEqual(tuple(source.shape), (1, 256, 256))
+        self.assertTrue(torch.allclose(outpaint + source, torch.ones_like(source)))
+        self.assertEqual(canvas[:, placement.box_y, placement.box_x].mean().item(), 1.0)
+        self.assertEqual(canvas[:, 0, 0].sum().item(), 0.0)
+
+    def test_feather_ramps_only_outside_source_box(self):
+        placement = compute_reframe_placement(16, 9, 256, 256, 0.5, 0.5, 0.5)
+        outpaint, source = build_reframe_masks(placement, feather=16)
+        inside_y = placement.box_y + placement.box_height // 2
+        inside_x = placement.box_x + placement.box_width // 2
+        self.assertEqual(outpaint[0, inside_y, inside_x].item(), 0.0)
+        self.assertEqual(source[0, inside_y, inside_x].item(), 1.0)
+        self.assertEqual(outpaint[0, 0, 0].item(), 1.0)
+
+    def test_zoom_in_can_extend_source_beyond_target(self):
+        placement = compute_reframe_placement(1920, 1080, 1088, 1920, 0.5, 0.5, 2.0)
+
+        self.assertEqual(placement.box_width, 2176)
+        self.assertEqual(placement.box_height, 1224)
+        self.assertEqual(placement.box_x, -544)
+        self.assertEqual(placement.box_y, 348)
+        self.assertEqual(placement.source_scale, 2.0)
+
+    def test_zoomed_source_is_cropped_safely_to_target_canvas(self):
+        frames = torch.ones((2, 8, 16, 3), dtype=torch.float32)
+        placement = compute_reframe_placement(16, 8, 256, 256, 0.5, 0.5, 4.0)
+        canvas = prepare_reframe_canvas(frames, placement)
+        outpaint, source = build_reframe_masks(placement, feather=32)
+
+        self.assertEqual(tuple(canvas.shape), (2, 256, 256, 3))
+        self.assertTrue(torch.allclose(canvas, torch.ones_like(canvas)))
+        self.assertTrue(torch.allclose(source, torch.ones_like(source)))
+        self.assertTrue(torch.allclose(outpaint, torch.zeros_like(outpaint)))
+
+    def test_target_fit_returns_exact_dimensions(self):
+        images = torch.rand((2, 96, 160, 3), dtype=torch.float32)
+
+        fitted, = LTXReframeTargetFit().fit(images, target_width=128, target_height=128)
+
+        self.assertEqual(tuple(fitted.shape), (2, 128, 128, 3))
+        self.assertGreaterEqual(fitted.min().item(), 0.0)
+        self.assertLessEqual(fitted.max().item(), 1.0)
+
+    def test_rectilinear_correction_is_temporally_deterministic(self):
+        horizontal = torch.linspace(0.0, 1.0, 65).view(1, 1, 65, 1)
+        frame = horizontal.expand(1, 33, 65, 3)
+        images = frame.expand(2, -1, -1, -1).clone()
+
+        corrected, correction_mask = rectilinear_correct_images(
+            images,
+            correction_strength=0.1,
+            center_x=0.5,
+            center_y=0.5,
+            protect_source=0.0,
+            protection_feather=0,
+            interpolation="bilinear",
+            chunk_size=1,
+        )
+
+        self.assertEqual(tuple(corrected.shape), tuple(images.shape))
+        self.assertEqual(tuple(correction_mask.shape), (1, 33, 65))
+        self.assertTrue(torch.allclose(corrected[0], corrected[1]))
+        self.assertGreater(corrected[0, 16, 0, 0].item(), images[0, 16, 0, 0].item())
+        self.assertLess(corrected[0, 16, -1, 0].item(), images[0, 16, -1, 0].item())
+        self.assertAlmostEqual(
+            corrected[0, 16, 32, 0].item(),
+            images[0, 16, 32, 0].item(),
+            places=5,
+        )
+
+    def test_rectilinear_source_mask_can_fully_protect_original_pixels(self):
+        images = torch.rand((2, 24, 40, 3), dtype=torch.float32)
+        source_mask = torch.ones((1, 24, 40), dtype=torch.float32)
+
+        corrected, correction_mask = rectilinear_correct_images(
+            images,
+            correction_strength=0.15,
+            center_x=0.5,
+            center_y=0.5,
+            protect_source=1.0,
+            protection_feather=0,
+            interpolation="bicubic",
+            chunk_size=2,
+            source_mask=source_mask,
+        )
+
+        self.assertTrue(torch.allclose(corrected, images))
+        self.assertEqual(torch.count_nonzero(correction_mask).item(), 0)
+
+    def test_rectilinear_node_bypass_is_pixel_exact(self):
+        images = torch.rand((3, 16, 20, 3), dtype=torch.float32)
+
+        bypassed, correction_mask = LTXReframeRectilinearCorrection().correct(
+            images,
+            enabled=False,
+            correction_strength=0.05,
+            center_x=0.5,
+            center_y=0.5,
+            protect_source=1.0,
+            protection_feather=96,
+            interpolation="bicubic",
+            chunk_size=8,
+        )
+
+        self.assertIs(bypassed, images)
+        self.assertEqual(tuple(correction_mask.shape), (1, 16, 20))
+        self.assertEqual(torch.count_nonzero(correction_mask).item(), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_model_loader_ui.mjs b/tests/test_model_loader_ui.mjs
new file mode 100644
index 0000000..d4b1388
--- /dev/null
+++ b/tests/test_model_loader_ui.mjs
@@ -0,0 +1,89 @@
+import assert from "node:assert/strict";
+
+let registeredExtension = null;
+globalThis.window = {
+  comfyAPI: {
+    app: {
+      app: {
+        registerExtension(extension) {
+          registeredExtension = extension;
+        },
+      },
+    },
+  },
+};
+
+await import(`../web/js/model_loader.js?test=${Date.now()}`);
+assert.ok(registeredExtension, "the model-loader frontend extension should register");
+
+class UnrelatedNode {}
+await registeredExtension.beforeRegisterNodeDef(UnrelatedNode, { name: "OtherNode" });
+assert.equal(
+  UnrelatedNode.prototype.onNodeCreated,
+  undefined,
+  "unrelated node types should not be patched",
+);
+
+let createdCalls = 0;
+let configuredCalls = 0;
+
+class FakeLoaderNode {}
+FakeLoaderNode.prototype.onNodeCreated = function () {
+  createdCalls += 1;
+};
+FakeLoaderNode.prototype.onConfigure = function () {
+  configuredCalls += 1;
+};
+
+await registeredExtension.beforeRegisterNodeDef(FakeLoaderNode, {
+  name: "DiffusionGemmaModelLoader",
+});
+
+const hiddenWidgetNames = new Set([
+  "backend",
+  "dtype",
+  "quantization",
+  "local_files_only",
+  "unload_policy",
+  "max_memory_gb",
+]);
+const node = new FakeLoaderNode();
+node.widgets = [
+  { name: "model_path", value: "models/diffusiongemma" },
+  { name: "temperature", value: "invalid" },
+  ...[...hiddenWidgetNames].map((name) => ({ name, value: "advanced" })),
+];
+node.size = [520, 640];
+node.computeSize = () => [360, 240];
+node.setSize = (size) => {
+  node.size = [...size];
+};
+
+node.onNodeCreated();
+assert.equal(createdCalls, 1, "the original creation hook should still run");
+assert.equal(
+  node.widgets.find((entry) => entry.name === "temperature").value,
+  0.45,
+  "invalid saved temperatures should migrate to the stable default",
+);
+assert.deepEqual(node.size, [520, 240], "hiding widgets should preserve the user's width");
+
+for (const entry of node.widgets) {
+  if (hiddenWidgetNames.has(entry.name)) {
+    assert.equal(entry.hidden, true, `${entry.name} should be hidden`);
+    assert.equal(entry.__dgHidden, true, `${entry.name} should only be hidden once`);
+    assert.deepEqual(entry.computeSize(), [0, -4]);
+  } else {
+    assert.notEqual(entry.hidden, true, `${entry.name} should remain visible`);
+  }
+}
+
+node.widgets.find((entry) => entry.name === "temperature").value = 0.7;
+node.size = [610, 700];
+node.computeSize = () => [380, 260];
+node.onConfigure({});
+assert.equal(configuredCalls, 1, "the original configuration hook should still run");
+assert.equal(node.widgets.find((entry) => entry.name === "temperature").value, 0.7);
+assert.deepEqual(node.size, [610, 260], "configuration should keep the saved width");
+
+console.log("DiffusionGemma model-loader UI checks passed");
diff --git a/tests/test_package_metadata.py b/tests/test_package_metadata.py
new file mode 100644
index 0000000..4a891eb
--- /dev/null
+++ b/tests/test_package_metadata.py
@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+from pathlib import Path
+import tomllib
+import unittest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+class PackageMetadataTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.metadata = tomllib.loads(
+            (ROOT / "pyproject.toml").read_text(encoding="utf-8")
+        )
+
+    def test_registry_identity_and_release_version_are_current(self) -> None:
+        self.assertEqual(self.metadata["project"]["version"], "0.2.0")
+        self.assertEqual(self.metadata["tool"]["comfy"]["PublisherId"], "exportanything")
+
+    def test_registry_compatibility_starts_at_the_tested_baseline(self) -> None:
+        self.assertEqual(
+            self.metadata["tool"]["comfy"]["requires-comfyui"],
+            ">=0.30.1",
+        )
+
+    def test_registry_archive_excludes_ci_and_large_demo_media(self) -> None:
+        ignored = {
+            line.strip()
+            for line in (ROOT / ".comfyignore").read_text(encoding="utf-8").splitlines()
+            if line.strip() and not line.lstrip().startswith("#")
+        }
+        self.assertTrue({".github/", "tests/", "*.mp4", "*.gif"} <= ignored)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_splatstage_planner.py b/tests/test_splatstage_planner.py
new file mode 100644
index 0000000..6837d62
--- /dev/null
+++ b/tests/test_splatstage_planner.py
@@ -0,0 +1,437 @@
+from __future__ import annotations
+
+import importlib.util
+import json
+import sys
+import unittest
+import uuid
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+COMFY_ROOT = ROOT.parents[1]
+SCHEMA = ROOT / "schemas" / "splatstage_show_blueprint.schema.json"
+
+DIRECTOR_NODE_IDS = {
+    "DiffusionGemmaModelLoader",
+    "DiffusionGemmaContextHub",
+    "DiffusionGemmaH3ReferenceContext",
+    "DiffusionGemmaTargetProfile",
+    "DiffusionGemmaCoTGenerator",
+    "DiffusionGemmaJSONSplitter",
+    "DiffusionGemmaGenerationGate",
+}
+EXPERIMENTAL_NODE_IDS = {
+    "DiffusionGemmaSplatStagePlanner",
+    "DiffusionGemmaSplatStageEfficientPlanner",
+}
+
+
+def load_nodes_module():
+    if str(COMFY_ROOT) not in sys.path:
+        sys.path.insert(0, str(COMFY_ROOT))
+    package_name = f"diffusiongemma_splatstage_fixture_{uuid.uuid4().hex}"
+    spec = importlib.util.spec_from_file_location(
+        package_name,
+        ROOT / "__init__.py",
+        submodule_search_locations=[str(ROOT)],
+    )
+    if spec is None or spec.loader is None:
+        raise RuntimeError("could not load DiffusionGemma Prompt Builder")
+    package = importlib.util.module_from_spec(spec)
+    sys.modules[package_name] = package
+    spec.loader.exec_module(package)
+    return sys.modules[f"{package_name}.nodes"]
+
+
+class SplatStagePlannerMusicTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.nodes = load_nodes_module()
+        cls.schema = json.loads(SCHEMA.read_text(encoding="utf-8"))
+
+    def test_strict_schema_is_bundled_with_the_package(self) -> None:
+        self.assertEqual(self.nodes._splatstage_blueprint_schema_path(), SCHEMA)
+
+    def test_prompt_requests_a_full_song_before_the_video_excerpt(self) -> None:
+        prompt = self.nodes._splatstage_blueprint_prompt(
+            "make a pop-music video",
+            42,
+            self.schema,
+            duration_seconds=20.0,
+            aspect_ratio="16:9",
+        )
+        self.assertIn("complete 90-second source song", prompt)
+        self.assertIn("final video uses a 20-second excerpt", prompt)
+        self.assertIn("two separately generated performer clips", prompt)
+        self.assertIn("one performer object, one background object, and one FX object", prompt)
+        self.assertIn("never arrays or clip-specific lists", prompt)
+        self.assertIn("16 to 36 non-empty lines", prompt)
+        self.assertIn("SAM3 object-category bundle", prompt)
+        self.assertIn("wide-brimmed mariachi hat", prompt)
+        self.assertIn("electric guitar", prompt)
+        self.assertIn("positive_prompt must use affirmative visual language only", prompt)
+        self.assertIn('"schema":"splatstage.show_blueprint"', prompt)
+        self.assertIn('"version":1', prompt)
+        self.assertIn(
+            'Never put the schema document identifier "splatstage.show_blueprint@1"',
+            prompt,
+        )
+
+    def test_planner_contract_supports_ten_or_twenty_seconds(self) -> None:
+        spec = self.nodes.DiffusionGemmaSplatStagePlanner.INPUT_TYPES()["required"][
+            "duration_seconds"
+        ][1]
+        self.assertEqual(spec["min"], 10.0)
+        self.assertEqual(spec["max"], 20.0)
+        self.assertEqual(spec["default"], 20.0)
+
+    def test_full_song_lyric_budget_accepts_48_lines_but_not_49(self) -> None:
+        def blueprint(line_count: int) -> dict:
+            return {
+                "schema": "splatstage.show_blueprint",
+                "version": 1,
+                "project_title": "Song",
+                "creative_summary": "A compact performance video.",
+                "music": {
+                    "ace_tags": "bright pop",
+                    "vocal_mode": "vocal_hook",
+                    "lyrics": "\n".join(
+                        f"Original lyric line {index}" for index in range(line_count)
+                    ),
+                    "bpm": 120,
+                    "key": "C major",
+                    "time_signature": "4",
+                    "language": "en",
+                },
+                "performer": {
+                    "label": "Singer",
+                    "positive_prompt": "One singer performing",
+                    "negative_prompt": "crowd",
+                    "segmentation_prompt": "the singer",
+                },
+                "background": {
+                    "positive_prompt": "Abstract lights",
+                    "negative_prompt": "people",
+                },
+                "fx": {
+                    "positive_prompt": "Prism sparks",
+                    "negative_prompt": "people, text",
+                    "blend_mode": "screen",
+                    "key_mode": "luminance",
+                    "opacity": 0.5,
+                },
+                "editing": {
+                    "background_density": "musical",
+                    "fx_density": "musical",
+                    "music_reactivity": 0.5,
+                },
+                "finish": {"style_preset": "broadcast_clean"},
+            }
+
+        self.assertEqual(
+            self.nodes._splatstage_blueprint_errors(
+                blueprint(48), self.schema
+            ),
+            [],
+        )
+        self.assertTrue(
+            any(
+                "at most 48" in error
+                for error in self.nodes._splatstage_blueprint_errors(
+                    blueprint(49), self.schema
+                )
+            )
+        )
+
+    def test_segmentation_prompt_uses_separate_sam3_object_categories(self) -> None:
+        valid = {
+            "schema": "splatstage.show_blueprint",
+            "version": 1,
+            "project_title": "Mariachi",
+            "creative_summary": "A cyberpunk mariachi performance.",
+            "music": {
+                "ace_tags": "cyberpunk mariachi",
+                "vocal_mode": "instrumental",
+                "lyrics": "[Instrumental]",
+                "bpm": 128,
+                "key": "G minor",
+                "time_signature": "4",
+                "language": "en",
+            },
+            "performer": {
+                "label": "Lead mariachi",
+                "positive_prompt": "One cyberpunk mariachi playing a glowing guitar",
+                "negative_prompt": "crowd",
+                "segmentation_prompt": (
+                    "cyberpunk mariachi musician, black charro suit, "
+                    "wide-brimmed mariachi hat, glowing electric guitar"
+                ),
+            },
+            "background": {
+                "positive_prompt": "Martian landscape",
+                "negative_prompt": "people",
+            },
+            "fx": {
+                "positive_prompt": "isolated sparks",
+                "negative_prompt": "people, text",
+                "blend_mode": "screen",
+                "key_mode": "luminance",
+                "opacity": 0.5,
+            },
+            "editing": {
+                "background_density": "musical",
+                "fx_density": "musical",
+                "music_reactivity": 0.5,
+            },
+            "finish": {"style_preset": "broadcast_clean"},
+        }
+        self.assertEqual(
+            self.nodes._splatstage_blueprint_errors(valid, self.schema),
+            [],
+        )
+        leaked = json.loads(json.dumps(valid))
+        leaked["fx"]["positive_prompt"] = "Neon sparks, no words, text-free overlay"
+        polarity_errors = self.nodes._splatstage_blueprint_errors(leaked, self.schema)
+        self.assertTrue(any("affirmative visual language" in error for error in polarity_errors))
+        invalid = json.loads(json.dumps(valid))
+        invalid["performer"]["segmentation_prompt"] = (
+            "the musician in the charro suit and guitar"
+        )
+        self.assertTrue(
+            any(
+                "separate the performer" in error
+                for error in self.nodes._splatstage_blueprint_errors(
+                    invalid, self.schema
+                )
+            )
+        )
+
+    def test_two_option_lane_lists_are_seed_addressed_to_one_coherent_contract(self) -> None:
+        candidate = {
+            "performer": [{"label": "firefighter"}, {"label": "country singer"}],
+            "background": [{"name": "fire station"}, {"name": "burning skyline"}],
+            "fx": [{"name": "embers"}, {"name": "smoke"}],
+        }
+        normalized, report = self.nodes._normalize_splatstage_lane_cardinality(
+            candidate,
+            selection_index=5,
+        )
+        self.assertEqual(normalized["performer"]["label"], "country singer")
+        self.assertEqual(normalized["background"]["name"], "burning skyline")
+        self.assertEqual(normalized["fx"]["name"], "smoke")
+        self.assertEqual(report["performer"]["selected_index"], 1)
+        self.assertIsInstance(candidate["performer"], list)
+
+    def test_malformed_or_oversized_lane_lists_remain_invalid(self) -> None:
+        candidate = {
+            "performer": [{"label": "a"}, {"label": "b"}, {"label": "c"}],
+            "background": [],
+            "fx": ["not an object"],
+        }
+        normalized, report = self.nodes._normalize_splatstage_lane_cardinality(
+            candidate,
+            selection_index=0,
+        )
+        self.assertEqual(normalized, candidate)
+        self.assertEqual(report, {})
+
+    def test_schema_document_id_alias_is_canonicalized_only_for_matching_version(self) -> None:
+        candidate = {
+            "schema": "splatstage.show_blueprint@1",
+            "version": 1,
+            "project_title": "Alias candidate",
+        }
+        normalized, report = self.nodes._normalize_splatstage_blueprint_identity(
+            candidate,
+            self.schema,
+        )
+        self.assertEqual(normalized["schema"], "splatstage.show_blueprint")
+        self.assertEqual(report["input_schema"], "splatstage.show_blueprint@1")
+        self.assertEqual(candidate["schema"], "splatstage.show_blueprint@1")
+
+        wrong_version = dict(candidate, version=2)
+        unchanged, report = self.nodes._normalize_splatstage_blueprint_identity(
+            wrong_version,
+            self.schema,
+        )
+        self.assertEqual(unchanged, wrong_version)
+        self.assertEqual(report, {})
+
+        unknown_alias = dict(candidate, schema="splatstage.show_blueprint@999")
+        unchanged, report = self.nodes._normalize_splatstage_blueprint_identity(
+            unknown_alias,
+            self.schema,
+        )
+        self.assertEqual(unchanged, unknown_alias)
+        self.assertEqual(report, {})
+
+
+class EfficientSplatStagePlannerTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.nodes = load_nodes_module()
+        cls.schema = json.loads(
+            (ROOT / "schemas" / "splatstage_efficient_show_blueprint.schema.json").read_text(
+                encoding="utf-8"
+            )
+        )
+
+    @staticmethod
+    def valid_blueprint() -> dict:
+        return {
+            "schema": "splatstage.efficient_show_blueprint",
+            "version": 1,
+            "project_title": "Cyber Mariachi",
+            "creative_summary": "A cyberpunk mariachi performance on Mars.",
+            "music": {
+                "ace_tags": "cyber mariachi pop, vocal hook",
+                "vocal_mode": "vocal_hook",
+                "lyrics": "[Verse]\nRed dust in the light\n[Chorus]\nPlay it to the stars",
+                "bpm": 124,
+                "key": "C major",
+                "time_signature": "4",
+                "language": "en",
+            },
+            "performer": {
+                "identity_description": "One cyberpunk mariachi guitarist in a black charro suit",
+                "positive_prompt": "Full-body performance beneath magenta concert light",
+                "negative_prompt": "crowd, duplicate person, text, logos",
+                "segmentation_prompt": "mariachi guitarist, mariachi hat, charro suit, electric guitar",
+                "variations": [
+                    "Low-angle strumming with a slow dolly",
+                    "Energetic side view with rhythmic footwork",
+                    "Centered hero framing with sweeping arm motion",
+                ],
+            },
+            "background": {
+                "positive_prompt": "Martian neon concert architecture and red dust",
+                "negative_prompt": "people, performers, crowd, text, logos",
+                "variations": [
+                    "Wide canyon stage with drifting haze",
+                    "Graphic orbital city vista with slow parallax",
+                ],
+            },
+            "fx": {
+                "positive_prompt": "Isolated luminous prismatic guitar-wave geometry on a pure black background",
+                "negative_prompt": "people, faces, text, words, letters, logos, symbols",
+                "blend_mode": "screen",
+                "key_mode": "luminance",
+                "opacity": 0.55,
+            },
+            "editing": {"density": "musical", "music_reactivity": "medium"},
+            "finish": {"style_preset": "broadcast_clean"},
+        }
+
+    def test_schema_requires_exact_candidate_counts_and_affirmative_prompts(self) -> None:
+        value = self.valid_blueprint()
+        self.assertEqual(
+            self.nodes._splatstage_efficient_blueprint_errors(value, self.schema), []
+        )
+        value["performer"]["variations"] = value["performer"]["variations"][:2]
+        value["fx"]["positive_prompt"] = "Glowing particles, no words, text-free"
+        errors = self.nodes._splatstage_efficient_blueprint_errors(value, self.schema)
+        self.assertTrue(any("variations" in error for error in errors))
+        self.assertTrue(any("affirmative" in error for error in errors))
+
+    def test_prompt_locks_full_song_short_assets_and_semantic_union(self) -> None:
+        prompt = self.nodes._splatstage_efficient_blueprint_prompt(
+            "make a pop-music video", 42, self.schema
+        )
+        self.assertIn("complete 90-second song", prompt)
+        self.assertIn("three separate five-second performer videos", prompt)
+        self.assertIn("canonical performer identity first", prompt)
+        self.assertIn("only concrete silhouette-extending items actually present", prompt)
+        self.assertIn("richly art-directed luminous motif", prompt)
+        self.assertIn("limited color palette", prompt)
+        self.assertIn("rhythmic energy or", prompt)
+        self.assertIn("clear spatial composition", prompt)
+        self.assertIn("Do not name TouchDesigner", prompt)
+        self.assertIn("locked TouchDesigner-inspired", prompt)
+        self.assertIn("affirmative visual language", prompt)
+
+    def test_fx_creative_slot_rejects_software_or_interface_language(self) -> None:
+        value = self.valid_blueprint()
+        value["fx"]["positive_prompt"] = (
+            "TouchDesigner application screenshot with a cyan particle vortex"
+        )
+        errors = self.nodes._splatstage_efficient_blueprint_errors(value, self.schema)
+        self.assertTrue(
+            any("locked TouchDesigner-inspired" in error for error in errors)
+        )
+
+    def test_efficient_segmentation_rejects_free_standing_generic_categories(self) -> None:
+        value = self.valid_blueprint()
+        value["performer"]["segmentation_prompt"] = (
+            "mariachi guitarist, face and head, hair and worn headwear, "
+            "costume and worn accessories, held instrument and attached props"
+        )
+        errors = self.nodes._splatstage_efficient_blueprint_errors(value, self.schema)
+        self.assertTrue(any("identity-specific items" in error for error in errors))
+
+    def test_efficient_segmentation_uses_identity_and_specific_items_only(self) -> None:
+        value = self.valid_blueprint()
+        value["performer"]["segmentation_prompt"] = (
+            "cyberpunk mariachi guitarist, face and head, hair and worn headwear, "
+            "black charro suit, electric guitar"
+        )
+        normalized = self.nodes._normalize_splatstage_efficient_segmentation(value)
+        segmentation = normalized["performer"]["segmentation_prompt"]
+        self.assertTrue(
+            segmentation.startswith("One cyberpunk mariachi guitarist in a black charro suit")
+        )
+        self.assertNotIn("face and head", segmentation)
+        self.assertNotIn("hair and worn headwear", segmentation)
+        self.assertIn("black charro suit", segmentation)
+        self.assertIn("electric guitar", segmentation)
+        self.assertLessEqual(len(segmentation.split(",")), 6)
+        self.assertEqual(
+            self.nodes._splatstage_efficient_blueprint_errors(normalized, self.schema),
+            [],
+        )
+        self.assertEqual(
+            normalized["performer"]["positive_prompt"],
+            value["performer"]["positive_prompt"],
+        )
+
+    def test_efficient_planner_is_registered_with_safe_widget_contract(self) -> None:
+        self.assertIs(
+            self.nodes.NODE_CLASS_MAPPINGS["DiffusionGemmaSplatStageEfficientPlanner"],
+            self.nodes.DiffusionGemmaSplatStageEfficientPlanner,
+        )
+        required = self.nodes.DiffusionGemmaSplatStageEfficientPlanner.INPUT_TYPES()[
+            "required"
+        ]
+        self.assertEqual(required["max_new_tokens"][1]["default"], 3072)
+        self.assertGreaterEqual(required["max_new_tokens"][1]["min"], 1024)
+        self.assertNotIn("duration_seconds", required)
+
+
+class ExportedNodeSurfaceTests(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls) -> None:
+        cls.nodes = load_nodes_module()
+
+    def test_export_map_has_seven_director_and_two_experimental_nodes(self) -> None:
+        expected = DIRECTOR_NODE_IDS | EXPERIMENTAL_NODE_IDS
+        self.assertEqual(set(self.nodes.NODE_CLASS_MAPPINGS), expected)
+        self.assertEqual(set(self.nodes.NODE_DISPLAY_NAME_MAPPINGS), expected)
+
+    def test_experimental_planners_are_visibly_separate(self) -> None:
+        for node_id in DIRECTOR_NODE_IDS:
+            self.assertEqual(
+                self.nodes.NODE_CLASS_MAPPINGS[node_id].CATEGORY,
+                self.nodes.CATEGORY,
+            )
+        for node_id in EXPERIMENTAL_NODE_IDS:
+            self.assertEqual(
+                self.nodes.NODE_CLASS_MAPPINGS[node_id].CATEGORY,
+                self.nodes.EXPERIMENTAL_MOTION_CATEGORY,
+            )
+            self.assertTrue(
+                self.nodes.NODE_DISPLAY_NAME_MAPPINGS[node_id].endswith("(Experimental)")
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()