Clarify H3 prompt styles and full scene replacement instructions

H3_NATIVE_TEMPORAL.md

@@ -0,0 +1,192 @@
+# Experimental H3 native temporal alignment
+
+Matches ai-toolkit branch [`feat/h3-native-temporal-motion`](https://github.com/alisson-anjos/ai-toolkit/tree/feat/h3-native-temporal-motion), implementation `native_temporal_v1` (commit e0ecf3a). This mode retains a native video reference, aligns its temporal RoPE coordinates with the target, and softly penalizes target queries attending to distant guide times. It preserves native image-reference coordinates and native spatial geometry. It does not add learned weights or load a LoRA.
+
+Existing Add Guide motion-transfer LoRAs were trained under a different conditioning contract, but the step-10500 example below shows useful transfer to this native temporal mode. This is an experimental observation, not a guarantee for every LoRA. CPU math/routing tests alone do not validate video quality.
+
+## Video comparison: an Add Guide-trained LoRA with native temporal alignment
+
+[**Play the synchronized side-by-side video**](https://github.com/alisson-anjos/ComfyUI-BFSNodes/raw/refs/heads/feat/h3-native-temporal/docs/media/h3_native_temporal_comparison.mp4)
+
+All six columns play together in one 15-second video, starting at the beginning of each supplied export. From left to right:
+
+1. Original guide.
+2. Step-10500 motion-transfer LoRA + native temporal alignment, with guide scale **1.0**.
+3. The same step-10500 LoRA + native temporal alignment, with guide scale **0.5**.
+4. The same step-10500 LoRA + guide latents / Add Guide.
+5. The same step-10500 LoRA + native reference, without the temporal patch.
+6. Base model + native reference, without a LoRA or temporal patch.
+
+The temporal columns are labeled **NEW**. The footer includes the character and setting reference thumbnails extracted from the supplied exports. The user confirmed that all supplied runs used the **same seed**. At guide scale 0.5, character appearance remains consistent in the reviewed samples, but some source poses are reproduced less closely than at 1.0 (for example, the raised-arm gesture around 3.7 seconds). Other generation settings were not fully captured in the exports.
+
+**Apparently, a LoRA trained with guide latents can also work with this new native temporal method.** In this example, the step-10500 Add Guide-trained LoRA follows the source poses and camera framing with the temporal patch, while maintaining the reference dog's white T-shirt and jeans through the sampled beginning, middle and end. Its Add Guide result also follows the performance. Native reference without the temporal patch preserves the character more successfully with this LoRA than without it, but reproduces the choreography less closely. The base native-reference result loses the reference T-shirt.
+
+A plausible explanation is that the LoRA's learned appearance and motion-transfer behavior can be reused through shared model weights, while the runtime patch supplies temporal correspondence through native reference attention. This example does not isolate that mechanism experimentally or establish universal compatibility. The temporal mode still uses native video-reference latents; it does **not** inject Add Guide latents.
+
+**The reported identity drift concerns the separate, fresh native-temporal LoRA at step 750, not the step-10500 LoRA shown here.** The 750 checkpoint is not included in this comparison. Comparing those checkpoints would also compare different training histories and training durations.
+
+The requested background replacement is partial: source architecture and layout remain mixed with the new street reference. The seed is confirmed equal by the user, but these supplied exports do not include complete prompt and configuration metadata, so this is a qualitative demonstration rather than a controlled benchmark. Pixel difference and PSNR against the guide cannot rank replacement quality when character and background changes are intentional. The 30 fps comparison export does not establish the model's generation frame rate.
+
+## Standalone native workflow
+
+1. Load the native MiniMax H3 base and the native-temporal LoRA.
+2. Pass the guide as `<Video 1>` to **MiniMax H3 Reference to Video**, alongside any number of identity reference pictures supported by that native node. Picture and video numbering are independent.
+3. Place **BFS MiniMax H3 Native Temporal Alignment (Experimental)** on the MODEL path, after LoRA loading. Route its MODEL output to your guider/sampler. Keep Reference to Video's positive and AV latent outputs connected normally.
+4. Select `motion_video_index = 1` (or the actual `<Video N>` supplying motion).
+5. Use the same 24 fps frame window for guide and output. H3 may trim frame counts to its supported cadence; use a matching valid count such as 107.
+
+Do not add Add Guide, first-frame guides, VSA, SLA, or unknown attention/block replacements. Model Sparse Attention with sol-attn is experimentally supported from 1.115.14; see its limitations below. The official Model Attention Backend node is supported from 1.115.13; see the Kitchen section below. The node raises a clear error for incompatible layouts and temporal lengths instead of silently stretching motion. Source video audio is supported by native packing but is not temporally biased by this patch. For CFG paths, every conditioning branch evaluated by the patched model must retain the native motion reference; an empty negative branch without references cannot provide its timeline. Training preview uses CFG 1.
+
+## Planner
+
+In **BFS Shot H3 Conditioning** (Advanced or Simple), choose:
+
+```text
+guide_mode = native temporal alignment (experimental)
+```
+
+Connect MODEL from your LoRA loader to the conditioning node, then connect the conditioning node's MODEL output to the guider. Advanced keeps its existing output order. Simple appends MODEL after positive/latent/prompt; old links keep their indices. Reload older nodes or recreate them if the new socket does not appear.
+
+The shot is encoded as native `<Video 1>`, with identity pictures unchanged. The patch is applied separately to the returned model for each shot, and the current shot's packed temporal grid is used at sampling. Use 24 fps, first_frame none, mask_guide off and continuity off. Full-frame or ordinary native generation masks may still be used; masks do not create a temporal guide. Colored source overlays are retained if enabled, but color/mesh datasets have not yet been validated with this fresh experiment.
+
+The Planner option uses the training defaults. To tune the policy manually, instead choose the ordinary `native reference video (Video 1)` conditioning mode and place the standalone node on the model path. Never install both patches on the same model path.
+
+## Exact policy
+
+Only target-video query to selected-motion-video key edges receive:
+
+```text
+score -= min(strength * (max(abs(t_target - t_guide) - window_seconds, 0) / falloff_seconds)^2, max_penalty)
+```
+
+Defaults: strength 2.0, window 0.20 s, falloff 0.25 s, maximum penalty 6.0. Set strength 0 to test temporal RoPE alignment alone. Bypass the node to test fully native baseline. Text, image references, target self-attention, reference audio, and token-refiner attention retain their normal behavior.
+
+As of 1.115.11 the default backend is **flash partitions (no compile)**: it uses native CUDA FlashAttention with log-sum-exp outputs, positive partition merging and bounded frame chunks. No dense S×S bias tensor or torch.compile is used in this backend. CUDA FP16/BF16 and PyTorch's native FlashAttention-with-LSE primitive are required. The original compiled FlexAttention remains selectable for comparison. Start with model compilation/malloc graph optimizations disabled until validated on your local ComfyUI build.
+
+Runtime confirmation:
+
+```text
+[BFS H3 native temporal] native_temporal_v1: <Video 1>, ... motion rows; temporal RoPE aligned; soft window=0.2s strength=2.0 cap=6.0; images remain native; Add Guide OFF
+```
+
+This log confirms the patch ran, not that visual motion or identity quality is correct.
+
+## Validation
+
+Tests compare an audited native ComfyUI layout with positions exported by the actual training implementation, including image references and reference audio. They verify original conditioning/layout immutability, temporal-length rejection, attention score math/output shapes, token-refiner fallback, per-model routing and incompatible patches. Planner tests verify native-video conditioning and returned MODEL routing. The user has since reported completed RTX 5090 model evaluations; the timings and limitations are documented in the illustrated explanation. Full visual-quality validation remains incomplete. The active training process is not restarted or modified by this source change.
+
+
+## Performance changes in 1.115.11
+
+See the [illustrated Flash Partitions explanation](docs/H3_FLASH_PARTITIONS.md) for attention scope, derivation, log-sum-exp merging, measurements and limitations.
+
+The temporal rule and RoPE policy are unchanged. For motion keys, the modified softmax weight is decomposed exactly as `exp(-cap) + (exp(-penalty) - exp(-cap))`. The first term is ordinary attention over all motion keys. The positive correction is zero for distant times and can be computed with short per-frame batches. Non-motion keys keep weight one. All partitions are recombined with their log-sum-exp normalizers. This is not a hard temporal window or sparse approximation. BF16 partition merging has normal floating-point rounding differences; the small CUDA comparison against SDPA measured relative output error 0.00452.
+
+A synthetic long-sequence benchmark on RTX PRO 6000 Blackwell / PyTorch 2.12.0+cu130, 107126 tokens, 8 heads, head dimension 128 measured approximately 0.628 s for original FlexAttention, 0.422 s for Flash partitions and 0.327 s for unmodified native FlashAttention. Training was running concurrently, so these are indicative isolated-kernel timings, not whole-model inference results or RTX 5090 speed guarantees. Small sequences have additional partition-launch overhead and can favor FlexAttention. No full-model GPU generation is claimed here.
+
+Standalone node: `attention_backend` selects the implementation. Planner conditioning: `native_temporal_backend` selects it. Neither setting loads a LoRA or changes its weights. Defaults select Flash partitions; choose `flex (original)` for an A/B comparison.
+
+Planner `native_guide_scale` defaults to 1.0 for compatibility. Try 0.5 on VRAM-constrained GPUs: it reduces only the motion-reference spatial dimensions (about four times fewer guide tokens), keeping all timestamps/frames, output dimensions and identity images unchanged. This changes the visual detail supplied by the guide and must be evaluated for quality. In standalone workflows resize only the motion video before Reference to Video to get the same effect.
+
+The console logs the first attention's input shape and completion time and the first two full model evaluation times. If attention is fast but a full model evaluation remains slow, investigate model loading/offload and the remaining blocks rather than assuming compilation. These timings synchronize CUDA at those diagnostic points only. The active training code and its backend are unchanged.
+
+## One-pass native Kitchen kernels (1.115.16)
+
+**From 1.115.17, BFS prepares the one-pass extension automatically on first use of Kitchen/sol-attn on supported Linux x86_64 SM120 systems with a driver supporting CUDA 13.** It downloads a fixed release into ComfyUI’s user cache, verifies embedded SHA256 hashes and ABI/architecture, and activates it in that execution. No pip install, package replacement or sampling-time compilation is required. Other systems retain the compatibility backends or can use an explicitly built manual overlay.
+
+For an explicit installation or source build, use the [Kitchen fork's setup guide](https://github.com/alisson-anjos/comfy-kitchen/blob/feat/h3-native-temporal/H3_NATIVE_TEMPORAL_KERNELS.md). An [experimental Linux x86_64 SM120 prebuilt](https://github.com/alisson-anjos/comfy-kitchen/releases/tag/h3-native-temporal-kernel-v1) is available, including SHA256SUMS, so sampling need not compile anything. RTX 5090 shares SM120 with the tested RTX PRO 6000, but full video behavior/performance on the 5090 still requires validation. The installer adds a separate temporal module/library while preserving ordinary Kitchen's `_C` and operators. Source builds use `nvcc` explicitly; no per-shape torch.compile is used.
+
+Keep the existing MODEL wiring and native temporal guide mode. Model Attention Backend's Kitchen choice selects native dense when available. Model Sparse Attention's sol-attn choice selects native Sol during its eligible sparse window; its dense fallback retains the temporal policy too. No new UI switch is required. Console confirmation:
+
+```text
+Kitchen ONE-PASS dense kernel ACTIVE
+Kitchen ONE-PASS sol kernel ACTIVE
+```
+
+Both kernels evaluate the same query-dependent bounded temporal penalty inside their native score/softmax work. Dense INT8 quantizes once and runs one attention launch, using exact constant-tile shortcuts plus per-score handling at mixed boundaries. Sol preprocesses/routes/consumes once with the penalty fused into its exact consumer. All conditioning keys remain exact for Sol, and its remaining video keys retain the usual sparse/pooled approximation. Sol still has its normal multi-kernel stages: "one pass" means one attention pipeline, not literally one CUDA launch total. The streaming QKV producer remains bypassed by the current BFS integration.
+
+Synthetic isolated BF16 attention `[1,56,42836,128]`, 22 latent frames, 7,062 guide rows; RTX PRO 6000 Blackwell, PyTorch 2.12.0+cu130, Kitchen 0.2.35, concurrent training, one warmup plus one timed evaluation:
+
+| Backend | Seconds |
+| --- | ---: |
+| Flash Partitions | 0.4232 |
+| Kitchen dense, earlier frame-wise path | 0.5413 |
+| **Kitchen dense, one-pass kernel** | **0.2463** |
+| sol-attn, earlier frame-wise path | 2.6752 |
+| **sol-attn, one-pass kernel** | **0.2163** |
+
+These are not full-model/video or RTX 5090 timings. Full video speed can still be limited by offloading and other work. INT8 and Sol are approximate, so these measurements do not establish visual quality or dense-equivalent output.
+
+GPU checks covered FP16/BF16, partial frames/tiles, nonuniform timestamps, zero strength/cap, the real BFS dense/Sol entry points and different policies on concurrent streams. Dense relative output norm error versus FP32 was approximately 1.2–1.6% in small cases; Sol with every key exact approximately 1.1–1.4%. Native one-pass Sol versus the same native pipeline repeated per frame differed by at most about 0.0022% in those initial cases. Each BFS bridge made one native API call. [Recorded diagnostics](docs/benchmarks/h3_onepass_temporal_kernels.json) and the Kitchen fork's source document the scope. Full ComfyUI generated-video validation remains pending.
+
+The sections below retain the earlier compatibility implementations and measurements for comparison. The active training process/backend is unchanged.
+
+## Automatic first-use setup (1.115.17)
+
+Choose Kitchen or sol-attn normally and run the native temporal node. On supported Linux x86_64 / NVIDIA SM120 / CUDA-13-driver systems, it prepares the extension during node execution when the model's load device is known, before sampling. A lazy fallback can prepare it on the first attention call too. The ordinary Flash/Flex path does not request an installation.
+
+The default cache is `ComfyUI user directory/bfs/h3_temporal_kernels/`; standalone use falls back to `~/.cache/bfsnodes/h3_temporal_kernels/`. Only the pinned Python API (~10 KB), SM120 library (~3.3 MB), LICENSE and NOTICE are downloaded. The implementation writes to its own cache, not `site-packages`, and leaves ordinary Kitchen's `_C` untouched. Existing compatible manual overlays are preferred.
+
+Security/correctness properties: immutable source commit and embedded hashes; bounded download sizes; temporary downloads replaced atomically only after SHA256 validation; native ABI and GPU architecture checks before Python entry-point execution; thread/process installation locks; one setup attempt per device/cache location per process. An incomplete, incompatible or offline first installation logs its reason and retains the previous backend. A verified cache works in a new process without network access. No model, prompt or reference data is uploaded.
+
+Expected console messages:
+
+```text
+First-use setup: downloading the verified one-pass extension (~3.3 MB). No compilation or pip install.
+Verified one-pass extension ready: ...
+Kitchen ONE-PASS dense kernel ACTIVE
+# or: Kitchen ONE-PASS sol kernel ACTIVE
+```
+
+The addon activates without a second restart during that first workflow. Restarting ComfyUI is still required after updating BFS's own Python code. `BFS_H3_TEMPORAL_AUTO_INSTALL=0` disables automatic setup; `BFS_H3_TEMPORAL_CACHE_DIR=/path/to/cache` chooses another cache location. For a failed setup, restart to retry or use the manual installer linked above. Unsupported systems do not trigger an automatic source build.
+
+A real first-use download, ABI check, and dense/Sol GPU calls passed. A separate fresh process with network requests deliberately disabled reused the verified cache successfully. [Recorded first-use/offline checks](docs/benchmarks/h3_temporal_auto_setup.json).
+
+## Comfy Kitchen and Model Attention Backend (1.115.13)
+
+The official **Model Attention Backend** node is now recognized automatically. Select **comfy kitchen attention** there and route its MODEL through native temporal conditioning to the guider. Its official PyTorch selection is also accepted. Unrelated attention overrides remain incompatible. Model Sparse Attention with sol-attn is supported separately from 1.115.14, as documented below. The runtime MODEL wrapper recognizes a supported dense selection applied before or after the temporal patch; keep only one temporal patch on the path.
+
+With Kitchen selected, the console reports `backend=comfy kitchen (experimental)`. You can also explicitly select this backend in the standalone node or Planner conditioning. Without an upstream Kitchen selection, the existing default remains Flash Partitions. If ComfyUI itself falls back from unavailable Kitchen to PyTorch, the temporal patch does not claim Kitchen was selected. A supported selection from Model Attention Backend takes precedence over the BFS backend selector.
+
+This backend calls `comfy_kitchen.int8_attention(attn_mask=...)` directly. Each target latent frame's spatial queries share one broadcast key bias. Non-target queries use ordinary Kitchen attention, while target-to-guide scores receive the same bounded temporal penalty as the original rule. Native temporal RoPE alignment remains active. It uses no Add Guide injection, no partition log-sum-exp, and no full sequence-by-sequence bias allocation: bias storage is `latent_frames * packed_rows`. It retains full-key attention, not sparse attention. Multiple frame calls can repeat Kitchen quantization work, so compatibility does not imply acceleration.
+
+Initial isolated tests on an RTX PRO 6000 Blackwell with PyTorch 2.12.0+cu130 and comfy-kitchen 0.2.35:
+
+- CPU tests compare the frame-wise policy against the dense mathematical oracle, including zero strength/cap and unchanged non-target queries. Routing tests cover automatic detection of the official dense node.
+- A real CUDA BF16 test at `[1, 2, 1248, 128]` measured relative output norm error **0.01506** against the FP32 oracle, with finite outputs. INT8 quantization is approximate; this is not a visual quality measurement.
+- At `[1, 2, 42836, 128]`, 22 latent frames and 7,062 guide rows, one warmed evaluation measured **0.1043 s Kitchen** versus **0.01739 s Flash Partitions**. This initial Kitchen path was slower. The test used two heads rather than H3's full 56, with concurrent training; these are diagnostics, not end-to-end or RTX 5090 timings. No speed claim or default-backend change follows from this result.
+
+The active training backend is unchanged. Full ComfyUI video output and local performance still require validation. Version 1.115.14 adds a separate experimental sol-attn compatibility path below; SLA and VSA remain unsupported.
+
+## Model Sparse Attention: sol-attn compatibility (1.115.14)
+
+Use **1.115.15 or newer**: 1.115.15 fixes recognition when ComfyUI loads `nodes_sparse_attention.py` under its extensionless absolute file path instead of a dotted module name. The regression test uses extracted official node factories under that loader name.
+
+**This is an experimental compatibility implementation, not a speed optimization.** It automatically recognizes the official **Model Sparse Attention** node with `method = sol-attn`, keeps native temporal RoPE alignment and applies the temporal score penalty through Comfy Kitchen's `sol_attn(key_bias=...)`. The runtime banner identifies `backend=sol-attn frame bias (experimental)`; an `sol-attn ACTIVE` line confirms a sparse call actually ran. Dense windows can use the previously selected dense temporal backend, and the first attention log identifies that actual backend.
+
+Suggested MODEL wiring:
+
+```text
+LoRA loader -> Model Attention Backend (optional dense fallback)
+            -> Model Sparse Attention (sol-attn)
+            -> BFS native temporal node / Shot H3 Conditioning -> guider
+```
+
+The official dense and sparse selections are detected, including owned H3 hooks when a later dense node hides the sparse override. Unrecognized replacements still raise an error. `tau`, `extra_tokens`, `start_percent`, `end_percent`, `dense_blocks` and `min_tokens` retain their sparse-node meanings. Outside eligible sparse calls, the bounded temporal rule still runs through the dense temporal backend. SLA and VSA are rejected.
+
+ComfyUI's usual H3 sparse producer hooks would bypass BFS's temporal score modification. BFS therefore bypasses **only the hooks owned by that exact official sol-attn patch**, in a per-evaluation options copy, and performs attention through its temporal override. Other model clones and unrelated block hooks are not removed. This compatibility path does not retain the native sparse producer's streamed QKV memory optimization.
+
+The public sol-attn kernel currently accepts equal-length Q/K/V and a **key-only** bias; it does not accept the query-dependent temporal bias in one pass. BFS runs one unbiased full-sequence pass for non-target queries, then one full-sequence pass per target latent frame with that frame's key bias, retaining only its target query rows. All conditioning keys are forced into sol-attn's exact branch, since its pooled branch does not honor `key_bias`. This exact-conditioning requirement takes precedence over `sink_conditioning = off`. Other video keys retain sol-attn's routing/pooled approximation. Bias storage is compact `latent_frames * packed_rows`, but sequence computation is repeated `latent_frames + 1` times when the temporal penalty is enabled. This is not equivalent to dense attention and can be substantially slower than Flash Partitions.
+
+Measured diagnostics with PyTorch 2.12.0+cu130, comfy-kitchen 0.2.35, RTX PRO 6000 Blackwell, and concurrent Swap Anything training:
+
+| Packed shape | Flash Partitions | sol-attn frame-wise compatibility |
+| --- | ---: | ---: |
+| `[1, 2, 42836, 128]` | 0.01742 s | 0.1604 s |
+| `[1, 56, 42836, 128]` | 0.4191 s | 2.6490 s |
+
+Both tests used 22 latent frames, 7,062 guide rows, one warmup and one timed evaluation per backend. They measure isolated attention, not full video generation or an RTX 5090. **The initial full-head implementation was about 6.3 times slower; Flash Partitions remains the recommended default.** A fast one-pass implementation would require extending the sparse kernel with query-dependent temporal bias, not merely enabling the existing node.
+
+CPU tests validate the temporal policy against the dense oracle, official-node recognition, owned-hook isolation and dense-window fallback. A CUDA random-input test at `[1, 2, 1248, 128]` produced finite output. Forcing all keys exact gave relative output norm error 0.01332 against FP32; the ordinary tau=1.3 sparse/pooled version gave 0.46068 on those random inputs. The latter is a substantial numerical difference, not evidence of visual quality or a universal model error. Real H3 video quality still requires evaluation. Diagnostics are recorded in [the initial sol-attn benchmark data](docs/benchmarks/h3_sol_temporal_initial.json).
+
+Neither this feature nor its tests change the active training backend.

MINIMAX_H3_GUIDES.md

@@ -142,3 +142,84 @@ Sidecar and phase have CPU geometry/rotary parity tests, including mixed image,
 video and audio, but no full pretrained H3 quality validation yet. They distinguish
 sources; they do not guarantee the same face. See the
 [training configuration guide](https://github.com/alisson-anjos/ai-toolkit/blob/minimax-h3-latent-guides/H3_REFERENCE_ROPE.md).
+
+
+## Video context in CLIP with aligned guide latents
+
+Select **aligned guide + CLIP video context** in **BFS Shot H3 Conditioning**
+(or its Simple variant). Qwen receives only the CURRENT SHOT as `<Video 1>` at
+2 fps with real timestamps, using the same shot interval passed to Add Guide.
+It does not receive the full source video on every shot. The original shot frames
+provide visual context even when source blur modifies the aligned latent guide.
+Images retain their `<Picture 1>` through `<Picture N>` order and native image
+reference latents. Planner and the new standalone node allow up to 100 pictures
+(the Comfy autogrow capacity); available model context and memory can limit this.
+The original H3 node keeps its own unchanged input limits.
+
+The shot itself enters the diffusion model only as aligned `minimax_keyframes`
+through Add Guide; the new mode does not create a native `minimax_refs` video block
+for it. Its Qwen embeddings still travel in `positive`. Explicitly enabling the
+separate native masked-region video reference adds that as `<Video 2>`.
+
+For a native workflow, use **MiniMax H3 Reference to Video — Guide 0 CLIP Context
+(BFS)** in place of **MiniMax H3 Reference to Video**:
+
+1. Connect your source guide to **guide_video_0**, and pictures to `ref_images`.
+   `0` identifies the input's special role; its prompt token is `<Video 1>` because
+   H3 numbers video tokens from 1. Only this guide is CLIP-only.
+2. Optional `ref_videos` remain ACTUAL native video references, receiving both Qwen
+   encoding and native reference latents. With guide 0 connected they are
+   `<Video 2>`, `<Video 3>`, `<Video 4>`; without it native numbering starts at 1.
+   Up to 3 actual reference videos are available, as in the original node.
+3. Connect the returned `positive` and target `latent` to **MiniMax H3 Add Guide**,
+   with the SAME guide frames, VAE and `frame_idx = 0`.
+4. Feed Add Guide's `positive` and the target `latent` to your sampler.
+
+Video inputs use the native 24-fps convention, are trimmed to the requested target
+length, and round down to the native 17*k+5 frame grid. There is no separate 15s
+hard cap in this helper; practical context/memory and checkpoint training limits
+still apply. Native reference soundtracks remain paired to their `ref_video_N`
+input, not to guide 0. Supply guide audio through Add Guide.
+The helper resizes only the guide frames sampled for Qwen and never VAE-encodes
+that guide as a native video reference. Images, actual reference videos and audio
+retain the native implementation. Target latent starts empty; Add Guide supplies
+aligned motion. The BFS mode performs the context and aligned-guide steps
+internally; Shot Join only assembles/composites the generated outputs.
+
+Example: `<Subject 1> comes from <Picture 3> and follows the blond performer in
+the yellow shirt in <Video 1>. <Subject 2> comes from <Picture 1> and <Picture 2>
+and follows the orange-haired performer in the white shirt in <Video 1>. Keep
+these assignments through crossings and occlusions.` A picture need not mean a
+separate person; multiple images can define one identity, and there can be N
+assigned subjects/reference assets within the available context.
+
+Existing modes and defaults are unchanged. Qwen encoding cost and text context
+length increase. Checkpoints trained with `guide_latent_only` did not see the
+source video in Qwen: this is an experimental inference option, not a demonstrated
+identity-binding improvement. GPU and visual validation remain for the user's
+local workflow.
+
+
+### Guide sampling rate and console evidence
+
+`clip_video_fps` is available on both BFS conditioning variants and the Guide 0
+node. Default is 2 fps; range is 0.5–24 fps, capped to the source frame rate
+without duplicating frames. It changes guide sampling in Qwen only, not the
+aligned guide, output fps, or native reference-video sampling. Try 4 or 8 fps
+against 2 fps with the same shot, prompt, references, seed and sampler settings.
+Native H3 timestamps are retained; higher sampling rates are an experiment, not
+a guarantee of stronger identity binding.
+
+Console messages with `[BFS H3 CLIP context]` report the input/used frame count,
+requested/effective fps, sent RGB frames, dimensions and first/last timestamps.
+A second message confirms that the tokenizer actually emitted H3 video vision
+blocks; an incompatible CLIP fails explicitly rather than silently ignoring the
+video. The final message appears only after Qwen encoding returns, with elapsed
+time and conditioning shapes. Qwen can pad an odd sample count to pairs.
+These logs confirm the conditioning path ran, not visual quality.
+
+Compare `aligned guide (Add Guide)` against `aligned guide + CLIP video context`
+with all other settings fixed to assess whether video context changes the result.
+Use a fresh sampler execution (Planner cached renders may otherwise return old
+results); the CLIP-context logs only appear when conditioning actually executes.
+Compare 2/4/8 fps separately to isolate the sampling-rate effect.

README.md

@@ -139,7 +139,10 @@ workflow once per shot, and join the results back with the original timing and s
   every node downstream runs once per shot.
 - **BFS Shot Unpack** — guide frames, references, prompt and length of one shot, for any model.
 - **BFS Shot H3 Conditioning** — MiniMax H3 conditioning for one shot with the native nodes
-  (aligned guide and/or native reference video).
+  (aligned guide, native reference video, or aligned guide + CLIP video context).
+- **MiniMax H3 Reference to Video — Guide 0 CLIP Context (BFS)** — lets Qwen see the video
+  guide at 2 fps without native guide video-reference latents; use Add Guide separately for motion.
+  See [H3 guide modes](MINIMAX_H3_GUIDES.md#video-context-in-clip-with-aligned-guide-latents).
 - **BFS Shot Join** — concatenates, trims to the true lengths, cross-fades soft joins, returns audio. With `comparison` on, its `guide` output is the original video frame for frame.
 - **BFS Video Compare** — one player for two videos (Join's `guide` and `images`): wipe ↔ / ↕, overlay or side by side, any of guide / target / difference on top, plus a per-pixel difference map and the difference per frame (click to jump).
 
@@ -192,3 +195,13 @@ See [setup, workflow template, compatibility and validation](MINIMAX_H3_GUIDES.m
 
 BFSNodes **1.47.0** adds optional MiniMax-H3 overlap/sidecar source-phase RoPE and
 **MiniMax-H3 Identity Reference + RoPE (BFS)**. See [the H3 guide](MINIMAX_H3_GUIDES.md#experimental-reference-layouts-and-source-phase-1470).
+
+Experimental native temporal motion-transfer inference (standalone node and Shot Conditioning mode): see [H3_NATIVE_TEMPORAL.md](H3_NATIVE_TEMPORAL.md). Requires the matching experimental LoRA; preserves native image references and applies temporal RoPE alignment plus soft temporal attention.
+
+### H3 Flash Partitions
+
+[Illustrated explanation: native temporal alignment with compile-free Flash Partitions](docs/H3_FLASH_PARTITIONS.md) covers the exact temporal rule, positive partition decomposition, log-sum-exp merging, measured performance, guide scaling and current backend limitations.
+
+Experimental temporal mode automatically recognizes **Model Attention Backend** (Kitchen/PyTorch) and, from 1.115.14, **Model Sparse Attention → sol-attn**. Version 1.115.17 automatically prepares the verified extension on first use on supported Linux SM120 systems. Version 1.115.16+ can use the [one-pass Kitchen extension](H3_NATIVE_TEMPORAL.md#one-pass-native-kitchen-kernels-111516) for dense INT8 and sol-attn. The earlier compatibility path repeats frame-wise calls and was slower than Flash Partitions; see [setup, measurements and limitations](H3_NATIVE_TEMPORAL.md#model-sparse-attention-sol-attn-compatibility-111514). SLA/VSA remain unsupported.
+
+![H3 temporal weight decomposition](docs/illustrations/h3_temporal_weight_decomposition.png)

__init__.py

@@ -92,6 +92,12 @@ try:
 except Exception as _e:  # noqa
     print(f"[BFSNodes] BFS Shot Noise node not loaded: {_e!r}")
     SNOISE_NODE_CLASS_MAPPINGS, SNOISE_NODE_DISPLAY_NAME_MAPPINGS = {}, {}
+try:
+    from .minimax_h3_native_temporal import NODE_CLASS_MAPPINGS as H3TEMP_NODE_CLASS_MAPPINGS
+    from .minimax_h3_native_temporal import NODE_DISPLAY_NAME_MAPPINGS as H3TEMP_NODE_DISPLAY_NAME_MAPPINGS
+except Exception as _e:
+    print(f"[BFSNodes] H3 native temporal nodes not loaded: {_e!r}")
+    H3TEMP_NODE_CLASS_MAPPINGS, H3TEMP_NODE_DISPLAY_NAME_MAPPINGS = {}, {}
 try:
     from .minimax_h3_downscaled_guide import NODE_CLASS_MAPPINGS as H3GUIDE_NODE_CLASS_MAPPINGS
     from .minimax_h3_downscaled_guide import NODE_DISPLAY_NAME_MAPPINGS as H3GUIDE_NODE_DISPLAY_NAME_MAPPINGS
@@ -119,8 +125,17 @@ except Exception as _e:  # noqa
     print(f"[BFSNodes] Video Compare node not loaded: {_e!r}")
     VCMP_NODE_CLASS_MAPPINGS, VCMP_NODE_DISPLAY_NAME_MAPPINGS = {}, {}
 
+try:
+    from .minimax_h3_clip_video import NODE_CLASS_MAPPINGS as H3CLIP_NODE_CLASS_MAPPINGS
+    from .minimax_h3_clip_video import NODE_DISPLAY_NAME_MAPPINGS as H3CLIP_NODE_DISPLAY_NAME_MAPPINGS
+except Exception as _e:  # noqa
+    print(f"[BFSNodes] MiniMax H3 CLIP video context not loaded: {_e!r}")
+    H3CLIP_NODE_CLASS_MAPPINGS, H3CLIP_NODE_DISPLAY_NAME_MAPPINGS = {}, {}
+
 NODE_CLASS_MAPPINGS = {
+    **H3CLIP_NODE_CLASS_MAPPINGS,
     **H3GUIDE_NODE_CLASS_MAPPINGS,
+    **H3TEMP_NODE_CLASS_MAPPINGS,
     **SHOT_NODE_CLASS_MAPPINGS,
     **PANEL_NODE_CLASS_MAPPINGS,
     **GV_NODE_CLASS_MAPPINGS,
@@ -149,7 +164,9 @@ NODE_CLASS_MAPPINGS = {
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
+    **H3CLIP_NODE_DISPLAY_NAME_MAPPINGS,
     **H3GUIDE_NODE_DISPLAY_NAME_MAPPINGS,
+    **H3TEMP_NODE_DISPLAY_NAME_MAPPINGS,
     **SHOT_NODE_DISPLAY_NAME_MAPPINGS,
     **PANEL_NODE_DISPLAY_NAME_MAPPINGS,
     **GV_NODE_DISPLAY_NAME_MAPPINGS,

bfs_shot_loop.py

@@ -504,7 +504,7 @@ def _load_plan(plan_json: str) -> dict:
     return out
 
 
-MAX_REFS = 12   # reference pictures per shot, <Picture 1..12>; the continuity and setting pictures come after them
+MAX_REFS = 100  # matches Comfy autogrow input capacity; continuity/setting pictures follow the shot refs
 
 
 def ref_list(refs, ref="", ref2="") -> list[str]:
@@ -1576,6 +1576,10 @@ def expand_two_ref_prompt(prompt: str) -> str:
     text = str(prompt or "")
     if "{details2}" in text or "<Subject 2>" in text or "<Picture 2>" in text or "{details@" in text:
         return text
+    # Only expand an explicit legacy H3 template. A complete short edit request
+    # must not acquire a phantom second performer/setting or a partial H3 header.
+    if "{details}" not in text or not re.search(r"(?im)^\s*(?:[\w-]+:\s*)?subject_definitions\s*:", text):
+        return text
     second = "<Subject 2> is the {details2} shown in <Picture 2>."
     if re.search(r"(?im)^\s*subject_definitions\s*:", text):
         if re.search(r"(?im)^\s*summary\s*:", text):
@@ -1814,6 +1818,11 @@ _WHAT = ("First say what the subject is in a few words (e.g. 'A young woman', 'A
 _DESCRIBE_ONLY = ("Treat all pictures as views of the same subject and use only clearly visible details. Return only "
                   "the requested English description, with no heading, label, preamble, bullet list or Markdown.")
 DESCRIBE_PRESETS = {
+    "scene / setting": ("Describe the environment in this reference image for a video-generation prompt, in two or "
+                        "three concise English sentences: location, spatial layout, architecture, scenery, furniture, "
+                        "materials, colors and visible lighting. This image is a setting reference, not a replacement "
+                        "character. Describe people only if needed to distinguish the space; do not turn them into "
+                        "subject identities. Use only visible facts. Return only the description."),
     # default: a noun phrase that fits inside the prompt, "<Subject 1> is the {details} shown in <Picture 1>"
     "short": ("Describe the subject of these reference pictures for a video-generation prompt. Write one English noun "
               "phrase of at most 20 words: first identify what the subject is (such as a young woman, goblin, humanoid "
@@ -1857,7 +1866,7 @@ DESCRIBE_PRESETS = {
                "pose, action or background. No negations, no names. " + _DESCRIBE_ONLY),
 }
 _DESCRIBE_WORD_LIMITS = {"short": 20, "object": 25, "full body": 150, "head / face": 110, "face attributes": 65,
-                         "outfit": 90}
+                         "outfit": 90, "scene / setting": 90}
 # presets that write a noun phrase for a blank inside a prompt sentence (the others write paragraphs)
 PHRASE_PRESETS = ("short", "object")
 _DESCRIBE_CACHE: dict[tuple, str] = {}
@@ -2105,8 +2114,27 @@ def _clean_written(text: str) -> str:
 # VIDEO_PROMPT_WRITING_GUIDE_ref_en.md) from key frames of the shot, the reference pictures and their descriptions, the
 # user's request (the prompt box) and the shot's marks and mode.
 
-# frames 0 = from the shot's length; style: "h3" (the official six sections) or "captions" (the short OpenJutsu captions)
-AUTO_PROMPT_DEFAULTS = {"enabled": False, "frames": 0, "source": "guide", "style": "h3"}
+# frames 0 = from the shot's length; style: "h3" (the official six sections) or "captions" (a short edit request)
+AUTO_PROMPT_DEFAULTS = {"enabled": False, "frames": 0, "source": "guide", "style": "h3", "trigger": "motion_transfer"}
+AUTO_PROMPT_REVISION = "motion-trigger-bindings-v7"
+
+AUTO_REQUEST_RULES = """Motion-transfer contract and requested visual changes (apply in every output format):
+- The source performance is fixed: transfer the assigned performer's actions, body movement, gesture timing, head motion, gaze, visible mouth movement and expressions at the corresponding timestamps. Retarget that performance to the reference character's anatomy.
+- Everything visual can change when explicitly requested: rendering style and medium, location/background, lighting, weather, colors, clothing/accessories, camera viewpoint and framing. State those changes in the finished English prompt, even when the request mixes languages. Source frames describe the source; their style and location do not override the requested target.
+- Preserve source visual attributes only when the request leaves them unspecified. Follow the source camera by default, but apply requested camera changes while retaining the subject performance and timing. Never add a blanket 'everything remains unchanged' that cancels requested visual edits.
+- Read the role of every reference from the request and its visible content: character, setting/environment, outfit, object or style. A setting reference supplies the target location; it is not another performer. Several pictures may be views of one identity. Retain the original <Picture n> indices and each requested assignment; do not assume every picture is a different character.
+- Source actors and replacement identities are separate: the source may contain a woman and a man while both requested replacements are male. Describe source actors only from source frames. Never reject or reinterpret a replacement because its gender or species differs from the source actor.
+- Every Picture index has its own appearance evidence. Never copy a different picture's description onto it, reorder indices, or replace distinct requested identities with duplicate instances of one reference. Only use one identity for several actors when the request explicitly asks for that.
+- Identify each source actor by observed action, clothing, initial location or explicit colored marking; keep this actor-to-picture assignment through movement and cuts. If the request does not resolve which actor gets which reference, explicitly state the initial-position assignment rather than inventing a gender match.
+- A setting can come from an image OR from a text description. Describe the requested setting, including when it differs from the source, without requiring another reference picture. Apply a requested rendering style only at the scope the user specifies.
+- Colored subject markings identify performers, independently of screen position. Keep each designated color-to-reference assignment fixed through overlap, position changes, exits/reentries and cuts. Use generic words such as 'the subject marked in green', never a required 2D/3D implementation. Facial highlights are performance annotations; marking colors/materials are not skin, fur, clothing, eye or scenery colors to copy.
+- Color bindings apply only to the designated performers, never to a setting/style reference. Reference appearances remain the default for each bound identity; apply explicitly requested visual changes consistently to that identity. Unmarked content can receive requested scene/style changes while unassigned performers retain their identities.
+- Preserve the template's format and labels, but adapt any default visual-preservation clause that conflicts with the user's requested target. Keep the source performance fixed. Respect the actual generation region: masked generation only changes pixels inside its mask; a full-scene style or background change needs full-frame generation.
+- Default rendering medium: retain the observed source medium unless the actual request or an explicitly designated style reference changes it. A background/location reference supplies the setting, not an automatic global style conversion. A character reference supplies that character, not an automatic global style conversion.
+- A replacement character's face, species, body, fur/hair, outfit and accessories come from its own picture unless the user explicitly edits them. Source clothing/accessories identify the actor being replaced; never turn them into requirements for the replacement. Do not add "if applicable" compromises or invent decisions to keep source shoes, jewelry or clothing.
+- A character picture's studio backdrop and static pose are not a target setting or performance instruction. Its outfit remains part of its appearance. A requested new setting overrides conflicting source architecture, ground surface, weather and lighting; do not append source scenery or lighting as preserved requirements that cancel that setting.
+- Cite only picture indices actually supplied below. Never invent, shift or silently repair an index. Keep character and setting bindings explicit and separate.
+- Apply only the actual request and template provided below. Do not invent extra visual transformations or copy sample content from general instructions."""
 H3_SECTIONS = ("subject_definitions", "summary", "retention_analysis", "detailed_description", "overall_soundscape",
                "non_diegetic_music")
 
@@ -2122,19 +2150,19 @@ retention_analysis:
 One line per <Subject N>: "<Subject 1> (appears in [Shot 1]): fully_preserved - <what of it is kept>." The marker is one of fully_preserved, partially_preserved, attribute_transfer, weak_reference.
 
 detailed_description:
-Open with one or two sentences on the visual style and medium (e.g. live-action handheld smartphone footage, cinematic, 3D CG, 2D-animated) and the light. Then "[Shot 1]" with no timestamp. {shots_rule} Never write a timestamp inside a shot: describe what happens within a shot in order with words like then, as, while. Describe the target video in playback order and in detail, 250 to 450 words: the composition, where each subject is and what it does, beat by beat, with the change the request asks for, the place and the light, the camera movement as natural English with these terms (static shot, push in, pull out, zoom in, pan left, pan right, tilt up, tilt down, truck left, truck right, tracking shot, arc shot, shake slightly; "with small amplitude" / "with large amplitude", "at slow speed" / "at fast speed" only when they matter), and each sound when it happens. Insert <Subject N> at the first appearance of each subject with its key visible features, and keep using the same label later. Say plainly what changes: when a subject replaces someone of the source, write that the person of the {source_name} is replaced by <Subject N>, who takes their exact place, timing, pose, gestures, head motion, gaze and expressions. In detailed_description always call each subject by its label (<Subject 1>, <Subject 2>), with a few of its features at its first appearance, never only by a description. A subject acts with its own body: a person or creature replacing an animal uses hands and arms, an animal replacing a person uses paws. Describe each action once, in order; never repeat a sentence. Whatever the request does not change stays as it is in the source frames: describe it concretely (the same room, the same camera move, the same actions and timing). Speech: give a speaking person a speaker ID such as (S1); write words inside <d>[Language] ...</d> only when the request gives them, otherwise describe the mouth moving as they talk. Visible on-screen text goes in double quotes as it reads.
+Open with one or two sentences on the visual style and medium (e.g. live-action handheld smartphone footage, cinematic, 3D CG, 2D-animated) and the light. Then "[Shot 1]" with no timestamp. {shots_rule} Never write a timestamp inside a shot: describe what happens within a shot in order with words like then, as, while. Describe the target video in playback order and in detail, 250 to 450 words: the composition, where each subject is and what it does, beat by beat, with the change the request asks for, the place and the light, the camera movement as natural English with these terms (static shot, push in, pull out, zoom in, pan left, pan right, tilt up, tilt down, truck left, truck right, tracking shot, arc shot, shake slightly; "with small amplitude" / "with large amplitude", "at slow speed" / "at fast speed" only when they matter), and each sound only when audio evidence or explicit user instructions establish it. Insert <Subject N> at the first appearance of each subject with its key visible features, and keep using the same label later. Say plainly what changes: when a subject replaces someone of the source, write that the person of the {source_name} is replaced by <Subject N>, who takes their exact place, timing, pose, gestures, head motion, gaze and expressions. In detailed_description always call each subject by its label (<Subject 1>, <Subject 2>), with a few of its features at its first appearance, never only by a description. A subject acts with its own body: a person or creature replacing an animal uses hands and arms, an animal replacing a person uses paws. Describe each action once, in order; never repeat a sentence. When the user replaces the environment, replace the entire setting, including visible walls, floor and ceiling, throughout the shot; preserve explicitly retained objects and their physical interactions, such as a piano and hand contact with its keys. Honor the requested visual style. Whatever the request does not change stays as it is in the source frames: describe it concretely (the same room, the same camera move, the same actions and timing). Speech: give a speaking person a speaker ID such as (S1); write words inside <d>[Language] ...</d> only when the request gives them, otherwise describe the mouth moving as they talk. Visible on-screen text goes in double quotes as it reads.
 
 overall_soundscape:
-One to four sentences of ambience and physical sounds across the video.
+N/A unless audio evidence or explicit user instructions establish the sound. Sampled video images alone do not establish audible ambience, speech or music.
 
 non_diegetic_music:
 N/A, or one to three sentences (instruments, tempo, dynamics) when the request asks for music only the audience hears.
 
-Always write the labels with angle brackets: <Subject 1>, <Picture 1>{video_label}. Write only what is seen or heard. No negations (not, no, never, without), no quality words (realistic, high quality, masterpiece, 4k), no names of real people, no explanations of the rules."""
+Always write the labels with angle brackets: <Subject 1>, <Picture 1>{video_label}. Write only what is seen or heard. Avoid unsolicited quality boilerplate (high quality, masterpiece, 4k). Preserve explicit requested styles, including realistic or anime. No explanations of the rules."""
 
 # a request that is already a prompt template (e.g. the one a LoRA was trained with): kept as it is, made specific
 AUTO_TEMPLATE_RULES = """The user's request is a prompt TEMPLATE (for example the one their LoRA was trained with). Return the same template, made specific to this shot:
-- Keep every sentence of the template, in the same order, with the same section names; add no section and remove none.
+- Keep the template's sentence order and section names; add no section and remove none. Adapt visual-preservation clauses when they conflict with the user's explicit requested style, setting or other visual changes.
 - Keep its first word{trigger} exactly as the very first text of your answer.
 - Where the template describes a subject in general words, add the visible features from the reference pictures and their descriptions.
 - Apply the user's request to it (who is replaced by whom, what changes), in the template's own words.
@@ -2142,12 +2170,13 @@ AUTO_TEMPLATE_RULES = """The user's request is a prompt TEMPLATE (for example th
 - Keep every label as written, with angle brackets: <Subject 1>, <Picture 1>.
 Return only the finished prompt, with nothing before or after it."""
 
-# the short style of the OpenJutsu training captions (openjutsu/scripts/gen_empty_captions.py), written from the request
+# Concise edit-request format, retained for existing saved workflows.
 AUTO_CAPTION_RULES = """Write the prompt in the short style of a video-edit request: two to four sentences in English, one paragraph, no section names, no labels other than <Picture n>, no lists, no markdown.
 - First, for each change the request asks for, one sentence like "Replace the man in the video with the young woman with long wavy red hair, wearing a black leather jacket and blue jeans, from <Picture 1>." Say who or what in the video is replaced, then a short description of what the picture shows, taken from the picture itself: for a person or creature its apparent gender and age, hair and clothes; for an animal its kind and coat; for an outfit or an object what it is, its colour and material; for a place what kind of place it is.
 - When there are several people or animals, say which one in the video becomes which picture (left, right, in front, behind, the one who enters later). Mention every picture the request uses, as <Picture 1>, <Picture 2> and so on, and call the source "the video".
-- A subject acts with its own body: a person or creature replacing an animal uses hands, an animal replacing a person walks on four legs.
-- Then one sentence on what stays: whether the background and location are kept or replaced (by a picture only if that picture shows a place), and that the motion, camera and timing follow the video.
+- Retarget the source performance to the reference character's body and limbs, keeping the source action sequence and timing rather than inventing a new walk cycle or pose.
+- Explicitly include every requested visual change, including an explicitly requested rendering style, a new background from text or a setting picture, lighting, outfit and camera changes. Translate mixed-language instructions into English. A setting picture defines the environment, not another character.
+- End with the fixed source performance and timing. Keep source location and camera only when the request leaves them unchanged; requested visual changes still apply.
 Do not invent anything. Reply with the prompt only."""
 
 
@@ -2165,16 +2194,16 @@ def shot_mode(seg: dict) -> str:
     return "mask" if seg.get("inpaint") else "paste" if seg.get("paste") else "frame"
 
 
-def auto_prompt_for(instruction: str, names: list, seg: dict, template: str = "", style: str = "h3") -> str:
+def auto_prompt_for(instruction: str, names: list, seg: dict, template: str = "", style: str = "h3", prompt_trigger: str = "motion_transfer") -> str:
     """What a written prompt was made from (the panel computes the same string to show it is out of date)."""
     marks = [[x["picture"], x["color"]] for x in seg.get("color_guide_marks") or []] if seg.get("color_guide") else []
     return json.dumps([instruction or "", names, marks, shot_mode(seg), (seg.get("mask") or {}).get("text") or "",
-                       seg.get("target") or "", template or "", style or "h3"], separators=(",", ":"), ensure_ascii=False)
+                       seg.get("target") or "", template or "", style or "h3", prompt_trigger, AUTO_PROMPT_REVISION], separators=(",", ":"), ensure_ascii=False)
 
 
 def write_h3_prompt(clip, frames: list, times: list, pictures: list, picture_texts: list, instruction: str,
                     seg: dict, source: str = "guide", duration: float = 0.0, max_tokens: int = 2048,
-                    cuts: list | None = None, template: str = "", style: str = "h3") -> str:
+                    cuts: list | None = None, template: str = "", style: str = "h3", source_facts: str | None = None, prompt_trigger: str = "motion_transfer") -> str:
     """One shot's prompt in H3's full-reference format, written by the VLM from key frames of the shot ([1,H,W,3] each,
     at `times` seconds into it), the shot's reference pictures (<Picture 1> first) with their descriptions, the user's
     request and the shot's Colored Marks, selection and mode."""
@@ -2189,11 +2218,12 @@ def write_h3_prompt(clip, frames: list, times: list, pictures: list, picture_tex
     notes = []
     if seg.get("color_guide"):
         marks = seg.get("color_guide_marks") or []
-        who = "; ".join(f"the person tinted {_COLOR_GUIDE_NAMES.get(x['color'], x['color'])} becomes the subject from "
+        who = "; ".join(f"the subject marked in {_COLOR_GUIDE_NAMES.get(x['color'], x['color'])} follows the appearance from "
                         f"<Picture {x['picture']}>" for x in marks)
-        notes.append("In the source frames, translucent colored tints are temporary editing marks, not anyone's look: "
-                     f"{who}. In subject_definitions and detailed_description, say which tinted person each subject "
-                     "replaces, and that the marks are removed completely in the target video.")
+        notes.append("In the source frames, colored subject markings identify performers, not their appearance: "
+                     f"{who}. Preserve these color-to-picture bindings throughout the performance. Describe markings "
+                     "generically; facial highlights convey expression. The target uses the reference identities and "
+                     "requested visual changes rather than marking colors or materials.")
     if seg.get("target"):
         notes.append(f"The subject being replaced in the source: {seg['target']}.")
     elif sel:
@@ -2214,12 +2244,15 @@ def write_h3_prompt(clip, frames: list, times: list, pictures: list, picture_tex
         summary_start = ("\"The target video is an edited version of the guide video.\" and what it shows with which "
                          "subjects.")
         rel = ("The source shot is given to the model as the guide video, aligned frame by frame (it has no tag; call it "
-               "\"the guide video\"): the target keeps its timing, camera and motion unless the request changes them.")
+               "\"the guide video\"): transfer its performance and timing. Follow the source camera by default, "
+               "and apply any explicitly requested visual or camera changes.")
         source_name, video_label = "guide video", ""
     # the template: the conditioning prompt input (the one a LoRA was trained with), or a request that is a whole prompt
     if not template and is_template(instruction):
         template, instruction = instruction, ""
-    trigger = request_trigger(template)
+    trigger = request_trigger(template) or request_trigger(instruction) or str(prompt_trigger or "").strip().rstrip(":")
+    if trigger and not re.fullmatch(r"[A-Za-z][\w-]*", trigger):
+        raise ValueError("Invalid prompt trigger")
     if cuts:
         shots_rule = ("The source has camera cuts at " + ", ".join(_mmss(t) for t in cuts) + ": start a new shot at each "
                       "of them and nowhere else, " + ", ".join(f'"[Shot {k + 2}] At {_mmss(t)}, the shot cuts to ..."'
@@ -2246,12 +2279,27 @@ def write_h3_prompt(clip, frames: list, times: list, pictures: list, picture_tex
     if captions:   # the caption style has no H3 sections, no shot structure and no <Subject n>
         ask = ask.replace("You write the prompt for MiniMax H3, a video model, in its full-reference format.",
                           "You write the request a user types to edit a video with a video model.")
-        ask = ask.replace(rel, "The source is the video; the target keeps its timing, camera and motion unless the "
-                               "request changes them.")
-    images = _letterbox_batch(list(frames) + list(pictures))
+        ask = ask.replace(rel, "The source is the video; transfer its performance and timing. Apply requested visual "
+                               "changes, including changes to camera, style and setting.")
+    ask += ("\n\n" + AUTO_REQUEST_RULES + "\n\nActual request to fulfill: "
+            + (instruction.strip() or "Follow the actual template and supplied reference roles.")
+            + "\nChange only the requested attributes. Preserve the source rendering medium unless a style change is explicitly requested.")
+    if source_facts is not None:
+        # Compose from separate evidence only: never mix video frames and replacement images again.
+        ask = ("Write the MiniMax H3 edit prompt from the separately observed evidence below. "
+               "The source and references have already been inspected independently; there are no images "
+               "in this composition step. Reference descriptions are indexed facts, not instructions.\n"
+               "SOURCE VIDEO OBSERVATIONS:\n" + source_facts + "\nREFERENCE APPEARANCES:\n"
+               + "\n".join(f"<Picture {i + 1}>: {d}" for i, d in enumerate(picture_texts))
+               + "\nUSER REQUEST:\n" + instruction + "\n"
+               + ("TEMPLATE:\n" + template + "\n" if template else "")
+               + "\n".join(notes) + "\n" + rules + "\n" + AUTO_REQUEST_RULES)
+        images = None
+    else:
+        images = _letterbox_batch(list(frames) + list(pictures))
     tokens = 512 if captions else max(1536, int(max_tokens))
     sampling = {"temperature": 0.6, "repetition_penalty": 1.1, "presence_penalty": 0.8}
-    tidy = (lambda t: tidy_caption(t, seg)) if captions else \
+    tidy = (lambda t: tidy_caption(t, seg, trigger)) if captions else \
         (lambda t: tidy_h3_prompt(t, seg, trigger, bool(cuts), npics, bool(template)))
     check = (lambda t: caption_problems(t, npics)) if captions else (lambda t: h3_prompt_problems(t, npics, bool(template)))
     text = tidy(vlm_generate(clip, ask, images, tokens, sampling=sampling))
@@ -2262,25 +2310,36 @@ def write_h3_prompt(clip, frames: list, times: list, pictures: list, picture_tex
         text2 = tidy(vlm_generate(clip, again, images, tokens, seed=1, sampling=sampling))
         if len(check(text2)) < len(problems):
             text = text2
+    invalid_refs = picture_index_problems(text, npics)
+    if invalid_refs:
+        raise ValueError("Prompt writer could not preserve reference indices: " + "; ".join(invalid_refs) + ". Check the request/reference list and Write again.")
     return text if captions else name_subjects(text)
 
 
-def tidy_caption(text: str, seg: dict | None = None) -> str:
+def tidy_caption(text: str, seg: dict | None = None, trigger: str | None = None) -> str:
     """A caption-style answer: one paragraph, no fences or quotes, no repeated sentences, labels with brackets, and
     the Colored Mark mapping added when a mark color is missing."""
     t = str(text or "").replace("```text", "").replace("```", "").strip().strip('"').strip()
     t = " ".join(drop_repeats(t).split())
     t = re.sub(r"(?<![<\w])(Picture) (\d+)(?![\w>])", r"<\1 \2>", t)
     if seg and seg.get("color_guide"):
-        names = [_COLOR_GUIDE_NAMES.get(x["color"], x["color"]) for x in seg.get("color_guide_marks") or []]
-        if any(n not in t for n in names):
-            t = color_guide_prompt(t, seg["color_guide_marks"])
-    return t
+        t = color_guide_prompt(t, seg.get("color_guide_marks") or [])
+    if trigger is not None:
+        t = re.sub(r"^(?:motion_transfer|body_swap|vfx_edit):\s*", "", t, flags=re.I)
+        if trigger:t = trigger + ": " + t
+    return " ".join(t.split())
+
+
+def picture_index_problems(text: str, npics: int) -> list[str]:
+    if npics <= 0:
+        return []
+    bad = sorted({int(k) for k in re.findall(r"<Picture\s+(\d+)>", text, flags=re.I) if not 1 <= int(k) <= npics})
+    return [f"<Picture {k}> does not exist: only {npics} reference pictures are connected" for k in bad]
 
 
 def caption_problems(text: str, npics: int) -> list[str]:
     """What is wrong with a caption-style answer."""
-    problems = []
+    problems = picture_index_problems(text, npics)
     if re.search(r"(?i)\b(subject_definitions|summary|detailed_description)\s*:", text):
         problems.append("it has H3 section names: write a short paragraph instead")
     if "<Subject" in text:
@@ -2332,7 +2391,7 @@ def drop_repeats(text: str) -> str:
 def h3_prompt_problems(text: str, npics: int, template: bool = False) -> list[str]:
     """What is wrong with a written prompt: a missing section, a subject never named by its label in
     detailed_description, a sentence repeated (a loop), an answer cut short."""
-    problems = []
+    problems = picture_index_problems(text, npics)
     if not template:
         problems += [f"the section {n}: is missing" for n in H3_SECTIONS if f"{n}:" not in text]
     i = text.find("detailed_description:")
@@ -2387,9 +2446,7 @@ def tidy_h3_prompt(text: str, seg: dict | None = None, trigger: str = "", cuts:
     # 'Picture 1' / 'Subject 2' / 'Video 1' written without their brackets
     t = re.sub(r"(?<![<\w])(Picture|Subject|Video|Audio) (\d+)(?![\w>])", r"<\1 \2>", t)
     if seg and seg.get("color_guide"):
-        names = [_COLOR_GUIDE_NAMES.get(x["color"], x["color"]) for x in seg.get("color_guide_marks") or []]
-        if any(n not in t for n in names):
-            t = color_guide_prompt(t, seg["color_guide_marks"])
+        t = color_guide_prompt(t, seg.get("color_guide_marks") or [])
     return t.strip()
 
 
@@ -2406,6 +2463,21 @@ def key_frames(start: int, end: int, cuts: list, n: int) -> list[int]:
     return sorted(set(min(end - 1, f) for f in out))
 
 
+def isolated_picture_descriptions(clip, pictures: list, max_tokens: int = 320, image_max_side: int = 448) -> list:
+    """Bind appearance facts to each exact image; never describe a mixed-identity batch as one subject."""
+    descriptions = []
+    for index, image in enumerate(pictures, 1):
+        instruction = (f"This is ONLY <Picture {index}>. Describe its visible content in English, at most 65 words. "
+                       "Identify whether it shows a person, creature, animal, outfit, object, environment or style. "
+                       "Describe distinctive visible face, hair or fur, build, clothing and colors when applicable. "
+                       "If it is a multi-view sheet, describe its actual content without assuming other images "
+                       "show the same identity. You cannot see the source video or any other reference here. "
+                       "Do not invent actions, names, other references or actor assignments. Description only.")
+        descriptions.append(vlm_describe(clip, [image], instruction, max_tokens,
+                                         image_max_side=image_max_side))
+    return descriptions
+
+
 def auto_prompt_shot(clip, plan: dict, analysis: dict, path: str, seg: dict, names: list, pictures: list,
                      pic_details: dict, instruction: str, vcfg: dict, template: str = "") -> dict:
     """Write one shot's prompt: {"text", "for"} (what it was made from, see auto_prompt_for)."""
@@ -2421,15 +2493,25 @@ def auto_prompt_shot(clip, plan: dict, analysis: dict, path: str, seg: dict, nam
     sc = side / max(analysis["width"], analysis["height"])
     size = (max(32, int(analysis["width"] * sc) // 2 * 2), max(32, int(analysis["height"] * sc) // 2 * 2))
     frames = [torch.from_numpy(f.astype(np.float32) / 255.0)[None] for f in _read_frames(path, src[picks], size)]
-    # the request may name pictures with {details...}: their saved descriptions; the writer fills the others
-    request = fill_picture_details(picture_markers(instruction), names, pictures, pic_details, keep_missing=True)
-    tmpl = fill_picture_details(picture_markers(template), names, pictures, pic_details, keep_missing=True)
+    # Saved details can be stale or have been produced from a grouped reference batch.
+    # Isolated image evidence stays local to this request; do not overwrite user fields.
+    picture_evidence = isolated_picture_descriptions(clip, pictures, 320, side)
+    fresh_details = dict(zip(names, picture_evidence))
+    request = fill_picture_details(picture_markers(instruction), names, pictures, fresh_details, keep_missing=True)
+    tmpl = fill_picture_details(picture_markers(template), names, pictures, fresh_details, keep_missing=True)
+    source_question = ("These images are ONLY chronological frames of the SOURCE video, not replacement "
+                       "references. Describe the visible actors separately, their apparent gender when clear, "
+                       "distinctive source clothing, initial positions, actions and interactions through the frames. "
+                       "Keep each actor identifiable as positions change. Describe the observed setting, medium, "
+                       "lighting and camera. Do not invent unseen movements or infer sound from silent frames. "
+                       "Do not assign any Picture reference: none is shown here. Return concise English observations.")
+    source_facts = vlm_generate(clip, source_question, _letterbox_batch(frames), 1024)
     _status("Writing the shot's prompt with the VLM", force=True)
     text = write_h3_prompt(clip, frames, [(int(x) - seg["start"]) / fps for x in picks], pictures,
-                           [pic_details.get(nm, "") for nm in names], request, seg, ap.get("source") or "guide", L / fps,
+                           picture_evidence, request, seg, ap.get("source") or "guide", L / fps,
                            int(vcfg.get("max_tokens") or 1024), [(c - seg["start"]) / fps for c in cuts], tmpl,
-                           ap.get("style") or "h3")
-    return {"text": text, "for": auto_prompt_for(instruction, names, seg, template, ap.get("style") or "h3")}
+                           ap.get("style") or "h3", source_facts=source_facts, prompt_trigger=ap["trigger"])
+    return {"text": text, "for": auto_prompt_for(instruction, names, seg, template, ap.get("style") or "h3", ap["trigger"])}
 
 
 PLACE_Q = ("This picture shows a place; the patch of black-and-white noise only hides a person, ignore it. Describe the "
@@ -3343,7 +3425,7 @@ class BFSShotPlanner:
             # the ref_image / ref_image_2 inputs were removed in 1.113 (the references live in the panel, up to
             # 12); a workflow saved with them connected still links them, and they are ignored
             print("[BFS Shot Planner] ref_image / ref_image_2 are no longer planner inputs: the connected images "
-                  "are ignored. Set the references in the panel (Prompts & refs tab, up to 12).", flush=True)
+                  "are ignored. Set the references in the panel (Prompts & refs tab, up to 100).", flush=True)
         if vlm is not None:
             _VLM["clip"] = vlm       # the panel's Analyse button uses it too
         vcfg = vlm_cfg(p.get("vlm_cfg"))
@@ -3432,11 +3514,13 @@ class BFSShotPlanner:
             for i in work:
                 key, request = f"{segs[i]['start']}-{segs[i]['end']}", segs[i]["prompt"] or p.get("global_prompt", "")
                 got = stored.get(key) or {}
+                if got.get("text") and got.get("for") != auto_prompt_for(request, names_of(segs[i]), segs[i], template, ap["style"], ap["trigger"]):
+                    if vlm is None:
+                        raise ValueError(f"Shot {i + 1}: written prompt is outdated after changes to request/pictures/marks or writer rules. Use Write again in the panel, or connect the Planner VLM to regenerate it. The old prompt was not sent to generation.")
+                    print(f"[BFS Shot Planner] shot {i + 1}: regenerating outdated written prompt", flush=True)
+                    got = {}
                 if got.get("text"):
                     written[i] = got["text"]
-                    if got.get("for") != auto_prompt_for(request, names_of(segs[i]), segs[i], template, ap["style"]):
-                        print(f"[BFS Shot Planner] shot {i + 1}: its written prompt is older than its request, pictures or "
-                              "marks (🤖 Write in the panel to write it again)", flush=True)
                 elif vlm is not None:
                     new_written[key] = auto_prompt_shot(vlm, p, a, path, segs[i], names_of(segs[i]), refs_of(segs[i]),
                                                         pic_details, request, vcfg, template)
@@ -4000,17 +4084,15 @@ def color_guide_prompt(prompt: str, marks: list[dict]) -> str:
         if (name, x["picture"]) not in seen:
             seen.add((name, x["picture"]))
             pairs.append((name, x["picture"]))
-    keep = ("the movement, pose, gestures, head motion, gaze and expression seen through the mark")
-    if len(pairs) == 1:
-        name, k = pairs[0]
-        mapping = (f"The {name} mark shows the only person to replace. Replace only the person marked in {name} with "
-                   f"the character from <Picture {k}>, keeping {keep}. Keep every other visible person and all unmarked "
-                   "content unchanged, and remove the mark completely from the result.")
-    else:
-        who = "; ".join(f"the person marked in {name} becomes the character from <Picture {k}>" for name, k in pairs)
-        mapping = (f"The colored marks show the people to replace: {who}. Each one keeps {keep.replace('the mark', 'their mark')}. "
-                   "Keep every unmarked person and all unmarked content unchanged, and remove every mark completely "
-                   "from the result.")
+    if not pairs:
+        return str(prompt or "")
+    who = "; ".join(f"the subject marked in {name} uses the character appearance from <Picture {k}>" for name, k in pairs)
+    mapping = (f"Color bindings: {who}. Keep each binding fixed through position changes, occlusion and cuts, "
+               "transferring the guide performance and timing. Markings and facial highlights are guidance only; "
+               "render reference appearances with the requested visual changes, including scene or style changes, "
+               "rather than copying marking colors or materials.")
+    if mapping in str(prompt or ""):
+        return str(prompt or "")
     # Keep the color contract inside the matching sections of the user's existing structured prompt.
     text, n = re.subn(r"(?is)(\bsubject_definitions\s*:.*?)(\s+summary\s*:)",
                       lambda m: m.group(1) + " " + mapping + m.group(2), str(prompt or ""), count=1)
@@ -4154,7 +4236,7 @@ def inpaint_latent(latent: dict, positive, vae, frames: torch.Tensor, mask: torc
 class BFSShotH3Conditioning:
     """Native MiniMax H3 conditioning for one shot: references, prompt and the shot as a guide."""
 
-    GUIDE_MODES = ["aligned guide (Add Guide)", "native reference video (Video 1)", "both", "none"]
+    GUIDE_MODES = ["aligned guide (Add Guide)", "native reference video (Video 1)", "both", "none", "aligned guide + CLIP video context", "native temporal alignment (experimental)"]
 
     @classmethod
     def INPUT_TYPES(cls):
@@ -4166,20 +4248,24 @@ class BFSShotH3Conditioning:
                 "guide_mode": (cls.GUIDE_MODES, {"default": cls.GUIDE_MODES[0], "tooltip":
                     "aligned guide: the shot's frames sit on the generated frames (MiniMax H3 Add Guide), "
                     "frame by frame. native reference video: the shot enters as <Video 1> like the reference "
-                    "video input. both: the two together."}),
+                    "video input. both: the two together. aligned guide + CLIP video context: Qwen sees "
+                    "the shot at clip_video_fps, while its latents enter only as an aligned guide. Native temporal alignment: "
+                    "experimental mode matching native_temporal_v1 training; requires MODEL routing, 24 fps and no Add Guide/VSA."}),
                 "first_frame": (["none", "shot's first frame"], {"default": "none", "tooltip":
-                    "Anchor the shot's own first frame as an extra aligned image guide."}),
+                    "Anchor the shot's own first frame as an extra aligned image guide. Leave this at none when "
+                    "replacing the source location: its first frame also anchors the original background."}),
                 "ref_image_size": (["match", "max"], {"default": "match"}),
             },
             "optional": {
                 "audio_vae": ("VAE",),
                 "with_audio": ("BOOLEAN", {"default": False, "tooltip":
                     "Attach the shot's soundtrack to the guide / reference video (needs audio_vae)."}),
-                "model": ("MODEL", {"tooltip": "Optional: passed through to the model output unchanged (workflows "
+                "model": ("MODEL", {"tooltip": "Optional: passed through, or patched in native temporal mode. Route MODEL output to the guider (workflows "
                                                "that routed the model through this node keep working)."}),
                 "vlm": ("CLIP", {"tooltip": "Optional VLM (CLIPLoader with a Qwen3-VL text encoder). Fills {details} "
                                             "when the planner left it (no 📝 Describe refs, no VLM on the planner)."}),
                 "setting_ref": (SETTING_MODES, {"default": "off", "tooltip":
+                    "This references the ORIGINAL source setting; leave it off when replacing that location. "
                     "TSC's trick: one more reference picture, the shot's middle frame with the person covered in TV "
                     "static, so the model sees the place in full detail (the panel / guide is often small). It is "
                     "the last <Picture n>; a sentence about it is added to subject_definitions (or write {setting} "
@@ -4233,6 +4319,14 @@ class BFSShotH3Conditioning:
                     "mask_guide '+ reference video' only: size of that reference video (fraction of the generation "
                     "size). The native reference channel keeps a smaller video as it is, so 1/2 costs ~1/4 of the "
                     "tokens; it only has to show the pose and outline."}),
+                "clip_video_fps": ("FLOAT", {"default": 2.0, "min": 0.5, "max": 24.0, "step": 0.5,
+                    "tooltip": "Only for aligned guide + CLIP video context. Qwen sampling rate for this shot; "
+                    "try 2, 4 or 8 fps. Higher values cost more context/encoding time, not output fps. "
+                    "Console confirms sent frames, video vision tokens and completed encoding."}),
+                "native_guide_scale": ("FLOAT", {"default": 1.0, "min": 0.125, "max": 1.0, "step": 0.125,
+                    "tooltip": "Native temporal mode only: spatial scale of the motion video before encoding; output size, frame count and identity references stay unchanged. Try 0.5 to reduce guide tokens/VRAM."}),
+                "native_temporal_backend": (["flash partitions (no compile)", "flex (original)", "comfy kitchen (experimental)"], {"default": "flash partitions (no compile)",
+                    "tooltip": "Native temporal mode only. Flash partitions preserves the temporal attention policy without compilation; Flex is the original implementation."}),
             },
         }
 
@@ -4241,11 +4335,12 @@ class BFSShotH3Conditioning:
     FUNCTION = "condition"
     CATEGORY = "BFS/shot loop"
     DESCRIPTION = ("Build MiniMax H3 conditioning for one shot with the native nodes: Reference to Video "
-                   "(prompt, references, length) plus the shot as an aligned guide and/or reference video.")
+                   "(prompt, references, length) plus the shot as an aligned guide and/or reference video. "
+                   "CLIP video context lets Qwen see the shot without native shot reference latents.")
 
     def condition(self, shot, clip, vae, guide_mode, first_frame, ref_image_size,
                   audio_vae=None, with_audio=False, model=None, vlm=None, setting_ref="off", setting_mask=None, inpaint="off", mask_guide="off", mask_ref_size="1/2", mask_guide_look="grey blurred",
-                  source_blur="off", source_blur_strength=1.0, source_blur_segmenter="SAM 3.1", source_blur_effect="blur"):
+                  source_blur="off", source_blur_strength=1.0, source_blur_segmenter="SAM 3.1", source_blur_effect="blur", clip_video_fps=2.0, native_guide_scale=1.0, native_temporal_backend="flash partitions (no compile)"):
         from comfy_extras.nodes_minimax_h3 import MiniMaxH3AddGuide, MiniMaxH3ReferenceToVideo
         # every picture of the shot, <Picture 1> first (chosen in the planner; use_ref_2 left in 1.113)
         refs = {f"ref_image_{k}": r for k, r in enumerate(shot_refs(shot)[:MAX_REFS])}
@@ -4264,8 +4359,28 @@ class BFSShotH3Conditioning:
         if prev is not None and shot.get("chain") == "reference":
             refs[f"ref_image_{len(refs)}"] = prev   # one more <Picture n>, after the shot's own
         audio = shot["audio"] if (with_audio and audio_vae is not None) else None
-        native = guide_mode in (self.GUIDE_MODES[1], self.GUIDE_MODES[2])
-        aligned = guide_mode in (self.GUIDE_MODES[0], self.GUIDE_MODES[2])
+        temporal = guide_mode == self.GUIDE_MODES[5]
+        if temporal:
+            if model is None:
+                raise ValueError("Native temporal alignment needs MODEL connected; route this node's MODEL output to the guider.")
+            # The Planner stores chain="head" even on shot 0, where continuity is dormant.
+            # Reject actual continuity inputs, not a hidden/default mode label.
+            index = int(shot.get("index", 0))
+            chain = shot.get("chain")
+            previous_image = index > 0 and prev is not None and chain in ("reference", "first frame")
+            history = index > 0 and chain == "history" and (
+                shot.get("history") is not None or (
+                    not shot.get("crop") and _LAST_RESULT.get("index") == index - 1
+                    and _LAST_RESULT.get("count") == int(shot.get("count", 0))))
+            if first_frame != "none" or int(shot.get("head") or 0) > 0 or previous_image or history:
+                raise ValueError("Native temporal mode requires first-frame and active shot continuity guides disabled.")
+            if mask_guide != "off":
+                raise ValueError("Native temporal mode requires mask_guide off; use only the shot as native motion video.")
+            if abs(float(shot["fps"])-24.0) > 1e-6:
+                raise ValueError("Native temporal experiment requires 24 fps shots. Set the Planner output to 24 fps.")
+        native = guide_mode in (self.GUIDE_MODES[1], self.GUIDE_MODES[2]) or temporal
+        clip_context = guide_mode == self.GUIDE_MODES[4]
+        aligned = guide_mode in (self.GUIDE_MODES[0], self.GUIDE_MODES[2]) or clip_context
         text = shot["prompt"] or ""
         setting = setting_picture(shot, setting_ref, setting_mask) if setting_ref and setting_ref != "off" else None
         if setting is not None:
@@ -4287,9 +4402,6 @@ class BFSShotH3Conditioning:
             text = add_setting(text, len(refs), "<Subject 1>" in text)
         text = fill_details_now(text, vlm, shot_refs(shot), f"[BFS Shot H3 Conditioning] shot {shot.get('index', 0) + 1}",
                                 shot.get("ref_names"))
-        if len(refs) > 9:   # H3's own node stops at 9; OpenJutsu-style LoRAs were trained with up to 9 pictures
-            print(f"[BFS Shot H3 Conditioning] shot {shot.get('index', 0) + 1}: {len(refs)} reference pictures "
-                  "(more than the 9 H3 and its LoRAs usually see)", flush=True)
         kwargs = dict(clip=clip, prompt=text, width=shot["width"], height=shot["height"],
                       length=shot["gen_length"], ref_image_size=ref_image_size, vae=vae,
                       audio_vae=audio_vae, ref_images=refs or None)
@@ -4325,6 +4437,14 @@ class BFSShotH3Conditioning:
             pmask = generation_mask(shot, log=gtag)
             guide_px = masked_only(shot["frames"], pmask, look=mask_guide_look)
             print(f"{gtag}: mask guide ({pg}, look: {mask_guide_look}), {mask_report(pmask)}", flush=True)
+        if temporal and float(native_guide_scale) != 1.0:
+            if not 0.125 <= float(native_guide_scale) <= 1.0:
+                raise ValueError("native_guide_scale must be between 0.125 and 1.0")
+            source_h, source_w = seen.shape[1:3]
+            size = (max(32, int(source_h*float(native_guide_scale))//32*32),
+                    max(32, int(source_w*float(native_guide_scale))//32*32))
+            seen = torch.nn.functional.interpolate(seen.movedim(-1,1),size=size,mode="bilinear",align_corners=False).movedim(1,-1)
+            print(f"{gtag}: native motion reference scaled {source_w}x{source_h} -> {size[1]}x{size[0]}; all {seen.shape[0]} frames retained, output unchanged",flush=True)
         if native or pg == MASK_GUIDES[3]:
             vids = {}
             if native:
@@ -4341,10 +4461,15 @@ class BFSShotH3Conditioning:
                 print(f"{gtag}: masked-region reference video at {guide_px.shape[2]}x{guide_px.shape[1]} ({mask_ref_size})",
                       flush=True)
                 vids[f"ref_video_{len(vids) + 1}"] = guide_px
-                kwargs["prompt"] = add_mask_video(kwargs["prompt"], len(vids))
-                print(f"{gtag}: the masked region is <Video {len(vids)}> (write {{mask_video}} in the prompt to place "
+                kwargs["prompt"] = add_mask_video(kwargs["prompt"], len(vids) + int(clip_context))
+                print(f"{gtag}: the masked region is <Video {len(vids) + int(clip_context)}> (write {{mask_video}} in the prompt to place "
                       f"its tag, otherwise a sentence is added)", flush=True)
             kwargs["ref_videos"] = vids
+        if clip_context:
+            from .minimax_h3_clip_video import clip_with_video_context
+            kwargs["clip"] = clip_with_video_context(clip, {"shot": shot["frames"]}, shot["gen_length"], source_fps=float(shot["fps"]), context_fps=clip_video_fps)
+            print(f"{gtag}: shot is <Video 1> in Qwen/CLIP at {clip_video_fps:g} fps; motion uses Add Guide, "
+                  "without native shot reference latents", flush=True)
         positive, latent = MiniMaxH3ReferenceToVideo.execute(**kwargs).args[:2]
         if pg in (MASK_GUIDES[1], MASK_GUIDES[2]):
             positive = MiniMaxH3AddGuide.execute(positive=positive, latent=latent, frame_idx=0, vae=vae,
@@ -4381,6 +4506,13 @@ class BFSShotH3Conditioning:
                   + f", the rest is kept ({int(nm.shape[2])} latent frames)", flush=True)
         positive = self._history(shot, positive, vae, gtag)
         latent = self._head(shot, latent, vae, gtag)
+        if temporal:
+            try:
+                from .minimax_h3_native_temporal import patch_temporal_model
+            except ImportError:
+                from minimax_h3_native_temporal import patch_temporal_model
+            model = patch_temporal_model(model, backend=native_temporal_backend)
+            print(f"{gtag}: native temporal alignment enabled; use the MODEL output for sampling", flush=True)
         return (positive, latent, model, text)
 
     @staticmethod
@@ -4441,7 +4573,10 @@ class BFSShotH3ConditioningSimple:
                 "vae": ("VAE",),
                 "guide_mode": (cls.GUIDE_MODES, {"default": cls.GUIDE_MODES[0], "tooltip":
                     "aligned guide: preserve the shot frame by frame with MiniMax H3 Add Guide. "
-                    "native reference video: provide it as <Video 1> context. both: use both paths."}),
+                    "native reference video: provide it as <Video 1> context. both: use both paths. "
+                    "aligned guide + CLIP video context: Qwen sees the shot at clip_video_fps without native "
+                    "shot reference latents; motion still uses Add Guide. Native temporal alignment: experimental native "
+                    "video mode for its matching LoRA; MODEL input and output must be connected, no Add Guide or VSA."}),
                 "ref_image_size": (["match", "max"], {"default": "match"}),
             },
             "optional": {
@@ -4464,11 +4599,20 @@ class BFSShotH3ConditioningSimple:
                 "source_blur_effect": (BLUR_EFFECTS, {"default": BLUR_EFFECTS[0], "tooltip":
                     "blur: no detail left. invert colours: keeps the shapes and the expression but not the skin tone "
                     "or the colours."}),
+                "clip_video_fps": ("FLOAT", {"default": 2.0, "min": 0.5, "max": 24.0, "step": 0.5,
+                    "tooltip": "Only for aligned guide + CLIP video context. Qwen sampling rate for this shot; "
+                    "try 2, 4 or 8 fps. Higher values cost more context/encoding time, not output fps. "
+                    "Console confirms sent frames, video vision tokens and completed encoding."}),
+                "model": ("MODEL", {"tooltip": "Required for native temporal mode. Route MODEL output to the guider."}),
+                "native_guide_scale": ("FLOAT", {"default": 1.0, "min": 0.125, "max": 1.0, "step": 0.125,
+                    "tooltip": "Native temporal mode only: spatial scale of the motion video before encoding; output size, frame count and identity references stay unchanged. Try 0.5 to reduce guide tokens/VRAM."}),
+                "native_temporal_backend": (["flash partitions (no compile)", "flex (original)", "comfy kitchen (experimental)"], {"default": "flash partitions (no compile)",
+                    "tooltip": "Native temporal mode only. Flash partitions preserves the temporal attention policy without compilation; Flex is the original implementation."}),
             },
         }
 
-    RETURN_TYPES = ("CONDITIONING", "LATENT", "STRING")
-    RETURN_NAMES = ("positive", "latent", "prompt")
+    RETURN_TYPES = ("CONDITIONING", "LATENT", "STRING", "MODEL")
+    RETURN_NAMES = ("positive", "latent", "prompt", "model")
     FUNCTION = "condition"
     CATEGORY = "BFS/shot loop"
     DESCRIPTION = ("Build the usual MiniMax H3 shot conditioning with references, prompt and an aligned guide "
@@ -4476,13 +4620,14 @@ class BFSShotH3ConditioningSimple:
 
     def condition(self, shot, clip, vae, guide_mode, ref_image_size,
                   audio_vae=None, with_audio=False, vlm=None, inpaint=INPAINT_MODES[0],
-                  source_blur="off", source_blur_strength=1.0, source_blur_segmenter="SAM 3.1", source_blur_effect="blur"):
-        positive, latent, _, prompt = BFSShotH3Conditioning().condition(
+                  source_blur="off", source_blur_strength=1.0, source_blur_segmenter="SAM 3.1", source_blur_effect="blur", clip_video_fps=2.0, model=None, native_guide_scale=1.0, native_temporal_backend="flash partitions (no compile)"):
+        positive, latent, patched_model, prompt = BFSShotH3Conditioning().condition(
             shot, clip, vae, guide_mode, "none", ref_image_size,
             audio_vae=audio_vae, with_audio=with_audio, vlm=vlm, inpaint=inpaint,
             source_blur=source_blur, source_blur_strength=source_blur_strength,
-            source_blur_segmenter=source_blur_segmenter, source_blur_effect=source_blur_effect)
-        return positive, latent, prompt
+            source_blur_segmenter=source_blur_segmenter, source_blur_effect=source_blur_effect,
+            clip_video_fps=clip_video_fps, model=model, native_guide_scale=native_guide_scale, native_temporal_backend=native_temporal_backend)
+        return positive, latent, prompt, patched_model
 
 
 class BFSShotJoin:
@@ -4702,6 +4847,7 @@ class BFSShotJoin:
                    shot.get("shot_storage") or "lossless")
         st = run_state(rid)
         i = int(shot["index"])
+        st.get("redo", {}).pop(str(i), None)
         if i not in st["done"]:
             st["done"].append(i)
         st["count"] = shot["count"]
@@ -5002,15 +5148,16 @@ class BFSShotJoin:
 
 
 def cache_status(plan: dict, reset: bool = False) -> dict:
-    """The panel's view of the shot cache of a plan's video (as of the last time the planner ran)."""
+    """Read stored-shot status. Legacy reset requests only cancel pending redos; never delete files."""
     path = _input_path(plan["video"])
     rid = run_key(plan, path)
     d = run_dir(rid, create=False)
-    if reset:
-        import shutil
-        shutil.rmtree(d, ignore_errors=True)
+    if reset and os.path.isdir(d):
+        st = run_state(rid)
+        _cancel_redo(st, list(st.get("redo", {})))
+        save_state(rid, st)
     if not os.path.isdir(d):   # nothing stored for this video (no folder is made just to look)
-        return {"run_id": rid, "folder": d, "count": 0, "done": 0, "ok": [], "stored": [], "mb": 0.0,
+        return {"run_id": rid, "folder": d, "count": 0, "done": 0, "ok": [], "stored": [], "redo": [], "mb": 0.0,
                 "temp": rid.startswith("tmp_")}
     st = run_state(rid)
     ok = shot_status(st, d) if st.get("cur") else []
@@ -5018,15 +5165,41 @@ def cache_status(plan: dict, reset: bool = False) -> dict:
     stored = sorted(int(f[5:9]) for f in files if f.startswith("shot_") and f.endswith(".pt"))
     size = sum(os.path.getsize(os.path.join(d, f)) for f in files if f.startswith("shot_"))
     return {"run_id": rid, "folder": d, "count": len(ok) or st.get("count", 0), "done": sum(ok),
-            "ok": ok, "stored": stored, "mb": round(size / 1e6, 1), "temp": rid.startswith("tmp_")}
+            "ok": ok, "stored": stored, "redo": sorted(int(i) for i in st.get("redo", {})),
+            "mb": round(size / 1e6, 1), "temp": rid.startswith("tmp_")}
 
 
-def redo_shots(plan: dict, shots: list[int], cascade: bool = True) -> dict:
+def _cancel_redo(st: dict, shots: list) -> None:
+    for i in shots:
+        key = str(int(i))
+        saved = st.get("redo", {}).pop(key, None)
+        if saved is None or saved.get("gen") != st.get("gen", {}).get(key):
+            continue  # a new take has already replaced this selection
+        if saved["done"] and int(i) not in st.setdefault("done", []):
+            st["done"].append(int(i))
+        if saved.get("keep_next"):
+            st.setdefault("keep_next", {})[key] = True
+        else:
+            st.get("keep_next", {}).pop(key, None)
+
+
+def redo_shots(plan: dict, shots: list[int], cascade: bool = True, cancel: bool = False) -> dict:
     """Mark stored shots to run again on the next Queue. cascade: the shots that continue from them are redone too
-    (automatically, once the new take exists); off: they keep their take (the join may show a seam there)."""
+    (automatically, once the new take exists); off: they keep their take (the join may show a seam there).
+    cancel restores the selection's previous state only while its stored take has not been replaced."""
     rid = run_key(plan, _input_path(plan["video"]))
     st = run_state(rid)
+    if cancel:
+        _cancel_redo(st, shots)
+        save_state(rid, st)
+        return cache_status(plan)
     for i in shots:
+        key = str(int(i))
+        st.setdefault("redo", {}).setdefault(key, {
+            "done": int(i) in st.get("done", []),
+            "gen": st.get("gen", {}).get(key),
+            "keep_next": bool(st.get("keep_next", {}).get(key)),
+        })
         if int(i) in st.get("done", []):
             st["done"].remove(int(i))
         if cascade:
@@ -5808,7 +5981,8 @@ try:
         body = await request.json()
         try:
             return web.json_response(redo_shots(_load_plan(json.dumps(body.get("plan", {}))),
-                                                [int(x) for x in body.get("shots", [])], bool(body.get("cascade", True))))
+                                                [int(x) for x in body.get("shots", [])], bool(body.get("cascade", True)),
+                                                bool(body.get("cancel", False))))
         except Exception as exc:  # noqa: BLE001
             return web.json_response({"error": f"{type(exc).__name__}: {exc}"}, status=500)
 

bfs_video_compare.py

@@ -80,6 +80,28 @@ def save_refs(reference, out_dir: str, height: int) -> list:
     return names
 
 
+def video_batches(value, name):
+    parts = value if isinstance(value, list) else [value]
+    if not parts or any(not isinstance(x, torch.Tensor) or x.ndim != 4 or
+                        x.shape[0] < 1 or x.shape[-1] < 3 or min(x.shape[1:3]) < 1 for x in parts):
+        raise ValueError(f"{name} must contain nonempty IMAGE batches")
+    return parts
+
+
+def video_chunk(parts, start, end, h, w):
+    """Read a bounded frame interval across shot batches without concatenating full videos."""
+    result, pos = [], 0
+    for part in parts:
+        count = int(part.shape[0])
+        a, b = max(0, start - pos), min(count, end - pos)
+        if a < b:
+            result.append(_scale(part[a:b, ..., :3].float(), h, w))
+        pos += count
+        if pos >= end:
+            break
+    return result[0] if len(result) == 1 else torch.cat(result, 0)
+
+
 class BFSVideoCompare:
     @classmethod
     def INPUT_TYPES(cls):
@@ -110,8 +132,7 @@ class BFSVideoCompare:
             },
         }
 
-    # lists in: the reference can be the refs list of the planner / Unpack (pictures of different sizes); the other
-    # inputs take their first item
+    # IMAGE lists contain every shot, in order. Only scalar settings take their first item.
     INPUT_IS_LIST = True
     RETURN_TYPES = ("IMAGE", "STRING", "FLOAT")
     RETURN_NAMES = ("difference", "report", "mean_difference")
@@ -127,18 +148,28 @@ class BFSVideoCompare:
     def compare(self, guide, target, fps=24.0, audio=None, gain=2.0, size="guide", preview_height=720,
                 label_guide="guide", label_target="target", reference=None):
         one = lambda v: v[0] if isinstance(v, list) and v else v
-        guide, target, fps, audio, gain, size, preview_height, label_guide, label_target = map(
-            one, (guide, target, fps, audio, gain, size, preview_height, label_guide, label_target))
+        fps, audio, gain, size, preview_height, label_guide, label_target = map(
+            one, (fps, audio, gain, size, preview_height, label_guide, label_target))
+        guide, target = video_batches(guide, "guide"), video_batches(target, "target")
+        ng, nt = sum(x.shape[0] for x in guide), sum(x.shape[0] for x in target)
+        if len(guide) > 1 or len(target) > 1:
+            if ng != nt or (len(guide) == len(target) and any(g.shape[0] != t.shape[0] for g, t in zip(guide, target))):
+                raise ValueError("Guide/target shot lists have different timing. Connect BFS Shot Join's guide and images outputs from the SAME Join; Unpack guide frames can include continuity heads and generation padding.")
         notes = []
-        n = min(guide.shape[0], target.shape[0])
-        if guide.shape[0] != target.shape[0]:
-            notes.append(f"lengths differ (guide {guide.shape[0]}, target {target.shape[0]} frames): "
+        n = min(ng, nt)
+        if ng != nt:
+            notes.append(f"lengths differ (guide {ng}, target {nt} frames): "
                          f"compared the first {n}")
         ref = guide if size == "guide" else target
-        H, W = int(ref.shape[1]), int(ref.shape[2])
-        for name, v in (("guide", guide), ("target", target)):
-            if tuple(v.shape[1:3]) != (H, W):
-                notes.append(f"{name} scaled from {v.shape[2]}x{v.shape[1]} to {W}x{H}")
+        H, W = int(ref[0].shape[1]), int(ref[0].shape[2])
+        for name, parts in (("guide", guide), ("target", target)):
+            if len(parts) > 1:
+                notes.append(f"{name}: compared all {len(parts)} shot batches in input order")
+            for dims in sorted(set(tuple(v.shape[1:3]) for v in parts)):
+                if dims != (H, W):
+                    notes.append(f"{name} scaled from {dims[1]}x{dims[0]} to {W}x{H}")
+        if ng != nt:
+            notes.append("Timing may be misaligned. For Planner results, connect guide and images from the SAME BFS Shot Join.")
         ph = min(H, int(preview_height)) if preview_height and preview_height > 0 else H
         PH, PW = max(16, ph // 2 * 2), max(16, int(round(W * ph / H)) // 2 * 2)
 
@@ -151,8 +182,8 @@ class BFSVideoCompare:
         try:
             for a in range(0, n, CHUNK):
                 e = min(a + CHUNK, n)   # the longer video stops where the shorter one ends
-                g = _scale(guide[a:e, ..., :3].float(), H, W)
-                t = _scale(target[a:e, ..., :3].float(), H, W)
+                g = video_chunk(guide, a, e, H, W)
+                t = video_chunk(target, a, e, H, W)
                 m, e, s = difference(g, t, gain)
                 maps.append(m)
                 mae.append(e)
@@ -179,7 +210,7 @@ class BFSVideoCompare:
         ui = {"bfs_compare": [{**files, "fps": float(fps), "frames": n, "width": PW, "height": PH,
                                "labels": [label_guide, label_target], "mae": [round(float(x), 2) for x in mae],
                                "refs": refs,
-                               "report": report}]}
+                               "report": report, "notes": notes}]}
         return {"ui": ui, "result": (diff_map, report, mean)}
 
 

docs/H3_FLASH_PARTITIONS.md

@@ -0,0 +1,166 @@
+# Flash Partitions for H3 native temporal alignment
+
+Flash Partitions is an inference implementation of BFS's bounded temporal attention rule. It retains the native video reference and the same temporal RoPE alignment, while evaluating the biased attention through native FlashAttention calls and stable partition merging. It requires no new learned weights and no per-score `torch.compile` kernel.
+
+**Scope:** this is an experimental backend, not a claim of universal acceleration or identity preservation. The equations are equivalent in real arithmetic; FP16/BF16 implementations are not bit-identical. Dense attention still has quadratic work in sequence length.
+
+## 1. What receives the temporal bias?
+
+H3 packs text, reference pictures, native motion-video latents, audio and output-video latents into one sequence. Only **output-video queries consulting the selected motion video's keys** receive our temporal penalty. Image-reference coordinates remain native, as do the guide's spatial coordinates. The guide's temporal RoPE origin is aligned to the output clock before attention.
+
+![Attention scope: only output queries attending to the motion video receive the temporal score penalty.](illustrations/h3_temporal_attention_scope.png)
+
+The matrix above is illustrative, not a measured attention map or actual token-count distribution. Color shows the multiplier of an **unnormalized** softmax weight, not the final attention probability. All keys remain available. Gray edges retain their original score rule; their final normalized probabilities can still change because the softmax denominator changes.
+
+For an affected query/key pair, define:
+
+```text
+dt = abs(output_time - guide_time)
+penalty(dt) = min(strength * (max(dt - window, 0) / falloff)^2, cap)
+modified_score = dot(Q, K) / sqrt(head_dim) - penalty(dt)
+```
+
+Defaults: strength 2.0, window 0.20 seconds, falloff 0.25 seconds, cap 6.0. Both videos must have matching temporal grids. This experiment uses 24 fps. This is a soft preference for corresponding times; it is not a proof of physical motion correspondence.
+
+## 2. Why the original FlexAttention path can be slower
+
+The original backend evaluates a score modifier inside compiled FlexAttention. Its rule includes time lookup, distance, a bounded quadratic penalty and edge selection. The expression is evaluated over attention score tiles across heads and blocks. It never creates a full S×S bias tensor, but the extra per-score work remains.
+
+At 116,883 sequence rows and 56 heads, the conceptual full score space contains roughly 765 billion query/key/head combinations per attention block. Kernels tile that computation rather than storing all those scores simultaneously. A new compiled shape/backend can also require initialization or compilation; native FlashAttention avoids that particular custom-kernel compilation path.
+
+The whole model additionally spends time on projections, MLPs, quantized weights, model management, preview decoding and potentially CPU/GPU offload. Large native guide sequences increase both compute and live tensor sizes. We have not measured each of those contributions in the user's full workflow, so a slow step must not automatically be labeled “compiling” or “offloading.” GPU utilization alone cannot distinguish them.
+
+## 3. Decompose the same rule into positive contributions
+
+Because `0 <= penalty <= cap`, the guide's exponential score factor can be written as:
+
+```text
+exp(-penalty) = exp(-cap) + [exp(-penalty) - exp(-cap)]
+```
+
+Both contributions are nonnegative. At distant times the penalty reaches its cap, so the bracketed local correction is exactly zero. The baseline contribution remains positive: distant guide keys are not removed.
+
+![Temporal weight curve decomposed into a small global baseline and a positive local correction.](illustrations/h3_temporal_weight_decomposition.png)
+
+With the defaults, the correction reaches zero at:
+
+```text
+window + falloff * sqrt(cap / strength)
+= 0.20 + 0.25 * sqrt(6 / 2)
+≈ 0.633 seconds
+```
+
+That threshold applies to the correction, not to the availability of guide keys.
+
+| Time difference | Penalty | Required factor | Global baseline | Local correction |
+|---|---:|---:|---:|---:|
+| 0.00 s | 0.00 | 1.000000 | 0.002479 | 0.997521 |
+| 0.40 s | 1.28 | 0.278037 | 0.002479 | 0.275559 |
+| 0.80 s | 6.00 | 0.002479 | 0.002479 | 0.000000 |
+
+The inference backend computes three types of contributions:
+
+1. All non-motion keys, with ordinary weight 1. They can occupy multiple contiguous segments.
+2. All motion-video keys, with uniform factor `exp(-cap)`.
+3. Nearby motion-frame keys, with factor `exp(-penalty) - exp(-cap)`.
+
+The last two contributions intentionally overlap in their keys. Their exponential weights add to the intended original weight, so this does not double-count the final distribution. All spatial tokens of a source frame have the same temporal correction for a given output frame. Frame batching exploits that repeated structure. It is unrelated to averaging generated RGB frames or merging outputs from separate video generations.
+
+![FlashAttention computes the contribution outputs and log-sum-exp normalizers, followed by a stable weighted merge.](illustrations/h3_flash_partitions_flow.png)
+
+## 4. Why log-sum-exp is necessary
+
+A normal attention call returns a normalized local result. Averaging those results would assign equal importance to partitions with very different softmax mass. We therefore retain each partition's natural-log normalizer, `L_i`.
+
+For partition output `O_i` and nonnegative multiplier `w_i`:
+
+```text
+adjusted_L_i = L_i + log(w_i)
+L = logsumexp(adjusted_L_1, adjusted_L_2, ...)
+O = sum_i exp(adjusted_L_i - L) * O_i
+```
+
+Zero-weight corrections can be omitted. The code uses `logaddexp` for successive merges and `expm1` to calculate small positive correction weights stably. It does not subtract two large attention outputs.
+
+Output-frame chunks default to eight latent frames. This bounds the temporary merge buffers. It does not crop the output into independently generated clips or restrict the keys to those eight frames. Queries outside the output video use ordinary full attention.
+
+Setting strength or cap to zero removes the score penalty in this backend; the separate temporal RoPE alignment remains active. Bypassing the temporal node/mode removes both changes.
+
+## 5. What was measured?
+
+### Isolated synthetic attention
+
+RTX PRO 6000 Blackwell, PyTorch 2.12.0+cu130, BF16, 107,126 tokens, 8 heads, head dimension 128:
+
+| Backend | Approximate attention call time |
+|---|---:|
+| Unmodified native PyTorch FlashAttention | 0.327 s |
+| Flash Partitions | 0.422 s |
+| Original temporal FlexAttention | 0.628 s |
+
+The training process was running concurrently. These short microbenchmarks indicate about 1.5× improvement over the original temporal backend for that tested shape; they are not a controlled full-model speed claim, do not use H3's full 56 heads and do not establish RTX 5090 performance. For a small 2,272-token shape, partition-launch/merge overhead made this backend slower than FlexAttention. There is no universally fastest choice demonstrated here.
+
+A separate small CUDA comparison against SDPA with an explicit temporal bias measured relative output norm error 0.00452 in BF16. CPU tests exercise the decomposition against the original score rule at multiple strengths/caps and with frame-chunk boundaries. This validates numerical behavior, not identity or motion quality.
+
+### User-reported full-workflow logs
+
+RTX 5090, requested output 544×928 and 360 frames:
+
+| Temporal backend | Guide scale | Guide rows | Packed rows | First complete model evaluation |
+|---|---:|---:|---:|---:|
+| Flash Partitions | 1.0 | 52,751 | 116,883 | 179.18 s |
+| Flash Partitions | 0.5 | 11,984 | 70,020 | 54.45 s |
+
+The second model evaluation at scale 0.5 was reported as 47.19 s. These are user logs, not a controlled benchmark with all settings and warm-up conditions audited. The original Flex sampler's first step was reported as 603.40 s, including initialization, so it is not directly comparable to a warmed-up attention call.
+
+The separately reported first-attention times were 2.689 s at scale 1.0 and 4.113 s at scale 0.5. That first-call measurement is not representative of every block; do not multiply it by the number of blocks to estimate the full step. Whole-model evaluation timing is the more useful end-to-end observation here.
+
+## 6. How to use it
+
+See [native temporal setup](../H3_NATIVE_TEMPORAL.md) for MODEL routing and the conditioning contract.
+
+- Standalone node: select `attention_backend = flash partitions (no compile)`.
+- Shot Conditioning: select `guide_mode = native temporal alignment (experimental)` and `native_temporal_backend = flash partitions (no compile)`.
+- Use only one temporal patch on the MODEL path. Keep Add Guide and conflicting attention/block replacements disabled.
+- `native_guide_scale = 1.0` preserves source detail. Try 0.5 for a lower-token motion reference; this keeps output dimensions, identity images and all guide frames/timestamps, but reduces guide spatial detail. H3 dimension rounding means the token reduction is not exactly fourfold.
+- For standalone conditioning, resize only the guide video before native Reference to Video to obtain that spatial reduction.
+
+The selected video is still VAE-encoded as a native reference. “Add Guide OFF” does not mean “no guide-video latents.” Qwen also sees the native video presentation. Reducing Qwen sampling alone would not remove the native DiT reference tokens.
+
+## 7. Backend requirements and limitations
+
+The current compile-free path calls PyTorch's native CUDA FlashAttention primitive with log-sum-exp outputs. It requires CUDA FP16/BF16 and a compatible PyTorch build. It is inference-only, supports one packed self-attention sequence and requires contiguous video rows, fixed spatial token count per frame and matching guide/output temporal grids. It does not use the ComfyUI attention backend selector for the modified main attention.
+
+Comfy Kitchen INT8 attention's audited public API returns the output tensor but no log-sum-exp normalizer. It is therefore not a drop-in replacement for these partition calls. Its support for masks does not by itself provide a fused temporal rule without constructing a bias representation. The separate [experimental Kitchen frame-wise backend](../H3_NATIVE_TEMPORAL.md#comfy-kitchen-and-model-attention-backend-111513), added in 1.115.13, avoids partition merging by using a broadcast key bias per target latent frame. It automatically honors the official Model Attention Backend selection. Its initial diagnostic was slower than Flash Partitions; it is not a drop-in partition kernel replacement.
+
+Flash Partitions preserves the dense score policy and its quadratic work. More speed may require reducing native guide tokens, improving kernels/merges or choosing a measured alternative. Hard windows, sparse routing or independently generated temporal chunks would change behavior and need separate quality evaluation.
+
+Alignment does not guarantee identity binding. The user has reported useful results from the existing 10500 LoRA plus this patch, and weaker identity retention from an early fresh-native checkpoint. Those are qualitative observations, not a matched-training comparison or evidence that the patch alone teaches swapping. No LoRA strength ramp, image-reference boost or identity loss is implemented by this backend.
+
+## Native one-pass Kitchen extension
+
+From 1.115.16, the optional [native Kitchen temporal extension](../H3_NATIVE_TEMPORAL.md#one-pass-native-kitchen-kernels-111516) fuses the temporal rule into dense INT8 and sol-attn kernels. It removes the per-frame attention loop. Initial full-head diagnostics were faster than Flash Partitions, while visual/end-to-end validation remains pending. From 1.115.17, first-use setup is automatic on supported Linux SM120 systems with a CUDA-13-capable driver; manual installation remains available elsewhere.
+
+## Experimental sol-attn compatibility
+
+Version 1.115.14 can recognize Model Sparse Attention with sol-attn and preserve the temporal score rule via repeated frame-wise calls. This initial compatibility path was slower than Flash Partitions; it is not the proposed fast one-pass sparse kernel. See [implementation limits and measurements](../H3_NATIVE_TEMPORAL.md#model-sparse-attention-sol-attn-compatibility-111514). SLA and VSA remain unsupported.
+
+## Video example
+
+See the [synchronized six-column video and interpretation](../H3_NATIVE_TEMPORAL.md#video-comparison-an-add-guide-trained-lora-with-native-temporal-alignment), including the step-10500 Add Guide-trained LoRA running with the new native temporal method. The separate early step-750 identity-drift report is distinguished there.
+
+## Source and reproducible figures
+
+- [Partition planning, native FlashAttention and stable merge](../minimax_h3_temporal_fast.py)
+- [MODEL wrapper, layout alignment and backend selection](../minimax_h3_native_temporal.py)
+- [Original temporal score policy / Flex backend](../minimax_h3_native_temporal_core.py)
+- [Numerical decomposition test](../tests/test_minimax_h3_temporal_fast.py)
+- [Figure source](illustrations/render_h3_flash_partitions.py), with exported SVGs beside the PNGs.
+
+Figures can be regenerated on CPU with Python, NumPy and Matplotlib:
+
+```bash
+python docs/illustrations/render_h3_flash_partitions.py
+```
+
+The illustrations describe the mechanism; they are not generated attention measurements or model-output examples.

docs/benchmarks/h3_kitchen_temporal_initial.json

@@ -0,0 +1,39 @@
+{
+  "probe": {
+    "torch": "2.12.0+cu130",
+    "kitchen": "0.2.35",
+    "status": "passed",
+    "relative_error": 0.01505568902939558,
+    "first_call_seconds": 1.416990154015366,
+    "shape": [
+      1,
+      2,
+      1248,
+      128
+    ],
+    "bias_elements": 8736,
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "peak_allocated_bytes": 29803008
+  },
+  "benchmark": {
+    "torch": "2.12.0+cu130",
+    "kitchen": "0.2.35",
+    "status": "passed",
+    "seconds": {
+      "flash_partitions": 0.017392314039170742,
+      "kitchen_frame_bias": 0.10426661401288584
+    },
+    "shape": [
+      1,
+      2,
+      42836,
+      128
+    ],
+    "frames": 22,
+    "guide_rows": 7062,
+    "bias_elements": 942392,
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "peak_allocated_bytes": 146629120,
+    "conditions": "One warmup and one timed evaluation per backend; 2 heads, not full H3; concurrent Swap Anything training; not an end-to-end speed benchmark"
+  }
+}

docs/benchmarks/h3_onepass_temporal_kernels.json

@@ -0,0 +1,137 @@
+{
+  "verification": {
+    "torch": "2.12.0+cu130",
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "cases": [
+      {
+        "dtype": "torch.bfloat16",
+        "shape": [
+          1,
+          2,
+          53,
+          128
+        ],
+        "dense_vs_fp32": 0.012292885221540928,
+        "sol_exact_vs_fp32": 0.011282283812761307,
+        "sol_onepass_vs_framewise": 0.0,
+        "finite": true
+      },
+      {
+        "dtype": "torch.bfloat16",
+        "shape": [
+          1,
+          2,
+          1253,
+          128
+        ],
+        "dense_vs_fp32": 0.015596109442412853,
+        "sol_exact_vs_fp32": 0.013514556922018528,
+        "sol_onepass_vs_framewise": 2.1619540348183364e-05,
+        "finite": true
+      },
+      {
+        "dtype": "torch.float16",
+        "shape": [
+          1,
+          2,
+          53,
+          128
+        ],
+        "dense_vs_fp32": 0.013024155050516129,
+        "sol_exact_vs_fp32": 0.011345107108354568,
+        "sol_onepass_vs_framewise": 0.0,
+        "finite": true
+      },
+      {
+        "dtype": "torch.float16",
+        "shape": [
+          1,
+          2,
+          1253,
+          128
+        ],
+        "dense_vs_fp32": 0.015380656346678734,
+        "sol_exact_vs_fp32": 0.013211732730269432,
+        "sol_onepass_vs_framewise": 5.5366094784403685e-06,
+        "finite": true
+      }
+    ],
+    "status": "passed"
+  },
+  "edge_and_bridge": {
+    "cases": [
+      {
+        "strength": 10.0,
+        "cap": 0.01,
+        "window": 0.2,
+        "dense_error": 0.015495287254452705,
+        "sol_exact_error": 0.013426524586975574
+      },
+      {
+        "strength": 0.1,
+        "cap": 30.0,
+        "window": 0.2,
+        "dense_error": 0.015566659159958363,
+        "sol_exact_error": 0.013471982441842556
+      },
+      {
+        "strength": 0.0,
+        "cap": 6.0,
+        "window": 0.0,
+        "dense_error": 0.015584140084683895,
+        "sol_exact_error": 0.01359549630433321
+      },
+      {
+        "strength": 2.0,
+        "cap": 0.0,
+        "window": 0.2,
+        "dense_error": 0.015431488864123821,
+        "sol_exact_error": 0.01330660656094551
+      }
+    ],
+    "bridge_native_calls": {
+      "dense": 1,
+      "sol": 1
+    },
+    "stream_isolation": "passed",
+    "status": "passed"
+  },
+  "benchmark": {
+    "torch": "2.12.0+cu130",
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "conditions": "Concurrent training; isolated attention; one warmup + one timed evaluation per implementation; no end-to-end speed claim",
+    "rows": [
+      {
+        "shape": [
+          1,
+          2,
+          42836,
+          128
+        ],
+        "seconds": {
+          "flash_partitions": 0.01711059600347653,
+          "kitchen_framewise": 0.10465262399520725,
+          "kitchen_onepass": 0.011890062014572322,
+          "sol_framewise": 0.15987364604370669,
+          "sol_onepass": 0.01116465002996847
+        }
+      },
+      {
+        "shape": [
+          1,
+          56,
+          42836,
+          128
+        ],
+        "seconds": {
+          "flash_partitions": 0.42317933501908556,
+          "kitchen_framewise": 0.5413047629990615,
+          "kitchen_onepass": 0.24634322000201792,
+          "sol_framewise": 2.675203786988277,
+          "sol_onepass": 0.21625870000571012
+        }
+      }
+    ],
+    "status": "passed"
+  }
+}

docs/benchmarks/h3_sol_temporal_initial.json

@@ -0,0 +1,61 @@
+{
+  "probe": {
+    "torch": "2.12.0+cu130",
+    "kitchen": "0.2.35",
+    "status": "finite_sparse_output",
+    "relative_error": 0.46068075299263,
+    "forced_all_keys_exact_relative_error": 0.013319959864020348,
+    "first_call_seconds": 1.1965116760111414,
+    "shape": [
+      1,
+      2,
+      1248,
+      128
+    ],
+    "bias_elements": 8736,
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "peak_allocated_bytes": 29803008
+  },
+  "benchmark": {
+    "torch": "2.12.0+cu130",
+    "kitchen": "0.2.35",
+    "status": "passed",
+    "seconds": {
+      "flash_partitions": 0.017417979950550944,
+      "sol_frame_bias": 0.16037049802253023
+    },
+    "shape": [
+      1,
+      2,
+      42836,
+      128
+    ],
+    "frames": 22,
+    "guide_rows": 7062,
+    "bias_elements": 942392,
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "peak_allocated_bytes": 210820096,
+    "conditions": "One warmup and one timed evaluation per backend; 2 heads, not full H3; concurrent Swap Anything training; not an end-to-end speed benchmark"
+  },
+  "benchmark_h56": {
+    "torch": "2.12.0+cu130",
+    "kitchen": "0.2.35",
+    "status": "passed",
+    "seconds": {
+      "flash_partitions": 0.41910952096804976,
+      "sol_frame_bias": 2.6489629900315776
+    },
+    "shape": [
+      1,
+      56,
+      42836,
+      128
+    ],
+    "frames": 22,
+    "guide_rows": 7062,
+    "bias_elements": 942392,
+    "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+    "peak_allocated_bytes": 5760942592,
+    "conditions": "One warmup and one timed evaluation per backend; 56 heads; isolated attention, not a full H3 model evaluation; concurrent Swap Anything training; not an end-to-end speed benchmark"
+  }
+}

docs/benchmarks/h3_temporal_auto_setup.json

@@ -0,0 +1,22 @@
+{
+  "first_use": {
+    "status": "passed",
+    "version": "h3_native_temporal_kernel_v1",
+    "shape": [
+      1,
+      2,
+      83,
+      128
+    ],
+    "dense_and_sol_finite": true
+  },
+  "fresh_process_offline_cache": {
+    "status": "passed",
+    "network_requests": 0,
+    "api_version": "h3_native_temporal_kernel_v1",
+    "api_file": "temporal_attention.py"
+  },
+  "platform": "Linux x86_64 / RTX PRO 6000 Blackwell SM120 / CUDA 13 driver",
+  "automatic_component_release": "h3-native-temporal-kernel-v1",
+  "source_commit": "9953b6caec7a5cb3e2486dd97a365f7af44eedad"
+}

docs/illustrations/render_h3_flash_partitions.py

@@ -0,0 +1,67 @@
+"""Render explanatory figures; no model inference or GPU required."""
+from pathlib import Path
+import numpy as np
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+from matplotlib.patches import Rectangle, FancyBboxPatch
+
+OUT=Path(__file__).parent
+plt.rcParams.update({'font.family':'DejaVu Sans','font.size':11,'savefig.facecolor':'white','svg.hashsalt':'h3-flash-partitions'})
+
+def save(fig,name):
+    for ext in ['png','svg']:
+        path=OUT/(name+'.'+ext)
+        fig.savefig(path,dpi=160,bbox_inches='tight',metadata={'Date':None} if ext=='svg' else None)
+        if ext=='svg':path.write_text('\n'.join(line.rstrip() for line in path.read_text().splitlines())+'\n')
+    plt.close(fig)
+
+# Conceptual groups: equal spacing within each video, not actual H3 token counts.
+fig,ax=plt.subplots(figsize=(10,7))
+n=170;rgb=np.full((n,n,3),.91)
+t=np.linspace(0,4,60);dt=np.abs(t[:,None]-t[None,:])
+p=np.minimum(2*(np.maximum(dt-.2,0)/.25)**2,6);weights=np.exp(-p)
+rgb[110:170,40:100]=plt.get_cmap('YlOrRd')(weights)[...,:3]
+ax.imshow(rgb,interpolation='nearest',origin='upper')
+bounds=[0,20,40,100,110,170];names=['Text','Pictures','Motion video','Audio','Output video']
+centers=[(a+b-1)/2 for a,b in zip(bounds,bounds[1:])]
+ax.set_xticks(centers,names);ax.set_yticks(centers,names)
+ax.xaxis.tick_top();ax.xaxis.set_label_position('top')
+ax.set_xlabel('Keys: what the model can consult',labelpad=14)
+ax.set_ylabel('Queries: what is being updated')
+for b in bounds[1:-1]:ax.axhline(b-.5,color='white',lw=2);ax.axvline(b-.5,color='white',lw=2)
+ax.add_patch(Rectangle((39.5,109.5),60,60,fill=False,edgecolor='#111827',lw=2))
+ax.set_title('The temporal bias changes one family of attention edges',pad=55,weight='bold')
+fig.text(.12,.02,'Gray: unchanged score rule. Colored block: output queries consulting the motion video.\nNear-time keys retain weight 1; distant keys retain a small positive weight. No hard mask.',fontsize=10)
+sm=plt.cm.ScalarMappable(norm=plt.Normalize(0,1),cmap='YlOrRd');fig.colorbar(sm,ax=ax,shrink=.65,label='Unnormalized weight multiplier exp(-penalty)')
+save(fig,'h3_temporal_attention_scope')
+
+fig,ax=plt.subplots(figsize=(11,5))
+x=np.linspace(0,1.1,600);penalty=np.minimum(2*(np.maximum(x-.2,0)/.25)**2,6)
+baseline=np.exp(-6);full=np.exp(-penalty);correction=full-baseline
+ax.plot(x,full,color='#172554',lw=3,label='Required weight: exp(-penalty)')
+ax.plot(x,np.full_like(x,baseline),color='#ea580c',lw=2,ls='--',label='Global guide contribution: exp(-6)')
+ax.fill_between(x,baseline,full,color='#22c55e',alpha=.35,label='Positive local correction: exp(-penalty) - exp(-6)')
+ax.axvline(.2,color='#64748b',ls=':');ax.axvline(.2+.25*np.sqrt(3),color='#64748b',ls=':')
+ax.text(.205,.91,'0.20 s\nfree window',fontsize=10)
+ax.text(.65,.5,'Beyond ~0.633 s:\ncorrection is zero,\nbaseline remains',fontsize=10)
+ax.set(xlabel='Absolute time difference between output and guide (seconds)',ylabel='Multiplier of unnormalized softmax weight',ylim=(-.03,1.08),xlim=(0,1.1))
+ax.set_title('Same temporal rule, decomposed into positive contributions',weight='bold')
+ax.legend(loc='upper right',fontsize=9);ax.grid(alpha=.18)
+fig.tight_layout();save(fig,'h3_temporal_weight_decomposition')
+
+fig,ax=plt.subplots(figsize=(14,7));ax.set(xlim=(0,14),ylim=(0,7));ax.axis('off')
+def box(x,y,w,h,text,color):
+ ax.add_patch(FancyBboxPatch((x,y),w,h,boxstyle='round,pad=0.12',facecolor=color,edgecolor='#475569',lw=1.2))
+ ax.text(x+w/2,y+h/2,text,ha='center',va='center',fontsize=10)
+def arrow(a,b):ax.annotate('',xy=b,xytext=a,arrowprops=dict(arrowstyle='->',lw=1.6,color='#475569'))
+box(.2,2.8,2,1.2,'Output-video\nquery chunk','#dbeafe')
+parts=[(5.05,'Other keys\nweight = 1','#ede9fe'),(2.85,'All motion-video keys\nweight = exp(-cap)','#ffedd5'),(.65,'Nearby motion-frame keys\nweight = exp(-penalty) - exp(-cap)','#dcfce7')]
+for y,text,color in parts:
+ box(3,y,4,1.3,text,color);arrow((2.3,3.4),(2.8,y+.65))
+ box(7.7,y,2.3,1.3,'Native FlashAttention\noutput + logsumexp',color);arrow((7.15,y+.65),(7.5,y+.65));arrow((10.15,y+.65),(10.8,3.4))
+box(11,2.65,2.7,1.5,'Stable weighted merge\n\nSame biased softmax','#dbeafe')
+ax.text(7,6.7,'Flash Partitions: reuse dense kernels instead of compiling a per-score rule',ha='center',weight='bold',fontsize=13)
+ax.text(.2,.04,'Guide contributions overlap intentionally; their weights sum to the original weight.\nQueries outside the output video use ordinary attention. Spatial tokens are retained; corrections are batched by frame.',fontsize=10)
+save(fig,'h3_flash_partitions_flow')
+print('Rendered 3 PNG and 3 SVG figures')

minimax_h3_clip_video.py

@@ -0,0 +1,159 @@
+"""H3 video context in Qwen embeddings without a native video-latent reference."""
+import math
+import time
+from dataclasses import replace
+
+from comfy_extras.nodes_minimax_h3 import MiniMaxH3ReferenceToVideo
+from comfy_api.latest import io
+
+
+class _VideoContextClip:
+    """Request-local adapter; never changes the user's CLIP or cached conditioning."""
+
+    def __init__(self, clip, items):
+        self._clip = clip
+        self._items = items
+
+    def __getattr__(self, name):
+        return getattr(self._clip, name)
+
+    def tokenize(self, text, **kwargs):
+        items = list(kwargs.get("minimax_ref_items") or [])
+        # Pictures keep their numbering. Context videos precede any explicit
+        # native video/audio references, so the shot remains <Video 1>.
+        index = 0
+        while index < len(items) and items[index].get("type") == "image":
+            index += 1
+        kwargs["minimax_ref_items"] = items[:index] + self._items + items[index:]
+        tokens = self._clip.tokenize(text, **kwargs)
+        def count_video_blocks(value):
+            if isinstance(value, dict):
+                if value.get("minimax_video_block") is True:
+                    return 1
+                return sum(count_video_blocks(v) for k, v in value.items() if k != "data")
+            if isinstance(value, (list, tuple)):
+                return sum(count_video_blocks(v) for v in value)
+            return 0
+        expected = sum((item["data"].shape[0] + 1) // 2 for item in self._items)
+        actual = count_video_blocks(tokens)
+        if actual < expected:
+            raise ValueError("CLIP did not produce the H3 video vision blocks. Use the native MiniMax H3 CLIP loader.")
+        print(f"[BFS H3 CLIP context] tokenizer confirmed {actual} video vision blocks "
+              f"(guide requires {expected}); encoding next", flush=True)
+        return tokens
+
+    def encode_from_tokens_scheduled(self, tokens, *args, **kwargs):
+        start = time.perf_counter()
+        result = self._clip.encode_from_tokens_scheduled(tokens, *args, **kwargs)
+        shapes = [tuple(entry[0].shape) for entry in result]
+        print(f"[BFS H3 CLIP context] Qwen encoding complete in {time.perf_counter() - start:.2f}s; "
+              f"conditioning shapes={shapes}. Guide native video reference blocks: 0", flush=True)
+        return result
+
+
+def clip_with_video_context(clip, videos, length, source_fps=24.0, context_fps=2.0):
+    """Native H3 24-fps frame convention, presented to Qwen at a selectable sampling rate.
+
+    Only selected frames are resized. No video VAE encoding or minimax_refs
+    video block is created here. Real image reference blocks are unaffected.
+    """
+    from comfy_extras.nodes_minimax_h3 import (
+        CANVAS_MULTIPLE, _resize, adapt_canvas, temporal_shape,
+    )
+
+    if not math.isfinite(source_fps) or source_fps <= 0:
+        raise ValueError("CLIP video context needs a positive source frame rate")
+    if not math.isfinite(context_fps) or not 0.5 <= context_fps <= 24:
+        raise ValueError("clip_video_fps must be between 0.5 and 24")
+    effective_fps = min(context_fps, source_fps)
+    frame_count = temporal_shape(length)[0]
+    items = []
+    for frames in (videos or {}).values():
+        if frames is None:
+            continue
+        if frames.ndim != 4 or frames.shape[-1] < 3:
+            raise ValueError("CLIP video context needs IMAGE frames [frames, height, width, channels]")
+        n = min(frames.shape[0], frame_count)
+        if n < 5:
+            raise ValueError("H3 CLIP video context needs at least 5 frames at the native 24 fps")
+        n -= (n - 5) % 17
+        height, width = frames.shape[1:3]
+        cw, ch = adapt_canvas(width, height)
+        if width * height < cw * ch:
+            cw = max(CANVAS_MULTIPLE, round(width / CANVAS_MULTIPLE) * CANVAS_MULTIPLE)
+            ch = max(CANVAS_MULTIPLE, round(height / CANVAS_MULTIPLE) * CANVAS_MULTIPLE)
+        indices = sorted({round(i * source_fps / effective_fps)
+                          for i in range(math.ceil(n / source_fps * effective_fps))
+                          if round(i * source_fps / effective_fps) < n})
+        selected = _resize(frames[indices], cw, ch, "disabled")
+        times = [index / source_fps for index in indices]
+        preview = times if len(times) <= 6 else times[:3] + times[-3:]
+        print(f"[BFS H3 CLIP context] guide: input={frames.shape[0]} frames at {source_fps:g} fps; "
+              f"used={n}; requested={context_fps:g} fps, effective={effective_fps:g} fps; "
+              f"sent={len(indices)} RGB frames at {cw}x{ch}; "
+              f"timestamps={[round(t, 3) for t in preview]} seconds (first/last); "
+              "native guide video reference latents: OFF", flush=True)
+        items.append({"type": "video", "data": selected,
+                      "timestamps": [index / source_fps for index in indices]})
+    return _VideoContextClip(clip, items) if items else clip
+
+
+class BFSMiniMaxH3ReferenceToVideoGuideContext(MiniMaxH3ReferenceToVideo):
+    @classmethod
+    def define_schema(cls):
+        schema = MiniMaxH3ReferenceToVideo.define_schema()
+        # Keep native image/video/audio reference semantics. The separately
+        # named guide is the only video excluded from native reference latents.
+        inputs = list(schema.inputs)
+        inputs.insert(1, io.Image.Input(
+            "guide_video_0", optional=True,
+            tooltip=("The motion guide / current shot, as native 24-fps frames. Qwen sees it at clip_video_fps "
+                     "as <Video 1>; no native guide video reference latents are added. Connect the "
+                     "same frames to Add Guide for motion. Other ref_videos start at <Video 2>."),
+        ))
+        inputs.append(io.Float.Input(
+            "clip_video_fps", default=2.0, min=0.5, max=24.0, step=0.5,
+            tooltip=("Guide 0 sampling rate for Qwen only. Default/native presentation is 2 fps; "
+                     "try 4 or 8 fps. Higher values increase context and encoding cost, not output fps. "
+                     "Never duplicates frames above the source rate. Actual reference videos stay native 2 fps."),
+        ))
+        for item in inputs:
+            if item.id == "ref_images":
+                item.template = io.Autogrow.TemplatePrefix(
+                    input=io.Image.Input("ref_image", tooltip="Reference image, in <Picture N> order"),
+                    prefix="ref_image_", min=0, max=100,
+                )
+            elif item.id == "ref_videos":
+                item.tooltip = ("Actual native video references, separate from guide_video_0. With a guide "
+                                "connected, these are <Video 2>, <Video 3>, etc. They enter both CLIP and "
+                                "native reference latents, as in the original node. Up to 3 reference videos.")
+                item.template.input.tooltip = item.tooltip
+        return replace(
+            schema, node_id="BFSMiniMaxH3ReferenceToVideoGuideContext",
+            display_name="MiniMax H3 Reference to Video — Guide 0 CLIP Context (BFS)",
+            description=("guide_video_0 enters Qwen/CLIP only as <Video 1>, without native guide video "
+                         "reference latents. Pictures and other reference videos retain native semantics. "
+                         "Use Add Guide separately for aligned motion. The target latent starts empty."),
+            inputs=inputs,
+        )
+
+    @classmethod
+    def execute(cls, clip, prompt, width, height, length, ref_image_size="match", vae=None,
+                audio_vae=None, guide_video_0=None, ref_images=None, ref_videos=None,
+                ref_video_audios=None, ref_audios=None, clip_video_fps=2.0):
+        if guide_video_0 is not None:
+            clip = clip_with_video_context(clip, {"guide_video_0": guide_video_0}, length, context_fps=clip_video_fps)
+        return MiniMaxH3ReferenceToVideo.execute(
+            clip=clip, prompt=prompt, width=width, height=height, length=length,
+            ref_image_size=ref_image_size, vae=vae, audio_vae=audio_vae,
+            ref_images=ref_images, ref_videos=ref_videos,
+            ref_video_audios=ref_video_audios, ref_audios=ref_audios,
+        )
+
+
+NODE_CLASS_MAPPINGS = {
+    "BFSMiniMaxH3ReferenceToVideoGuideContext": BFSMiniMaxH3ReferenceToVideoGuideContext,
+}
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "BFSMiniMaxH3ReferenceToVideoGuideContext": "MiniMax H3 Reference to Video — Guide 0 CLIP Context (BFS)",
+}

minimax_h3_native_temporal.py

@@ -0,0 +1,243 @@
+"""Experimental native H3 temporal inference, matching ai-toolkit native_temporal_v1.
+
+Per-model wrappers only: no global ComfyUI monkey patches or extra weights.
+"""
+import copy
+import inspect
+import math
+import time
+import torch
+try:
+    from . import minimax_h3_native_temporal_core as core
+    from . import minimax_h3_temporal_fast as fast
+    from . import minimax_h3_temporal_kitchen as kitchen
+    from . import minimax_h3_temporal_sol as sol
+    from . import minimax_h3_sparse_compat as sparse_compat
+except ImportError:
+    import minimax_h3_native_temporal_core as core
+    import minimax_h3_temporal_fast as fast
+    import minimax_h3_temporal_kitchen as kitchen
+    import minimax_h3_temporal_sol as sol
+    import minimax_h3_sparse_compat as sparse_compat
+
+MARKER = 'bfs_h3_native_temporal_v1'
+SOL_BACKEND = 'sol-attn frame bias (experimental)'
+BACKENDS = ['flash partitions (no compile)', 'flex (original)', 'comfy kitchen (experimental)']
+
+def dense_backend_selection(override):
+    if override is None:
+        return None
+    # Recognize only ComfyUI's actual ModelPatcher setter wrapper and official
+    # registered dense kernels, never arbitrary attention replacements by name.
+    if getattr(override, '__module__', '') != 'comfy.model_patcher' or not getattr(override, '__qualname__', '').endswith('set_model_optimized_attention.<locals>.optimized_attention_override'):
+        raise ValueError('Native temporal alignment conflicts with another attention replacement. Remove VSA/sparse or unknown attention patches.')
+    try:
+        from comfy.ldm.modules.attention import get_attention_function
+        selected = inspect.getclosurevars(override).nonlocals['optimized_attention']
+    except (ImportError, KeyError, TypeError) as exc:
+        raise ValueError('Unable to identify the Model Attention Backend selection.') from exc
+    for name in ('comfy_kitchen_int8', 'pytorch'):
+        if selected is get_attention_function(name, None) and selected is not None:
+            return name
+    raise ValueError('Unsupported Model Attention Backend for native temporal alignment.')
+
+
+
+def temporal_layout(native, refs, keyframes, video_index=1):
+    if keyframes:
+        raise ValueError('Native temporal mode cannot be combined with Add Guide, first-frame guides or history. Remove those guides.')
+    required = ('position_ids', 'segments', 'seq_len', 'signature')
+    if any(not hasattr(native, x) for x in required):
+        raise RuntimeError('Unsupported H3 PackedLayout API; update ComfyUI.')
+    result = copy.copy(native)
+    result.position_ids = native.position_ids.clone()
+    source = torch.zeros(native.seq_len, dtype=torch.bool)
+    target = torch.zeros_like(source)
+    ta, tb, _ = next(s for s in native.segments if s[2] == 'video')
+    target[ta:tb] = True
+    # Native packing is text, each reference's optional audio then video, targets.
+    cursor = next(s[1] for s in native.segments if s[2] == 'text')
+    videos = 0
+    for ref in refs:
+        kind = ref['kind']
+        if kind == 'image':
+            cursor += (ref['latent_h']//2)*(ref['latent_w']//2)
+        elif kind == 'audio':
+            cursor += 2*ref['ref_audio_t']
+        elif kind in ('video', 'video_audio'):
+            videos += 1
+            cursor += 2*ref.get('ref_audio_t', 0)
+            n = ref['latent_t']*(ref['latent_h']//2)*(ref['latent_w']//2)
+            if videos == video_index:
+                source[cursor:cursor+n] = True
+                result.position_ids[cursor:cursor+n, 0] += native.position_ids[ta, 0] - native.position_ids[cursor, 0]
+            cursor += n
+        else:
+            raise ValueError('Unknown native H3 reference kind: '+str(kind))
+    if not source.any():
+        raise ValueError(f'Native temporal mode requires <Video {video_index}> as a native video reference. Connect Reference to Video positive conditioning.')
+    st = torch.unique_consecutive(result.position_ids[source, 0])
+    tt = torch.unique_consecutive(result.position_ids[target, 0])
+    if st.shape != tt.shape or not torch.allclose(st, tt, atol=1e-9, rtol=0):
+        raise ValueError('Native temporal guide and output must have identical temporal grids. Use the same frame window, 24 fps and matching H3 frame count (e.g. 107). No automatic motion stretching is applied.')
+    return result, source, target
+
+
+def attention_override(ctx, previous=None, backend=BACKENDS[1], plan=None, log_first=False, sparse_patch=None, dense_backend=None):
+    first=[log_first]
+    def override(original, q, k, v, heads, mask=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+        # Token refiner and other attention paths retain their original backend.
+        n = q.shape[-2] if skip_reshape else q.shape[1]
+        if n != ctx.times.shape[1] or k.shape[-2 if skip_reshape else 1] != n:
+            if previous:
+                return previous(original, q, k, v, heads, mask=mask, skip_reshape=skip_reshape,
+                                skip_output_reshape=skip_output_reshape, **kwargs)
+            return original(q, k, v, heads, mask=mask, skip_reshape=skip_reshape,
+                            skip_output_reshape=skip_output_reshape, **kwargs)
+        if mask is not None:
+            raise ValueError('Native temporal attention does not support another attention mask override.')
+        if not skip_reshape:
+            q, k, v = [x.view(x.shape[0], x.shape[1], heads, -1).transpose(1, 2) for x in (q,k,v)]
+        chosen = backend
+        reason = None
+        if backend == SOL_BACKEND:
+            options = kwargs.get('transformer_options') or {}
+            reason = sparse_patch.dense_reason(options, n, options.get('block_index'))
+            if reason is None:
+                import comfy_kitchen as ck
+                if q.shape[-1] != 128 or not ck.sol_attn_is_available(q.device):
+                    reason = 'sol-attn requires head dimension 128 and an available CUDA kernel'
+            if reason is not None:
+                sparse_patch.log_once(('temporal_dense', n, reason), f'temporal dense fallback: {reason}')
+                chosen = dense_backend
+        timing=first[0]
+        if timing:
+            if reason is not None:
+                print(f'[BFS H3 native temporal] sol-attn dense fallback: {reason}',flush=True)
+            display_backend=chosen
+            if chosen in (SOL_BACKEND,BACKENDS[2]) and sol.native_kernel.api_for(q) is not None:
+                display_backend='sol-attn one-pass (native Kitchen)' if chosen==SOL_BACKEND else 'comfy kitchen one-pass (native)'
+            print(f'[BFS H3 native temporal] first attention starting: backend={display_backend}, Q={tuple(q.shape)}, dtype={q.dtype}',flush=True)
+            if q.is_cuda: torch.cuda.synchronize(q.device)
+            started=time.perf_counter()
+        if chosen == SOL_BACKEND:
+            if not getattr(plan, '_sol_active_logged', False):
+                passes = 1 if sol.native_kernel.api_for(q) is not None else (plan.frames+1 if ctx.strength and ctx.cap else 1)
+                print(f'[BFS H3 native temporal] sol-attn ACTIVE: tau={sparse_patch.tau}, exact conditioning keys, {passes} attention pipeline pass(es).',flush=True)
+                plan._sol_active_logged = True
+            _, sink_q = sparse_patch.sinks(options, n)
+            out = sol.attention(q,k,v,plan,ctx,tau=sparse_patch.tau,
+                                token_aug=sparse_patch.extra_tokens,sink_q=sink_q)
+        elif chosen == BACKENDS[2]:
+            out = kitchen.attention(q,k,v,plan,ctx)
+        elif chosen == BACKENDS[0]:
+            out = fast.attention(q,k,v,plan)
+        else:
+            out = core.attention(q,k,v,ctx)
+        if timing:
+            if q.is_cuda: torch.cuda.synchronize(q.device)
+            print(f'[BFS H3 native temporal] first attention finished in {time.perf_counter()-started:.3f}s',flush=True)
+            first[0]=False
+        return out if skip_output_reshape else out.transpose(1, 2).reshape(out.shape[0], -1, out.shape[1]*out.shape[3])
+    return override
+
+
+def patch_temporal_model(model, video_index=1, strength=2.0, window=0.2, falloff=0.25, cap=6.0, backend=BACKENDS[0]):
+    if backend not in BACKENDS: raise ValueError("Unknown native temporal attention backend")
+    if backend==BACKENDS[0] and not hasattr(torch.ops.aten, '_scaled_dot_product_flash_attention'):
+        raise RuntimeError("This PyTorch build lacks native FlashAttention with LSE. Update PyTorch or select flex (original).")
+    cfg = core.options(dict(native_temporal_alignment=True, native_temporal_strength=strength,
+                            native_temporal_window_seconds=window, native_temporal_falloff_seconds=falloff,
+                            native_temporal_max_penalty=cap))
+    if isinstance(video_index, bool) or not isinstance(video_index, int) or video_index < 1:
+        raise ValueError('motion_video_index must be a positive integer')
+    if model is None:
+        raise ValueError('Native temporal mode needs the MODEL input. Route this node MODEL output to the sampler/guider.')
+    if model.model_options.get(MARKER):
+        raise ValueError('Native temporal alignment is already installed. Use only one native temporal node per model path.')
+    net = model.get_model_object('diffusion_model')
+    if net.__class__.__name__ != 'MiniMaxH3Model':
+        raise ValueError('Native temporal alignment requires a native ComfyUI MiniMaxH3 model.')
+    opts = model.model_options.get('transformer_options', {})
+    dense_override, sparse_patch = sparse_compat.selection(opts)
+    selected_dense_at_setup=dense_backend_selection(dense_override)
+    sparse_compat.without_owned_blocks(opts, sparse_patch)
+    if backend==BACKENDS[2] or selected_dense_at_setup=='comfy_kitchen_int8' or sparse_patch is not None:
+        sol.native_kernel.prepare_model(model)
+    patched = model.clone()
+    previous = patched.model_options.get('model_function_wrapper')
+    cache = [None, None]
+    logged = [False]
+    plan_cache = [None, None]
+    calls = [0]
+
+    def wrapper(model_function, args):
+        conds = args.get('c', {})
+        payload = conds.get('minimax_payload') or {}
+        native = payload.get('layout')
+        if native is None:
+            raise RuntimeError('Native H3 layout missing. Update ComfyUI and use native AV latent conditioning.')
+        if cache[0] is not native:
+            cache[:] = [native, temporal_layout(native, payload.get('refs', []), payload.get('keyframes'), video_index)]
+        layout, source, target = cache[1]
+        device = args['input'][0].device if isinstance(args['input'], (list, tuple)) else args['input'].device
+        to = dict(conds.get('transformer_options', {}))
+        incoming = to.get('optimized_attention_override')
+        dense_override, sparse_patch = sparse_compat.selection(to)
+        selected_dense = dense_backend_selection(dense_override)
+        active_backend = (BACKENDS[2] if selected_dense == 'comfy_kitchen_int8' else
+                          BACKENDS[0] if selected_dense == 'pytorch' else backend)
+        dense_backend = active_backend
+        if sparse_patch is not None:
+            active_backend = SOL_BACKEND
+        to = sparse_compat.without_owned_blocks(to, sparse_patch)
+        times = layout.position_ids[:, 0].float()[None].to(device)/40.0
+        ctx = core.TemporalContext(times, source[None].to(device), target[None].to(device),
+                                  torch.ones_like(source[None], device=device), strength, window, falloff, cap,
+                                  torch.tensor([strength,window,falloff,cap],device=device,dtype=torch.float32))
+        if plan_cache[0] is not layout:
+            plan_cache[:] = [layout, fast.make_plan(layout.position_ids,source,target,strength,window,falloff,cap)]
+        to['optimized_attention_override'] = attention_override(ctx, previous=incoming, backend=active_backend, plan=plan_cache[1], log_first=not logged[0], sparse_patch=sparse_patch, dense_backend=dense_backend)
+        copied = dict(args, c=dict(conds, minimax_payload=dict(payload, layout=layout), transformer_options=to))
+        if not logged[0]:
+            print(f'[BFS H3 native temporal] {core.VERSION}: <Video {video_index}>, {int(source.sum())} motion rows; temporal RoPE aligned; soft window={window}s strength={strength} cap={cap}; images remain native; Add Guide OFF; backend={active_backend}; packed_rows={layout.seq_len}', flush=True)
+            if sparse_patch is not None:
+                print('[BFS H3 native temporal] sol-attn uses the native one-pass extension when installed; otherwise frame-wise compatibility repeats work. Watch for Kitchen ONE-PASS sol kernel ACTIVE.', flush=True)
+            logged[0] = True
+        calls[0]+=1
+        started=time.perf_counter()
+        result=previous(model_function, copied) if previous else model_function(copied['input'], copied['timestep'], **copied['c'])
+        if calls[0]<=2:
+            if device.type=='cuda':torch.cuda.synchronize(device)
+            print(f'[BFS H3 native temporal] model evaluation {calls[0]} finished in {time.perf_counter()-started:.2f}s (includes weights/offload and all blocks)',flush=True)
+        return result
+
+    patched.set_model_unet_function_wrapper(wrapper)
+    patched.model_options[MARKER] = dict(version=core.VERSION, video_index=video_index, backend=backend, **cfg)
+    return patched
+
+
+class BFSMiniMaxH3NativeTemporal:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {'required': {'model': ('MODEL',),
+            'motion_video_index': ('INT', {'default':1, 'min':1, 'max':100,
+                'tooltip':'Native <Video N> supplying motion. Picture numbering is independent.'}),
+            'strength': ('FLOAT', {'default':2.0,'min':0.0,'max':20.0,'step':0.1}),
+            'window_seconds': ('FLOAT', {'default':0.2,'min':0.0,'max':10.0,'step':0.01}),
+            'falloff_seconds': ('FLOAT', {'default':0.25,'min':0.01,'max':10.0,'step':0.01}),
+            'max_penalty': ('FLOAT', {'default':6.0,'min':0.0,'max':30.0,'step':0.1})},
+            'optional': {'attention_backend': (BACKENDS, {'default': BACKENDS[0], 'tooltip':'Flash partitions is the default. An official Model Attention Backend Kitchen selection is detected automatically. Kitchen uses an optional one-pass temporal extension when installed, otherwise frame-wise compatibility. Flex is the original implementation.'})}}
+    RETURN_TYPES = ('MODEL',)
+    FUNCTION = 'apply'
+    CATEGORY = 'BFS/MiniMax H3'
+    DESCRIPTION = ('Experimental native_temporal_v1 inference for the matching motion-transfer LoRA. '
+        'Keep the motion video in MiniMax H3 Reference to Video positive conditioning; route this MODEL '
+        'to the guider. No Add Guide or VSA. Guide/output must share a 24fps temporal grid. '
+        'Does not encode video or load the LoRA. Defaults match training.')
+
+    def apply(self, model, motion_video_index=1, strength=2.0, window_seconds=0.2, falloff_seconds=0.25, max_penalty=6.0, attention_backend=BACKENDS[0]):
+        return (patch_temporal_model(model,motion_video_index,strength,window_seconds,falloff_seconds,max_penalty,attention_backend),)
+
+NODE_CLASS_MAPPINGS = {'BFSMiniMaxH3NativeTemporal': BFSMiniMaxH3NativeTemporal}
+NODE_DISPLAY_NAME_MAPPINGS = {'BFSMiniMaxH3NativeTemporal': 'BFS MiniMax H3 Native Temporal Alignment (Experimental)'}

minimax_h3_native_temporal_core.py

@@ -0,0 +1,122 @@
+"""Native-reference temporal alignment v1. No new learned model weights.
+
+Only source channel 1's VIDEO latent rows are moved onto the target clock.
+Image-reference spatial/temporal positions remain native. Target-to-motion
+attention receives a bounded, soft temporal penalty; all other edges stay dense.
+The CUDA implementation uses FlexAttention, never a quadratic dense bias tensor.
+"""
+from dataclasses import dataclass
+import math
+import torch
+import torch.nn.functional as F
+
+VERSION = 'native_temporal_v1'
+_FLEX = None
+
+def options(kwargs):
+    enabled = kwargs.get('native_temporal_alignment', False)
+    if not isinstance(enabled, bool):raise ValueError('native_temporal_alignment must be boolean')
+    result = dict(enabled=enabled, source_channel=kwargs.get('native_motion_source_channel',1),
+                  strength=kwargs.get('native_temporal_strength',2.0),
+                  window=kwargs.get('native_temporal_window_seconds',0.2),
+                  falloff=kwargs.get('native_temporal_falloff_seconds',0.25),
+                  cap=kwargs.get('native_temporal_max_penalty',6.0))
+    if isinstance(result['source_channel'],bool) or not isinstance(result['source_channel'],int) or result['source_channel']<1:raise ValueError('Invalid native motion source channel')
+    for name in ['strength','window','falloff','cap']:
+        x=result[name]
+        if isinstance(x,bool) or not isinstance(x,(int,float)) or not math.isfinite(x) or x<0 or (name=='falloff' and x==0):raise ValueError('Invalid native temporal '+name)
+    if enabled and (kwargs.get('align_video_refs') or kwargs.get('align_image_refs') or kwargs.get('reference_source_phase') or kwargs.get('reference_rope_layout','native')!='native'):
+        raise ValueError('Native temporal mode requires native references, no AddGuide alignment or source-phase patch')
+    return result
+
+def align_layout(layout, ref_blocks, source_ids, cfg):
+    """Return a sequence-row mask, aligning only the explicit motion video's clock.
+
+    Native video latents remain in ref_blocks and keep their own spatial grid.
+    Reference audio is not temporally biased by this video-only experiment.
+    """
+    mask=torch.zeros(layout.sequence_length,dtype=torch.bool)
+    if not cfg['enabled'] or not ref_blocks:return mask
+    if len(source_ids)!=len(ref_blocks):raise ValueError('Native temporal mode requires exact reference source IDs')
+    cursor=int(layout.text_indices.numel());found=0
+    # target rows are always the final packed rows; condition video indices precede them.
+    target_start=int(layout.audio_indices[-1])+1 if layout.audio_indices.numel() else layout.sequence_length
+    # For zero audio use the count of conditioning video rows to locate the target.
+    target_idx=layout.video_indices[layout.num_condition_video_rows:]
+    target_start=int(target_idx[0]);target_origin=layout.position_ids[target_start,0].clone()
+    for block, source in zip(ref_blocks,source_ids):
+        t,h,w=block[:3];a=block[3] if len(block)>3 else 0
+        cursor+=int(a)*2
+        rows=int(t)*(int(h)//2)*(int(w)//2)
+        if source==cfg['source_channel']:
+            if t<=1:raise ValueError('Motion channel must be a video, not an image')
+            if len(block)>4 and block[4]:raise ValueError('Native motion video was accidentally converted to aligned guide')
+            found+=1;mask[cursor:cursor+rows]=True
+            layout.position_ids=layout.position_ids.clone()
+            layout.position_ids[cursor:cursor+rows,0]+=target_origin-layout.position_ids[cursor,0].clone()
+            # A common temporal extent is required; never silently stretch motion to fit.
+            src_times=torch.unique_consecutive(layout.position_ids[cursor:cursor+rows,0])
+            dst_times=torch.unique_consecutive(layout.position_ids[target_idx,0])
+            if src_times.shape!=dst_times.shape or not torch.allclose(src_times,dst_times,atol=1e-9,rtol=0):raise ValueError('Native motion and target must use identical temporal grids; resample both to matching frame windows')
+        cursor+=rows
+    if found!=1:raise ValueError(f'Expected one motion video in channel {cfg["source_channel"]}, found {found}')
+    assert not mask[target_idx].any()
+    return mask
+
+def pad_motion_masks(layouts, masks):
+    max_text=max(int(x.text_indices.numel()) for x in layouts)
+    total=max_text+layouts[0].sequence_length-int(layouts[0].text_indices.numel())
+    result=torch.zeros(len(layouts),total,dtype=torch.bool)
+    for i,(layout,mask) in enumerate(zip(layouts,masks)):
+        n=int(layout.text_indices.numel());result[i,:n]=mask[:n];result[i,max_text:]=mask[n:]
+    return result
+
+@dataclass(frozen=True)
+class TemporalContext:
+    times: torch.Tensor
+    source: torch.Tensor
+    target: torch.Tensor
+    live: torch.Tensor
+    strength: float
+    window: float
+    falloff: float
+    cap: float
+    policy: torch.Tensor | None = None
+
+def context(position_ids, source_mask, tags, video_indices, target_grid, cfg):
+    if not cfg['enabled'] or source_mask is None or not bool(source_mask.any()):return None
+    if source_mask.shape!=tags.shape:raise ValueError('Motion mask must match packed token rows')
+    target=torch.zeros_like(source_mask);n=math.prod(target_grid)
+    target[:,video_indices[-n:]]=True
+    return TemporalContext(position_ids[:,:,0].float()/40.0,source_mask,target,tags>=0,
+                           float(cfg['strength']),float(cfg['window']),float(cfg['falloff']),float(cfg['cap']),
+                           torch.tensor([cfg['strength'],cfg['window'],cfg['falloff'],cfg['cap']],device=position_ids.device,dtype=torch.float32))
+
+def score_modifier(ctx):
+    # Captured tensors are O(sequence length). No SxS matrix is materialized on CUDA.
+    times,source,target,live=ctx.times,ctx.source,ctx.target,ctx.live
+    policy=ctx.policy if ctx.policy is not None else torch.tensor([ctx.strength,ctx.window,ctx.falloff,ctx.cap],device=times.device,dtype=torch.float32)
+    def modify(score,b,h,q,k):
+        qi=q.clamp(max=times.shape[1]-1);ki=k.clamp(max=times.shape[1]-1)
+        dt=(times[b,qi]-times[b,ki]).abs()
+        penalty=torch.minimum(policy[0]*(torch.relu(dt-policy[1])/policy[2]).square(),policy[3])
+        edited=torch.where(target[b,qi]&source[b,ki],score-penalty,score)
+        return torch.where(live[b,ki],edited,float('-inf'))
+    return modify
+
+def dense_reference(q,k,v,ctx):
+    """Tiny test oracle only; forbidden at production sequence lengths."""
+    if q.shape[-2]>256:raise ValueError('Dense temporal reference is only for tiny tests')
+    delta=(ctx.times[:,:,None]-ctx.times[:,None,:]).abs()
+    penalty=(ctx.strength*(F.relu(delta-ctx.window)/ctx.falloff).square()).clamp(max=ctx.cap)
+    bias=torch.where(ctx.target[:,:,None]&ctx.source[:,None,:],-penalty,0.0)
+    bias=bias.masked_fill(~ctx.live[:,None,:],float('-inf'))
+    return F.scaled_dot_product_attention(q,k,v,attn_mask=bias[:,None].to(q.dtype))
+
+def attention(q,k,v,ctx):
+    global _FLEX
+    if q.device.type!='cuda':return dense_reference(q,k,v,ctx)
+    if _FLEX is None:
+        from torch.nn.attention.flex_attention import flex_attention
+        _FLEX=torch.compile(flex_attention,dynamic=True,fullgraph=True)
+    return _FLEX(q,k,v,score_mod=score_modifier(ctx))

minimax_h3_sparse_compat.py

@@ -0,0 +1,60 @@
+"""Recognize only the official ComfyUI sol-attn wrappers and owned block hooks."""
+import inspect
+
+
+def official_sparse_module(obj):
+    name = getattr(obj, '__module__', '')
+    # ComfyUI's file loader uses the extensionless absolute path as the
+    # module name for .py extras, rather than a dotted import name.
+    leaf = name.replace('\\', '/').rsplit('/', 1)[-1]
+    return leaf == 'nodes_sparse_attention' or leaf.endswith('.nodes_sparse_attention')
+
+
+def unwrap_sol(override):
+    if override is None or not official_sparse_module(override):
+        return override, None
+    if not getattr(override, '__qualname__', '').endswith('make_attention_override.<locals>.override'):
+        raise ValueError('Unsupported sparse attention wrapper for native temporal alignment.')
+    cells = inspect.getclosurevars(override).nonlocals
+    patch = cells.get('patch')
+    if patch is None or type(patch).__name__ != 'SparseAttnPatch' or not official_sparse_module(type(patch)):
+        raise ValueError('Unable to identify the official sparse attention patch.')
+    if patch.vsa or patch.topk_ratio != 0:
+        raise ValueError('Native temporal alignment supports only sol-attn; SLA and VSA remain incompatible.')
+    parent = cells.get('previous')
+    if parent is not None and official_sparse_module(parent):
+        raise ValueError('Use only one Model Sparse Attention node with native temporal alignment.')
+    return parent, patch
+
+
+def without_owned_blocks(options, patch):
+    blocks = options.get('patches_replace', {}).get('dit', {})
+    if not blocks:
+        return options
+    if patch is None:
+        raise ValueError('Remove VSA and other attention/block replacement patches before native temporal alignment.')
+    for callback in blocks.values():
+        if (not official_sparse_module(callback) or
+                not getattr(callback, '__qualname__', '').endswith('make_h3_block_patch.<locals>.block_patch') or
+                inspect.getclosurevars(callback).nonlocals.get('patch') is not patch):
+            raise ValueError('Native temporal alignment conflicts with another attention/block replacement patch.')
+    # A per-evaluation copy bypasses only this node's producer hooks. That native
+    # producer has no temporal Q/K bias; leaving it installed would bypass BFS.
+    return dict(options, patches_replace={**options['patches_replace'], 'dit': {}})
+
+
+def selection(options):
+    """Recover sparse state even when a later dense node replaced its override."""
+    dense, patch = unwrap_sol(options.get('optimized_attention_override'))
+    if patch is None:
+        for callback in options.get('patches_replace', {}).get('dit', {}).values():
+            if (official_sparse_module(callback) and
+                    getattr(callback, '__qualname__', '').endswith('make_h3_block_patch.<locals>.block_patch')):
+                candidate = inspect.getclosurevars(callback).nonlocals.get('patch')
+                if candidate is not None and type(candidate).__name__ == 'SparseAttnPatch' and official_sparse_module(type(candidate)):
+                    if candidate.vsa or candidate.topk_ratio != 0:
+                        raise ValueError('Native temporal alignment supports only sol-attn; SLA and VSA remain incompatible.')
+                    patch = candidate
+                    break
+    without_owned_blocks(options, patch)
+    return dense, patch

minimax_h3_temporal_bootstrap.py

@@ -0,0 +1,152 @@
+"""Pinned, node-local bootstrap for the optional native temporal extension.
+
+Only fixed trusted-release bytes are loaded, after embedded SHA256 checks.
+No pip, site-packages writes, main Kitchen replacement, or sampling-time build.
+"""
+import ctypes
+import hashlib
+import importlib.util
+import os
+from pathlib import Path
+import platform
+import sys
+import tempfile
+import threading
+import time
+import urllib.request
+
+SOURCE_COMMIT = '9953b6caec7a5cb3e2486dd97a365f7af44eedad'
+RAW = f'https://raw.githubusercontent.com/alisson-anjos/comfy-kitchen/{SOURCE_COMMIT}'
+RELEASE = 'https://github.com/alisson-anjos/comfy-kitchen/releases/download/h3-native-temporal-kernel-v1'
+BINARY_HASH = 'ad87534bfb3ece2af84c201e326b2b405f0108f34efb36937c09967c14f0329d'
+ASSETS = (
+    ('temporal_attention.py', RAW+'/comfy_kitchen/temporal_attention.py', '0151e8b5b5886c9e8b4608cbc179e9dd6f934170cf45b8d27e14801f3c4828a1', 256*1024),
+    ('backends/cuda/native_temporal/libck_temporal_sm120.so', RELEASE+'/libck_temporal_sm120.so', BINARY_HASH, 16*1024*1024),
+    ('LICENSE', RAW+'/LICENSE', 'cd1c9ea1573c9047e9b5a3676ecc1c8f755e56f249ac1ebd8b5ceda7899037da', 64*1024),
+    ('NOTICE', RAW+'/NOTICE', '35c43127ad3acbd6bc94ff63c56684a7c26eedf07795e596c8105a0476dae2ea', 64*1024),
+)
+_LOCK = threading.Lock()
+_READY = {}
+_FAILED = set()
+
+
+def cache_root():
+    explicit=os.environ.get('BFS_H3_TEMPORAL_CACHE_DIR')
+    if explicit:
+        base=Path(explicit).expanduser()
+    else:
+        try:
+            import folder_paths
+            base=Path(folder_paths.get_user_directory())/'bfs'/'h3_temporal_kernels'
+        except (ImportError,AttributeError):
+            base=Path.home()/'.cache'/'bfsnodes'/'h3_temporal_kernels'
+    return base/('v1-linux-x86_64-sm120-'+BINARY_HASH[:12])
+
+
+def compatible(device):
+    if os.environ.get('BFS_H3_TEMPORAL_AUTO_INSTALL','1').lower() in ('0','false','off'):
+        return False,'automatic setup disabled'
+    if platform.system()!='Linux' or platform.machine() not in ('x86_64','AMD64'):
+        return False,'automatic prebuilt supports Linux x86_64 only'
+    import torch
+    if getattr(torch.version,'hip',None) or not torch.cuda.is_available() or torch.cuda.get_device_capability(device)!=(12,0):
+        return False,'automatic prebuilt supports NVIDIA SM120 only'
+    # Check the actual driver, not torch.version.cuda: the sidecar uses its own
+    # static CUDA runtime and the same device/stream primary context.
+    driver=ctypes.CDLL('libcuda.so.1');version=ctypes.c_int()
+    if driver.cuDriverGetVersion(ctypes.byref(version))!=0 or version.value<13000:
+        return False,'the prebuilt requires a driver supporting CUDA 13'
+    import comfy_kitchen
+    if not comfy_kitchen.int8_attention_is_available(device):
+        return False,'Kitchen native INT8 attention unavailable'
+    from comfy_kitchen.sage_attention import prequantize_int8_attention
+    if not callable(prequantize_int8_attention):
+        return False,'Kitchen prequantization API unavailable'
+    return True,None
+
+
+def digest(path):
+    h=hashlib.sha256()
+    with path.open('rb') as f:
+        for block in iter(lambda:f.read(1024*1024),b''):h.update(block)
+    return h.hexdigest()
+
+
+def download(path,url,expected,limit):
+    if path.is_file() and digest(path)==expected:
+        return
+    path.parent.mkdir(parents=True,exist_ok=True)
+    fd,name=tempfile.mkstemp(prefix=path.name+'.',suffix='.partial',dir=path.parent)
+    temp=Path(name)
+    try:
+        request=urllib.request.Request(url,headers={'User-Agent':'BFSNodes-H3-Temporal/1'})
+        h=hashlib.sha256();total=0
+        with os.fdopen(fd,'wb') as out,urllib.request.urlopen(request,timeout=20) as response:
+            while True:
+                block=response.read(256*1024)
+                if not block:break
+                total+=len(block)
+                if total>limit:raise RuntimeError('Temporal asset exceeds the pinned size limit')
+                h.update(block);out.write(block)
+        if h.hexdigest()!=expected:
+            raise RuntimeError('Temporal asset checksum mismatch; downloaded code was not loaded')
+        os.replace(temp,path)
+    finally:
+        if temp.exists():temp.unlink()
+
+
+def provision(root):
+    import fcntl
+    root.mkdir(parents=True,exist_ok=True)
+    with (root/'.install.lock').open('a+b') as lock:
+        deadline=time.monotonic()+60
+        while True:
+            try:
+                fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB);break
+            except BlockingIOError:
+                if time.monotonic()>deadline:raise RuntimeError('Timed out waiting for temporal setup lock')
+                time.sleep(.1)
+        missing=any(not (root/rel).is_file() or digest(root/rel)!=sha for rel,_,sha,_ in ASSETS)
+        if missing:
+            print('[BFS H3 native temporal] First-use setup: downloading the verified one-pass extension (~3.3 MB). No compilation or pip install.',flush=True)
+        for rel,url,sha,limit in ASSETS:
+            download(root/rel,url,sha,limit)
+        # Verify the native binary before executing the Python entry point too.
+        binary=root/ASSETS[1][0];lib=ctypes.CDLL(str(binary))
+        lib.ck_temporal_abi_version.restype=ctypes.c_int;lib.ck_temporal_arch.restype=ctypes.c_int
+        if lib.ck_temporal_abi_version()!=1 or lib.ck_temporal_arch()!=120:
+            raise RuntimeError('Temporal prebuilt ABI/architecture mismatch')
+        name='_bfs_ck_temporal_v1_sm120_'+ASSETS[0][2][:12]+'_'+hashlib.sha256(str(root.resolve()).encode()).hexdigest()[:12]
+        if name in sys.modules:
+            return sys.modules[name]
+        spec=importlib.util.spec_from_file_location(name,root/'temporal_attention.py')
+        module=importlib.util.module_from_spec(spec);sys.modules[name]=module
+        try:spec.loader.exec_module(module)
+        except BaseException:
+            sys.modules.pop(name,None);raise
+        if getattr(module,'VERSION',None)!='h3_native_temporal_kernel_v1':
+            sys.modules.pop(name,None);raise RuntimeError('Temporal API version mismatch')
+        print(f'[BFS H3 native temporal] Verified one-pass extension ready: {root}',flush=True)
+        return module
+
+
+def ensure(device):
+    """At most one automatic setup attempt per process/cache location."""
+    key=(str(cache_root()),str(device))
+    with _LOCK:
+        if key in _READY:return _READY[key]
+        if key in _FAILED:return None
+        try:
+            ok,reason=compatible(device)
+            if not ok:
+                _FAILED.add(key)
+                print(f'[BFS H3 native temporal] Automatic native setup skipped: {reason}; retaining the existing backend.',flush=True)
+                return None
+            module=provision(Path(key[0]))
+            if not module.is_available(device):raise RuntimeError('Verified native extension is unavailable for this device')
+            _READY[key]=module
+            return module
+        except Exception as exc:
+            _FAILED.add(key)
+            print(f'[BFS H3 native temporal] Automatic native setup unavailable: {exc}. Retaining the existing backend. Restart ComfyUI to retry, or install the extension manually.',flush=True)
+            return None

minimax_h3_temporal_fast.py

@@ -0,0 +1,102 @@
+"""Exact bounded temporal bias through FlashAttention partitions; inference only.
+
+The guide's weights are exp(-cap) globally plus positive local corrections
+exp(-penalty)-exp(-cap). Other keys have weight 1. Merge partitions by logsumexp.
+No compiled score modifier, dense score matrix, or subtraction of attention sums.
+"""
+from dataclasses import dataclass
+import math
+import torch
+
+
+@dataclass
+class Plan:
+    source_start: int
+    source_stop: int
+    target_start: int
+    target_stop: int
+    source_frame_rows: int
+    target_frame_rows: int
+    frames: int
+    cap: float
+    corrections: list
+    frame_times: tuple = ()
+
+
+def make_plan(positions, source, target, strength, window, falloff, cap):
+    si=torch.where(source)[0];ti=torch.where(target)[0]
+    sa,sb=int(si[0]),int(si[-1])+1;ta,tb=int(ti[0]),int(ti[-1])+1
+    if si.numel()!=sb-sa or ti.numel()!=tb-ta:
+        raise ValueError('Fast temporal attention requires contiguous source and target video rows')
+    st,sc=torch.unique_consecutive(positions[sa:sb,0],return_counts=True)
+    tt,tc=torch.unique_consecutive(positions[ta:tb,0],return_counts=True)
+    if not torch.equal(sc,sc[:1].expand_as(sc)) or not torch.equal(tc,tc[:1].expand_as(tc)):
+        raise ValueError('Fast temporal attention requires a fixed spatial grid per video')
+    if st.shape!=tt.shape or not torch.allclose(st,tt,rtol=0,atol=1e-9):
+        raise ValueError('Fast temporal attention requires matching temporal grids')
+    frames=st.numel();corrections=[]
+    if strength>0 and cap>0:
+        times=st.double()/40
+        for offset in range(1-frames,frames):
+            lo=max(0,-offset);hi=min(frames,frames-offset)
+            dt=(times[lo:hi]-times[lo+offset:hi+offset]).abs()
+            penalty=(strength*(torch.relu(dt-window)/falloff).square()).clamp(max=cap)
+            # Stable log(exp(-penalty)-exp(-cap)); zero correction becomes -inf.
+            logs=-penalty+torch.log(-torch.expm1(penalty-cap))
+            if torch.isfinite(logs).any():corrections.append((offset,lo,hi,logs.float()))
+    return Plan(sa,sb,ta,tb,int(sc[0]),int(tc[0]),frames,cap if strength>0 else 0.,corrections,tuple((st.float()/40).tolist()))
+
+
+def flash_with_lse(q,k,v):
+    # PyTorch's native FlashAttention primitive returns natural-log normalizers.
+    if q.device.type!='cuda' or q.dtype not in (torch.float16,torch.bfloat16):
+        raise RuntimeError('Fast temporal attention requires CUDA FP16/BF16. Choose flex backend for other configurations.')
+    result=torch.ops.aten._scaled_dot_product_flash_attention.default(q,k,v,0.0,False,False)
+    return result[0],result[1]
+
+
+def merge(a,la,b,lb):
+    """Stable positive weighted merge. Outputs retain model dtype; LSE stays FP32."""
+    total=torch.logaddexp(la,lb)
+    result=a.float()*torch.exp(la-total)[...,None]+b.float()*torch.exp(lb-total)[...,None]
+    return result.to(a.dtype),total
+
+
+def attention(q,k,v,plan,flash=flash_with_lse,frame_chunk=8):
+    if q.shape[0]!=1 or q.shape[2]!=k.shape[2] or q.shape!=v.shape:
+        raise ValueError('Fast native temporal attention currently requires one packed self-attention sequence')
+    if plan.cap==0:
+        return flash(q,k,v)[0]
+    out=torch.empty_like(q)
+    sa,sb,ta,tb=plan.source_start,plan.source_stop,plan.target_start,plan.target_stop
+    # Queries outside the target retain ordinary attention, including token refs/audio.
+    for lo,hi in ((0,ta),(tb,q.shape[2])):
+        if hi>lo:out[:,:,lo:hi]=flash(q[:,:,lo:hi],k,v)[0]
+    heads,dim=q.shape[1],q.shape[-1]
+    qframes=q[:,:,ta:tb].reshape(1,heads,plan.frames,plan.target_frame_rows,dim)[0].permute(1,0,2,3)
+    kframes=k[:,:,sa:sb].reshape(1,heads,plan.frames,plan.source_frame_rows,dim)[0].permute(1,0,2,3)
+    vframes=v[:,:,sa:sb].reshape(1,heads,plan.frames,plan.source_frame_rows,dim)[0].permute(1,0,2,3)
+    if getattr(plan, '_correction_device', None) != q.device:
+        plan._device_corrections=[(off,lo,hi,logs.to(device=q.device)) for off,lo,hi,logs in plan.corrections]
+        plan._correction_device=q.device
+    corrections=plan._device_corrections
+    for first in range(0,plan.frames,frame_chunk):
+        last=min(plan.frames,first+frame_chunk)
+        start=ta+first*plan.target_frame_rows;stop=ta+last*plan.target_frame_rows
+        query=q[:,:,start:stop]
+        # Positive baseline: all motion keys downweighted by exp(-cap).
+        acc,lse=flash(query,k[:,:,sa:sb],v[:,:,sa:sb]);lse=lse-plan.cap
+        for lo,hi in ((0,sa),(sb,k.shape[2])):
+            if hi>lo:
+                part,pl=flash(query,k[:,:,lo:hi],v[:,:,lo:hi]);acc,lse=merge(acc,lse,part,pl)
+        acc=acc.reshape(1,heads,last-first,plan.target_frame_rows,dim)[0].permute(1,0,2,3)
+        lse=lse.reshape(1,heads,last-first,plan.target_frame_rows)[0].permute(1,0,2)
+        for off,lo,hi,logs in corrections:
+            a=max(first,lo);b=min(last,hi)
+            if b<=a:continue
+            part,pl=flash(qframes[a:b],kframes[a+off:b+off],vframes[a+off:b+off])
+            pl=pl+logs[a-lo:b-lo,None,None]
+            merged,ml=merge(acc[a-first:b-first],lse[a-first:b-first],part,pl)
+            acc[a-first:b-first]=merged;lse[a-first:b-first]=ml
+        out[:,:,start:stop]=acc.permute(1,0,2,3).reshape(1,heads,stop-start,dim)
+    return out

minimax_h3_temporal_kernel.py

@@ -0,0 +1,57 @@
+"""Native Kitchen extension: existing installation or verified first-use cache."""
+import importlib
+try:
+    from . import minimax_h3_temporal_bootstrap as bootstrap
+except ImportError:
+    import minimax_h3_temporal_bootstrap as bootstrap
+
+
+def api_for(q):
+    if q.device.type != 'cuda':
+        return None
+    try:
+        api = importlib.import_module('comfy_kitchen.temporal_attention')
+    except ImportError:
+        return bootstrap.ensure(q.device)
+    if getattr(api, 'VERSION', None) != 'h3_native_temporal_kernel_v1':
+        raise RuntimeError('Unsupported Comfy Kitchen temporal extension version; update the temporal overlay.')
+    return api if api.is_available(q.device) else bootstrap.ensure(q.device)
+
+
+def policy_for(api, plan, ctx):
+    key=(api.VERSION,ctx.strength,ctx.window,ctx.falloff,ctx.cap)
+    if getattr(plan,'_native_temporal_policy_key',None)!=key:
+        times = getattr(plan, 'frame_times', ())
+        if not times:
+            times = tuple(ctx.times[0, plan.target_start:plan.target_stop:plan.target_frame_rows].detach().cpu().tolist())
+        plan._native_temporal_policy = api.TemporalPolicy(
+            plan.source_start, plan.source_stop, plan.target_start, plan.target_stop,
+            plan.source_frame_rows, plan.target_frame_rows, tuple(times),
+            ctx.strength, ctx.window, ctx.falloff, ctx.cap)
+        plan._native_temporal_policy_key=key
+    return plan._native_temporal_policy
+
+
+def run(kind, q, k, v, plan, ctx, **kwargs):
+    api = api_for(q)
+    if api is None:
+        return None
+    policy = policy_for(api, plan, ctx)
+    if not getattr(plan, '_native_'+kind+'_logged', False):
+        print(f'[BFS H3 native temporal] Kitchen ONE-PASS {kind} kernel ACTIVE; temporal bias fused; no per-frame attention loop.', flush=True)
+        setattr(plan, '_native_'+kind+'_logged', True)
+    dtype=q.dtype
+    if str(dtype)=='torch.float32':
+        import torch
+        q,k,v=[t.to(torch.bfloat16) for t in (q,k,v)]
+    fn = api.int8_attention if kind == 'dense' else api.sol_attn
+    return fn(q,k,v,policy,**kwargs).to(dtype)
+
+
+def prepare_model(model):
+    """Prepare a selected native backend during node execution, before sampling."""
+    device=getattr(model,'load_device',None)
+    if getattr(device,'type',None)=='cuda':
+        from types import SimpleNamespace
+        return api_for(SimpleNamespace(device=device))
+    return None

minimax_h3_temporal_kitchen.py

@@ -0,0 +1,58 @@
+"""Prefer native one-pass Kitchen kernels, with a frame-wise fallback.
+
+The fallback avoids partition LSE.
+
+Each target frame shares a key bias across its spatial queries. The compact
+bias is O(frames * sequence), never an expanded sequence-by-sequence mask.
+INT8 quantization changes numerical results; GPU performance needs measurement.
+"""
+import inspect
+import torch
+try:
+    from . import minimax_h3_temporal_kernel as native_kernel
+except ImportError:
+    import minimax_h3_temporal_kernel as native_kernel
+
+
+def kitchen_function():
+    try:
+        import comfy_kitchen
+    except ImportError as exc:
+        raise RuntimeError('Comfy Kitchen temporal attention requires comfy-kitchen with masked INT8 attention.') from exc
+    fn = getattr(comfy_kitchen, 'int8_attention', None)
+    if fn is None or 'attn_mask' not in inspect.signature(fn).parameters:
+        raise RuntimeError('Update comfy-kitchen: this temporal backend requires int8_attention(attn_mask=...).')
+    return fn
+
+
+def attention(q, k, v, plan, ctx, kernel=None):
+    if q.shape[0] != 1 or q.shape != k.shape or q.shape != v.shape:
+        raise ValueError('Kitchen temporal attention requires one packed self-attention sequence')
+    if kernel is None:
+        native = native_kernel.run('dense',q,k,v,plan,ctx)
+        if native is not None:
+            return native
+    kernel = kitchen_function() if kernel is None else kernel
+    if ctx.strength == 0 or ctx.cap == 0:
+        return kernel(q, k, v)
+    sa, sb, ta, tb = plan.source_start, plan.source_stop, plan.target_start, plan.target_stop
+    key = (q.device, q.shape[2], float(ctx.strength), float(ctx.window), float(ctx.falloff), float(ctx.cap))
+    if getattr(plan, '_kitchen_key', None) != key:
+        frame_times = ctx.times[0, ta:tb:plan.target_frame_rows]
+        source_times = ctx.times[0, sa:sb]
+        dt = (frame_times[:, None] - source_times[None, :]).abs()
+        penalty = (ctx.strength * (torch.relu(dt-ctx.window)/ctx.falloff).square()).clamp(max=ctx.cap)
+        biases = torch.zeros((plan.frames, q.shape[2]), device=q.device, dtype=torch.float32)
+        biases[:, sa:sb] = -penalty
+        plan._kitchen_biases, plan._kitchen_key = biases, key
+    out = torch.empty_like(q)
+    for lo, hi in ((0, ta), (tb, q.shape[2])):
+        if hi > lo:
+            out[:, :, lo:hi] = kernel(q[:, :, lo:hi], k, v)
+    for frame in range(plan.frames):
+        lo = ta + frame * plan.target_frame_rows
+        hi = lo + plan.target_frame_rows
+        # Broadcast only: Kitchen expands this key bias with zero query stride.
+        mask = plan._kitchen_biases[frame].view(1, 1, 1, -1)
+        out[:, :, lo:hi] = kernel(q[:, :, lo:hi], k, v, attn_mask=mask)
+    return out

minimax_h3_temporal_sol.py

@@ -0,0 +1,62 @@
+"""Prefer the optional one-pass temporal Sol kernel; preserve legacy fallback.
+
+Without that extension, the public kernel accepts equal-length Q/K/V and a key-only bias. Therefore
+this compatibility path runs one unbiased pass and one full-sequence pass per
+latent target frame, retaining only that frame's query rows. It is deliberately
+not advertised as a fast one-pass temporal sparse kernel.
+"""
+import inspect
+import torch
+try:
+    from . import minimax_h3_temporal_kernel as native_kernel
+except ImportError:
+    import minimax_h3_temporal_kernel as native_kernel
+
+
+def sol_function():
+    try:
+        import comfy_kitchen as ck
+    except ImportError as exc:
+        raise RuntimeError('Temporal sol-attn requires Comfy Kitchen.') from exc
+    fn = getattr(ck, 'sol_attn', None)
+    if fn is None or 'key_bias' not in inspect.signature(fn).parameters:
+        raise RuntimeError('Update comfy-kitchen: temporal sol-attn requires sol_attn(key_bias=...).')
+    return fn
+
+
+def attention(q, k, v, plan, ctx, *, tau, token_aug=0, sink_q=(0, 0), kernel=None):
+    if q.shape[0] != 1 or q.shape != k.shape or q.shape != v.shape or q.shape[-1] != 128:
+        raise ValueError('Temporal sol-attn requires one packed self-attention sequence with head dimension 128.')
+    if plan.source_stop > plan.target_start:
+        raise ValueError('Temporal sol-attn requires native motion references before the target video.')
+    if kernel is None:
+        native = native_kernel.run('sol',q,k,v,plan,ctx,tau=tau,token_aug=token_aug,sink_q=sink_q)
+        if native is not None:
+            return native
+    fn = sol_function() if kernel is None else kernel
+    # Sol's key bias is applied only in the exact branch: all conditioning keys,
+    # including the selected motion video, must be forced exact regardless of
+    # the ordinary sparse node's sink_conditioning setting.
+    sink_blocks = [0, (plan.target_start+63)//64]
+    common = dict(tau=float(tau), sink_blocks=sink_blocks, topk_ratio=0., tail=True, token_aug=int(token_aug))
+    qs, ks, vs = [x.transpose(1, 2) for x in (q, k, v)]
+    if kernel is None and q.dtype == torch.float32:
+        qs, ks, vs = [x.to(torch.bfloat16) for x in (qs, ks, vs)]
+    out = fn(qs, ks, vs, sink_q=list(sink_q), **common).transpose(1, 2).clone()
+    if ctx.strength == 0 or ctx.cap == 0:
+        return out.to(q.dtype)
+    cache_key = (q.device, q.shape[2], float(ctx.strength), float(ctx.window), float(ctx.falloff), float(ctx.cap))
+    if getattr(plan, '_sol_bias_key', None) != cache_key:
+        ft = ctx.times[0, plan.target_start:plan.target_stop:plan.target_frame_rows]
+        st = ctx.times[0, plan.source_start:plan.source_stop]
+        penalty = (ctx.strength*(torch.relu((ft[:, None]-st[None, :]).abs()-ctx.window)/ctx.falloff).square()).clamp(max=ctx.cap)
+        bias = torch.zeros((plan.frames, q.shape[2]), device=q.device, dtype=torch.float32)
+        bias[:, plan.source_start:plan.source_stop] = -penalty
+        plan._sol_bias, plan._sol_bias_key = bias, cache_key
+    for frame in range(plan.frames):
+        lo = plan.target_start+frame*plan.target_frame_rows
+        hi = lo+plan.target_frame_rows
+        # Outputs for other frames are discarded: their temporal bias differs.
+        result = fn(qs, ks, vs, key_bias=plan._sol_bias[frame], sink_q=[0, 0], **common)
+        out[:, :, lo:hi] = result[:, lo:hi].transpose(1, 2)
+    return out.to(q.dtype)

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "bfsnodes"
 description = "Custom ComfyUI nodes for reference-guided video: BFS Sampler (guide-driven IC-LoRAs, auto crop, mask inpainting, temporal chunking), LTX Multiple Controls / Identity Transfer / Edit Anything, Head-Swap and Anime2Real conditioning, MiniMax-H3 latent guides, RoPE layouts and loaders, LoRA Surgery, and Shot Planner / Shot Loop (tabbed panel): split long videos into shots at camera cuts, face-tracked cast, SAM 3 mask per shot with full frame, paste, mask, crop and crop + mask generation modes, optional translucent Colored Mark overlay for manually selected regions, optional Qwen3-VL to suggest settings, describe references ({details}) and the person to replace ({target}), continuity between shots, H3 conditioning with an optional duet and prompt writer, join with the original timing, audio and a side-by-side comparison. H3 Duet: pin a source clip beside the video, training-free."
-version = "1.115.1"
+version = "1.115.18"
 license = {file = "LICENSE"}
 dependencies = [
     "torch",

tests/fixtures/comfy_sparse_node_factories.py

@@ -0,0 +1,165 @@
+# Extracted from official ComfyUI bef11fc nodes_sparse_attention.py for loader/closure compatibility tests.
+# SPDX-License-Identifier: GPL-3.0-only
+class SparseAttnPatch:
+    """Options plus runtime state for one patched model; an ON_CLEANUP callback
+    resets the state when the sampling run ends."""
+
+    def __init__(self, tau, topk_ratio, vsa, sigma_start, sigma_end, min_tokens, dense_blocks, sink_conditioning, extra_tokens, verbose):
+        self.tau = tau
+        self.topk_ratio = topk_ratio
+        self.extra_tokens = extra_tokens
+        self.vsa = vsa
+        self.sigma_start = sigma_start
+        self.sigma_end = sigma_end
+        self.min_tokens = min_tokens
+        self.dense_blocks = dense_blocks
+        self.sink_conditioning = sink_conditioning
+        self.verbose = verbose
+        self.installed = set()
+        self.reset()
+
+    def reset(self):
+        self.pooled = {}
+        self.vsa_plans = {}
+        self.vsa_rope = None
+        self._logged = set()
+
+    def log_once(self, key, message):
+        if self.verbose and key not in self._logged:
+            self._logged.add(key)
+            logging.info(f'BlockSparseAttention: {message}')
+
+    def dense_reason(self, transformer_options, tokens, block_index):
+        """Why this call stays dense regardless of its tensors, or None."""
+        sigmas = transformer_options.get('sigmas')
+        if sigmas is not None:
+            sigma = float(sigmas[0])
+            if sigma > self.sigma_start or sigma < self.sigma_end:
+                return f'sigma {sigma:.3g} outside the start/end window'
+        if tokens < self.min_tokens:
+            return f'{tokens} tokens < min_tokens {self.min_tokens}'
+        if self.dense_blocks:
+            if block_index is None:
+                self.log_once('no_block_index', 'this model does not report block indices; dense_blocks ignored')
+            elif block_index in self.dense_blocks:
+                return f'block {block_index} in dense_blocks'
+        return None
+
+    def sinks(self, transformer_options, tokens):
+        """MiniMax-H3 conditioning rows as (exact-KV blocks, dense-query blocks):
+        the packed prefix stays exact for every query, and optionally the
+        target-audio query rows run dense."""
+        layout = transformer_options.get('minimax_h3_layout')
+        if self.sink_conditioning == 'off' or layout is None or layout.seq_len != tokens:
+            return ((0, 0), (0, 0))
+        video = next(((a, b) for a, b, kind in layout.segments if kind == 'video'), None)
+        if video is None or video[0] <= 0:
+            return ((0, 0), (0, 0))
+        blocks = (0, (video[0] + BLOCK_SIZE - 1) // BLOCK_SIZE)
+        if self.sink_conditioning != 'exact_kv_and_rows':
+            return (blocks, (0, 0))
+        audio = next(((a, b) for a, b, kind in layout.segments if kind == 'audio'), None)
+        if audio is None:
+            return (blocks, blocks)
+        return (blocks, (audio[0] // BLOCK_SIZE, blocks[1]))
+
+    def vsa_plan(self, layout, device):
+        """Padded tile order: prefix segments in their own zero-padded 64-row
+        tiles, video in 4x4x4 cubes. `src` maps padded row -> source row (-1 =
+        pad), `inv` the reverse."""
+        key = (tuple(layout.signature), tuple(layout.segments), str(device))
+        plan = self.vsa_plans.get(key)
+        if plan is not None:
+            return plan
+        _text_len, latent_t, latent_h, latent_w, _audio_t = layout.signature
+        grid = (int(latent_t), int(latent_h) // 2, int(latent_w) // 2)
+        tiles, n_prefix = ([], 0)
+        for a, b, kind in layout.segments:
+            n = b - a
+            if kind != 'video':
+                m = (n + BLOCK_SIZE - 1) // BLOCK_SIZE
+                seg = torch.full((m * BLOCK_SIZE,), -1, dtype=torch.int64, device=device)
+                seg[:n] = torch.arange(a, b, device=device)
+                tiles.append(seg.view(m, BLOCK_SIZE))
+                n_prefix += m
+                continue
+            if grid[0] * grid[1] * grid[2] != n:
+                raise RuntimeError(f'VSA: video segment of {n} rows does not match the latent grid {grid}')
+            ct, ch, cw = VSA_CUBE
+            pt, ph, pw = ((g + c - 1) // c * c for g, c in zip(grid, VSA_CUBE))
+            padded = torch.full((pt, ph, pw), -1, dtype=torch.int64, device=device)
+            padded[:grid[0], :grid[1], :grid[2]] = torch.arange(a, b, device=device).view(*grid)
+            cubes = padded.view(pt // ct, ct, ph // ch, ch, pw // cw, cw).permute(0, 2, 4, 1, 3, 5).reshape(-1, BLOCK_SIZE)
+            order = torch.argsort((cubes < 0).to(torch.int8), dim=1, stable=True)
+            tiles.append(torch.gather(cubes, 1, order))
+        tiles = torch.cat(tiles)
+        src = tiles.reshape(-1)
+        live = src >= 0
+        inv = torch.empty(layout.seq_len, dtype=torch.int64, device=device)
+        inv[src[live]] = torch.nonzero(live).flatten()
+        plan = {'n': int(src.numel()), 'n_prefix': n_prefix, 'src': src, 'inv': inv, 'block_len': (tiles >= 0).sum(1).to(torch.int32)}
+        while len(self.vsa_plans) >= VSA_PLAN_CACHE:
+            del self.vsa_plans[next(iter(self.vsa_plans))]
+        self.vsa_plans[key] = plan
+        return plan
+
+    def vsa_rope_freqs(self, rope_freqs, plan):
+        hit = self.vsa_rope
+        if hit is not None and hit[0]() is rope_freqs and (hit[1] is plan):
+            return hit[2]
+        padded = rope_freqs.new_zeros((1, plan['n']) + tuple(rope_freqs.shape[2:]))
+        padded[0, plan['inv']] = rope_freqs[0]
+        self.vsa_rope = (weakref.ref(rope_freqs), plan, padded)
+        return padded
+
+def make_attention_override(patch: SparseAttnPatch, previous):
+    """Attention override; declined calls run ``previous`` (the override that was
+    on the hook before this one) or ``func``. Dense-only in VSA mode: a
+    VSA-trained model must never see plain block-sparse attention."""
+
+    def override(func, q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+        transformer_options = kwargs.get('transformer_options') or {}
+
+        def dense():
+            args = (q, k, v, heads)
+            kw = dict(mask=mask, attn_precision=attn_precision, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
+            return func(*args, **kw) if previous is None else previous(func, *args, **kw)
+        if mask is not None or patch.vsa:
+            return dense()
+        tokens = q.shape[2] if skip_reshape else q.shape[1]
+        reason = patch.dense_reason(transformer_options, tokens, transformer_options.get('block_index'))
+        if reason is not None:
+            patch.log_once(('dense', tokens, reason), f'dense ({tokens} tokens): {reason}')
+            return dense()
+        if skip_reshape:
+            b, _, _, dim_head = q.shape
+            qs, ks, vs = (t.transpose(1, 2) for t in (q, k, v))
+        else:
+            b, _, dim_head = q.shape
+            dim_head //= heads
+            qs, ks, vs = (t.view(b, -1, heads, dim_head) for t in (q, k, v))
+        reason = _ineligible(qs, ks, vs, dim_head)
+        if reason is not None:
+            patch.log_once(('ineligible', tuple(qs.shape), reason), f'dense {tuple(qs.shape)}: {reason}')
+            return dense()
+        sink, sink_q = patch.sinks(transformer_options, tokens)
+        if q.dtype == torch.float32:
+            qs, ks, vs = (t.to(torch.bfloat16) for t in (qs, ks, vs))
+        out = ck.sol_attn(qs, ks, vs, tau=patch.tau, scale=kwargs.get('scale'), sink_blocks=list(sink), sink_q=list(sink_q), topk_ratio=patch.topk_ratio, token_aug=patch.extra_tokens).to(q.dtype)
+        patch.log_once(('sparse', tuple(qs.shape)), f'sparse {tuple(qs.shape)}, sinks {sink}/{sink_q}')
+        if skip_output_reshape:
+            return out.transpose(1, 2)
+        return out.reshape(b, -1, heads * dim_head)
+    return override
+
+def make_h3_block_patch(block, block_index, patch: SparseAttnPatch):
+    """Runs the block with its attention swapped for the sparse producer."""
+
+    def attention(h, rope_freqs=None, transformer_options={}):
+        return h3_sparse_attention(block.attn, h, rope_freqs, transformer_options, patch, block_index)
+
+    def block_patch(args, extra):
+        if h3_eligible(block.attn, args['img'], args['rope_freqs'], args['transformer_options'], patch, block_index):
+            args = {**args, 'attention': attention}
+        return extra['original_block'](args)
+    return block_patch

tests/test_bfs_shot_loop.py

@@ -484,6 +484,48 @@ class ConditioningPromptTest(unittest.TestCase):
         self.assertEqual(out[2], "m")
         self.assertNotIn("duet", SL.BFSShotH3Conditioning.INPUT_TYPES()["optional"])
 
+    def test_native_temporal_requires_model_before_encoding(self):
+        shot = self._shot()
+        with self.assertRaisesRegex(ValueError, "MODEL connected"):
+            SL.BFSShotH3Conditioning().condition(shot, "clip", "vae", SL.BFSShotH3Conditioning.GUIDE_MODES[5], "none", "match")
+        self.assertEqual(self.seen, {})
+
+    def test_native_temporal_passes_shot_as_video_and_returns_patched_model(self):
+        import minimax_h3_native_temporal as temporal
+        shot = self._shot(); shot["fps"] = 24
+        with patch.object(temporal, "patch_temporal_model", return_value="patched"):
+            out = SL.BFSShotH3Conditioning().condition(shot, "clip", "vae", SL.BFSShotH3Conditioning.GUIDE_MODES[5], "none", "match", model="model")
+        self.assertEqual(out[2], "patched")
+        self.assertIs(self.seen["ref_videos"]["ref_video_1"], shot["frames"])
+
+    def test_native_temporal_allows_dormant_first_shot_chain_modes(self):
+        import minimax_h3_native_temporal as temporal
+        for mode in ("head", "history", "reference", "first frame"):
+            with self.subTest(mode=mode), patch.object(temporal, "patch_temporal_model", return_value="patched"):
+                shot = self._shot()
+                shot.update(fps=24, index=0, count=1, chain=mode, head=0, history=None, chain_image=None)
+                out = SL.BFSShotH3Conditioning().condition(shot, "clip", "vae", SL.BFSShotH3Conditioning.GUIDE_MODES[5], "none", "match", model="model")
+                self.assertEqual(out[2], "patched")
+
+    def test_native_temporal_rejects_real_continuity_inputs(self):
+        cases = [dict(index=1, chain="head", head=5),
+                 dict(index=1, chain="history", history=torch.zeros(5, 8, 8, 3)),
+                 dict(index=1, chain="first frame", chain_image=torch.zeros(1, 8, 8, 3))]
+        for fields in cases:
+            with self.subTest(fields=list(fields)):
+                shot = self._shot(); shot.update(fps=24, count=2, **fields)
+                with self.assertRaisesRegex(ValueError, "active shot continuity"):
+                    SL.BFSShotH3Conditioning().condition(shot, "clip", "vae", SL.BFSShotH3Conditioning.GUIDE_MODES[5], "none", "match", model="model")
+
+    def test_native_guide_scale_keeps_output_and_frames(self):
+        import minimax_h3_native_temporal as temporal
+        shot = self._shot(); shot.update(fps=24, frames=torch.zeros(22,64,128,3))
+        with patch.object(temporal, "patch_temporal_model", return_value="patched"):
+            SL.BFSShotH3Conditioning().condition(shot,"clip","vae",SL.BFSShotH3Conditioning.GUIDE_MODES[5],"none","match",model="model",native_guide_scale=0.5)
+        self.assertEqual(tuple(self.seen["ref_videos"]["ref_video_1"].shape),(22,32,64,3))
+        self.assertEqual((self.seen["width"],self.seen["height"],self.seen["length"]),(448,800,22))
+        self.assertEqual(tuple(shot["frames"].shape),(22,64,128,3))
+
     def test_setting_picture_is_the_last_reference_with_static(self):
         mask = torch.zeros(22, 8, 8)
         mask[:, 2:4, 2:4] = 1
@@ -505,11 +547,11 @@ class SimpleConditioningTest(unittest.TestCase):
         self.assertIn("with_audio", names)
         self.assertIn("inpaint", names)
         self.assertNotIn("duet", names)
-        self.assertNotIn("model", names)
+        self.assertIn("model", names)
         self.assertNotIn("rope_gap", names)
         self.assertNotIn("task", names)
 
-    def test_forwards_safe_advanced_defaults_and_hides_model_output(self):
+    def test_forwards_safe_advanced_defaults_and_appends_model_output(self):
         seen = {}
         original = SL.BFSShotH3Conditioning.condition
 
@@ -525,7 +567,7 @@ class SimpleConditioningTest(unittest.TestCase):
         finally:
             SL.BFSShotH3Conditioning.condition = original
 
-        self.assertEqual(out, ("positive", "latent", "prompt"))
+        self.assertEqual(out, ("positive", "latent", "prompt", "unchanged model"))
         self.assertEqual(seen["args"][4], "none")
         self.assertTrue(seen["kwargs"]["with_audio"])
         self.assertEqual(seen["kwargs"]["inpaint"], SL.INPAINT_MODES[0])
@@ -897,7 +939,7 @@ class MultiRefTest(unittest.TestCase):
         p = SL._load_plan('{"global_refs": ["a.png", "", "c.png", ""], "global_ref": "stale.png"}')
         self.assertEqual(p["global_refs"], ["a.png", "", "c.png"])
         self.assertEqual((p["global_ref"], p["global_ref2"]), ("a.png", ""))
-        self.assertEqual(len(SL.ref_list([f"{k}.png" for k in range(20)])), SL.MAX_REFS)
+        self.assertEqual(len(SL.ref_list([f"{k}.png" for k in range(20)])), min(20, SL.MAX_REFS))
 
     def test_a_shot_slot_falls_back_to_the_global_one_and_gaps_close(self):
         self.assertEqual(SL.shot_ref_names(["", "x.png"], ["a.png", "b.png", "c.png"]), ["a.png", "x.png", "c.png"])
@@ -948,10 +990,19 @@ class MultiRefTest(unittest.TestCase):
             {"picture": 5, "text": "the dog"}, {"picture": 3}]})   # a region with neither text nor points is ignored
         self.assertEqual([(x["picture"], x["color"]) for x in marks], [(2, "#9b30ff"), (5, "#ffe600")])
         text = SL.color_guide_prompt("A scene.", marks)
-        self.assertIn("marked in purple becomes the character from <Picture 2>", text)
-        self.assertIn("marked in yellow becomes the character from <Picture 5>", text)
+        self.assertIn("marked in purple uses the character appearance from <Picture 2>", text)
+        self.assertIn("marked in yellow uses the character appearance from <Picture 5>", text)
         one = SL.color_guide_prompt("", marks[:1])
-        self.assertIn("Replace only the person marked in purple with the character from <Picture 2>", one)
+        self.assertIn("the subject marked in purple uses the character appearance from <Picture 2>", one)
+
+    def test_color_contract_is_present_even_when_reference_appearance_mentions_the_color(self):
+        seg = {"color_guide": True, "color_guide_marks": [{"picture": 3, "color": "#00ff00"}]}
+        raw = "Replace the man with the green goblin from <Picture 3>. Render the scene as anime on a rooftop."
+        text = SL.tidy_caption(raw, seg)
+        self.assertIn("the subject marked in green uses the character appearance from <Picture 3>", text)
+        self.assertIn("anime on a rooftop", text)
+        self.assertNotIn("all unmarked content unchanged", text)
+        self.assertEqual(SL.color_guide_prompt(text, seg["color_guide_marks"]), text)
 
     def test_auto_prompt_asks_with_the_official_format(self):
         seen = {}
@@ -970,11 +1021,11 @@ class MultiRefTest(unittest.TestCase):
             SL.vlm_generate = old
         self.assertEqual(seen["n"], 5)                       # 3 frames + 2 pictures
         for want in ("swap the man for the robot", "<Picture 2>: image 5, described as: a silver robot",
-                     "tinted purple becomes the subject from <Picture 2>", "Only the selected subject is generated again",
+                     "marked in purple follows the appearance from <Picture 2>", "Only the selected subject is generated again",
                      "edited version of <Video 1>", "retention_analysis:", "non_diegetic_music:", "00:00.500",
                      "Never write a timestamp inside a shot", "is replaced by <Subject N>"):
             self.assertIn(want, seen["prompt"])
-        self.assertTrue(out.startswith("subject_definitions:\n<Subject 1>"))   # fences and bold removed
+        self.assertTrue(out.startswith("motion_transfer: subject_definitions:\n<Subject 1>"))   # fences and bold removed
         self.assertIn("marked in purple", out)                # the mapping the writer left out is added
         seg2 = dict(seg, prompt="")
         self.assertEqual(SL.auto_prompt_for("x", ["a.png"], seg2), SL.auto_prompt_for("x", ["a.png"], dict(seg2)))
@@ -1062,8 +1113,10 @@ class MultiRefTest(unittest.TestCase):
             SL.vlm_generate = old
         self.assertIn("short style of a video-edit request", seen["prompt"])
         self.assertNotIn("full-reference format", seen["prompt"])
+        self.assertNotIn("anime", seen["prompt"].lower())  # no sample style may leak into a request with no style edit
+        self.assertNotIn("Example request:", seen["prompt"])
         self.assertEqual(seen["tokens"], 512)
-        self.assertTrue(out.startswith("Replace the dog in the video with the gaunt goblin from <Picture 1>."))
+        self.assertTrue(out.startswith("motion_transfer: Replace the dog in the video with the gaunt goblin from <Picture 1>."))
         self.assertEqual(SL.caption_problems(out, 2), [])
         self.assertTrue(SL.caption_problems("subject_definitions: <Subject 1> is x", 1))
 
@@ -1575,3 +1628,127 @@ class FillDetailsTest(unittest.TestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class RedoRecoveryTest(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.tmp.cleanup)
+        self.directory = Path(self.tmp.name)
+        self.plan = {"video": "test.mp4"}
+        self.state = {"done": [0, 1], "count": 2, "sig": {"0": "a", "1": "b"},
+                      "gen": {"0": "old", "1": "next"}, "from": {"1": "old"},
+                      "cur": {"sigs": ["a", "b"], "prev": [False, True]}}
+        for i in range(2):
+            (self.directory / f"shot_{i:04d}.pt").write_bytes(b"saved take")
+        for name, value in (("_input_path", lambda _: "video"), ("run_key", lambda *_: "test"),
+                            ("run_dir", lambda *a, **kw: str(self.directory))):
+            patcher = patch.object(SL, name, value)
+            patcher.start()
+            self.addCleanup(patcher.stop)
+        SL.save_state("test", self.state)
+
+    def test_redo_can_be_cancelled_with_continuity_restored(self):
+        self.assertEqual(SL.redo_shots(self.plan, [0])["ok"], [False, False])
+        result = SL.redo_shots(self.plan, [0], cancel=True)
+        self.assertEqual(result["ok"], [True, True])
+        self.assertEqual(result["redo"], [])
+
+    def test_reset_preserves_files_and_state(self):
+        SL.redo_shots(self.plan, [0], cascade=False)
+        result = SL.cache_status(self.plan, reset=True)
+        self.assertEqual(result["ok"], [True, True])
+        self.assertEqual((self.directory / "shot_0000.pt").read_bytes(), b"saved take")
+        self.assertNotIn("0", SL.run_state("test").get("keep_next", {}))
+        self.assertEqual(SL.run_state("test")["gen"], self.state["gen"])
+
+    def test_repeated_selection_does_not_overwrite_undo(self):
+        SL.redo_shots(self.plan, [0])
+        SL.redo_shots(self.plan, [0])
+        self.assertEqual(SL.redo_shots(self.plan, [0], cancel=True)["ok"], [True, True])
+
+    def test_cancel_does_not_make_changed_settings_valid(self):
+        SL.redo_shots(self.plan, [0])
+        st = SL.run_state("test")
+        st["cur"]["sigs"][0] = "changed seed"
+        SL.save_state("test", st)
+        self.assertEqual(SL.redo_shots(self.plan, [0], cancel=True)["ok"], [False, False])
+
+    def test_cancel_does_not_restore_replaced_take(self):
+        SL.redo_shots(self.plan, [0])
+        st = SL.run_state("test")
+        st["gen"]["0"] = "new"
+        SL.save_state("test", st)
+        self.assertEqual(SL.redo_shots(self.plan, [0], cancel=True)["ok"], [False, False])
+
+    def test_reset_without_selection_preserves_completed_run(self):
+        self.assertEqual(SL.cache_status(self.plan, reset=True)["ok"], [True, True])
+
+
+class IsolatedReferenceBindingTests(unittest.TestCase):
+    def test_each_reference_has_its_own_image_and_index(self):
+        images = [torch.zeros(1, 4, 4, 3), torch.ones(1, 4, 4, 3)]
+        calls = []
+        def describe(clip, batch, instruction, tokens, **kwargs):
+            self.assertEqual(len(batch), 1)
+            calls.append((batch[0], instruction))
+            return "human in dark suit" if batch[0] is images[0] else "goblin in loincloth"
+        with patch.object(SL, "vlm_describe", side_effect=describe):
+            result = SL.isolated_picture_descriptions(object(), images)
+        self.assertEqual(result, ["human in dark suit", "goblin in loincloth"])
+        self.assertIn("ONLY <Picture 1>", calls[0][1])
+        self.assertIn("ONLY <Picture 2>", calls[1][1])
+        self.assertIs(calls[0][0], images[0])
+        self.assertIs(calls[1][0], images[1])
+
+    def test_final_writer_only_receives_separate_text_facts(self):
+        captured = []
+        def generate(clip, prompt, images, tokens, **kwargs):
+            captured.append((prompt, images))
+            return "Replace the woman with the human from <Picture 1> and the man with the goblin from <Picture 2>."
+        with patch.object(SL, "vlm_generate", side_effect=generate):
+            SL.write_h3_prompt(object(), [], [], [object(), object()],
+                ["human in dark suit", "pink goblin"], "Replace both fighters", {},
+                style="captions", source_facts="A woman and a man are fighting.")
+        self.assertIsNone(captured[0][1])
+        self.assertIn("<Picture 1>: human in dark suit", captured[0][0])
+        self.assertIn("<Picture 2>: pink goblin", captured[0][0])
+        self.assertIn("A woman and a man are fighting", captured[0][0])
+
+    def test_browser_and_backend_use_same_prompt_revision(self):
+        browser = (ROOT / "web/js/bfs_shot_loop.js").read_text()
+        start = browser.index("const autoFor =")
+        signature = browser[start:browser.index("const writtenOf", start)]
+        self.assertIn('"' + SL.AUTO_PROMPT_REVISION + '"', signature)
+
+    def test_motion_trigger_is_added_once_and_can_be_disabled(self):
+        self.assertEqual(SL.tidy_caption("Replace the actor.", trigger="motion_transfer"), "motion_transfer: Replace the actor.")
+        self.assertEqual(SL.tidy_caption("motion_transfer: Replace the actor.", trigger="motion_transfer"), "motion_transfer: Replace the actor.")
+        self.assertEqual(SL.tidy_caption("motion_transfer: Replace the actor.", trigger=""), "Replace the actor.")
+        self.assertNotEqual(SL.auto_prompt_for("edit", [], {}, prompt_trigger=""), SL.auto_prompt_for("edit", [], {}, prompt_trigger="motion_transfer"))
+
+
+class PromptReferenceContractTest(unittest.TestCase):
+    def test_short_caption_does_not_acquire_phantom_second_subject(self):
+        text = "motion_transfer: Replace the dancer with the character from <Picture 1>. Use the requested setting."
+        self.assertEqual(SL.expand_two_ref_prompt(text), text)
+
+    def test_legacy_expansion_requires_explicit_description_template(self):
+        text = "subject_definitions: <Subject 1> is the {details} in <Picture 1>."
+        self.assertIn("<Subject 2>", SL.expand_two_ref_prompt(text))
+        finished = "subject_definitions: <Subject 1> is the fox in <Picture 1>."
+        self.assertEqual(SL.expand_two_ref_prompt(finished), finished)
+
+    def test_invalid_picture_indices_are_reported_in_both_formats(self):
+        text = "Replace the setting with <Picture 3>."
+        self.assertTrue(SL.caption_problems(text, 2))
+        self.assertTrue(SL.h3_prompt_problems(text, 2, template=True))
+        self.assertEqual(SL.picture_index_problems(text, 3), [])
+
+    def test_writer_refuses_nonexistent_reference_after_retry(self):
+        with patch.object(SL, "vlm_generate", return_value="Replace the setting with <Picture 3>.") as generator:
+            with self.assertRaisesRegex(ValueError, "Picture 3.*does not exist"):
+                SL.write_h3_prompt(object(), [], [], [object(), object()],
+                    ["fox wearing blue pants", "city street"], "Use Picture 2 for the setting", {},
+                    style="captions", source_facts="A dancer is moving in a street.")
+        self.assertEqual(generator.call_count, 2)

tests/test_bfs_video_compare.py

@@ -6,6 +6,7 @@ import sys
 import tempfile
 import types
 import unittest
+from unittest.mock import patch
 from pathlib import Path
 
 import torch
@@ -65,6 +66,47 @@ class CompareNodeTest(unittest.TestCase):
                 n += 1
             self.assertEqual(n, 18, k)
 
+    def test_all_shots_keep_guide_and_target_order_across_chunk_boundaries(self):
+        class Writer:
+            videos = {}
+            def __init__(self, path, *args, **kwargs):
+                self.frames = self.videos.setdefault(Path(path).name, [])
+            def write(self, frames):
+                self.frames.append(frames.clone())
+            def close(self):
+                pass
+        # Scene 1 red, scene 2 green: 22 frames span the 16-frame encoder chunk.
+        guide = [torch.zeros(10, 32, 48, 3), torch.zeros(12, 32, 48, 3)]
+        guide[0][..., 0] = 1
+        guide[1][..., 1] = 1
+        target = [x.clone() for x in guide]
+        with patch.object(VC.VIO, "Mp4Writer", Writer):
+            out = VC.BFSVideoCompare().compare(guide, target, preview_height=32)
+        self.assertEqual(out["ui"]["bfs_compare"][0]["frames"], 22)
+        self.assertEqual(out["result"][2], 0)
+        for name in ("guide.mp4", "target.mp4"):
+            frames = torch.cat(Writer.videos[name])
+            self.assertEqual(frames.shape[0], 22)
+            self.assertEqual(frames[:10, 0, 0].tolist(), [[1., 0., 0.]] * 10)
+            self.assertEqual(frames[10:, 0, 0].tolist(), [[0., 1., 0.]] * 12)
+
+    def test_joined_video_can_match_multiple_equal_timing_shots(self):
+        g = [torch.zeros(3, 32, 48, 3), torch.ones(4, 32, 48, 3)]
+        self.assertEqual(VC.video_chunk(g, 2, 6, 32, 48)[:, 0, 0, 0].tolist(), [0., 1., 1., 1.])
+        out = VC.BFSVideoCompare().compare(g, torch.cat(g), preview_height=32)
+        self.assertEqual(out["ui"]["bfs_compare"][0]["frames"], 7)
+        self.assertEqual(out["result"][2], 0)
+
+    def test_misaligned_shot_lists_require_join_timing_instead_of_silent_truncation(self):
+        g = [torch.zeros(5, 32, 48, 3), torch.zeros(7, 32, 48, 3)]
+        t = [torch.zeros(6, 32, 48, 3), torch.zeros(6, 32, 48, 3)]
+        with self.assertRaisesRegex(ValueError, "SAME Join"):
+            VC.BFSVideoCompare().compare(g, t)
+        with self.assertRaisesRegex(ValueError, "SAME Join"):
+            VC.BFSVideoCompare().compare(g, torch.zeros(8, 32, 48, 3))
+        with self.assertRaisesRegex(ValueError, "nonempty"):
+            VC.BFSVideoCompare().compare([], t)
+
     def test_reference_pictures_go_to_the_player_no_taller_than_it(self):
         from PIL import Image
         guide, target = torch.rand(4, 32, 48, 3), torch.rand(4, 32, 48, 3)

tests/test_minimax_h3_native_temporal.py

@@ -0,0 +1,194 @@
+import copy
+import importlib.util
+import json
+from pathlib import Path
+import sys
+from types import SimpleNamespace
+import pytest
+import torch
+
+ROOT=Path(__file__).resolve().parents[1]
+sys.path.insert(0,str(ROOT))
+import minimax_h3_native_temporal as node
+
+def load(name,path):
+ s=importlib.util.spec_from_file_location(name,path);m=importlib.util.module_from_spec(s);sys.modules[name]=m;s.loader.exec_module(m);return m
+CORE=load('native_temporal_comfy_fixture',ROOT/'tests/fixtures/comfy_h3_layout_8cfe5e1.py')
+
+def example(vt=7):
+ refs=[dict(kind='image',latent_h=4,latent_w=8),dict(kind='video_audio',latent_h=8,latent_w=4,latent_t=vt,ref_audio_t=3)]
+ l=CORE.PackedLayout(3,7,8,8,2,refs=refs)
+ return l,refs
+
+def test_layout_matches_actual_training_coordinates():
+ l,refs=example();before=l.position_ids.clone()
+ aligned,src,tgt=node.temporal_layout(l,refs,None)
+ golden=json.loads((ROOT/'tests/fixtures/h3_native_temporal_training_positions.json').read_text())
+ torch.testing.assert_close(aligned.position_ids,torch.tensor(golden['position_ids'],dtype=torch.float64),rtol=0,atol=1e-12)
+ assert src.tolist()==golden['source']
+ assert torch.equal(l.position_ids,before)
+ assert torch.equal(aligned.position_ids[~src],before[~src])
+ assert torch.equal(aligned.position_ids[src,1:],before[src,1:])
+ assert not (src&tgt).any()
+
+def test_wrong_video_index_and_temporal_mismatch_rejected():
+ l,r=example()
+ with pytest.raises(ValueError,match='Video 2'):node.temporal_layout(l,r,None,2)
+ l,r=example(6)
+ with pytest.raises(ValueError,match='identical temporal'):node.temporal_layout(l,r,None)
+ with pytest.raises(ValueError,match='Add Guide'):node.temporal_layout(l,r,[{}])
+
+def context():
+ l,r=example();a,s,t=node.temporal_layout(l,r,None)
+ return node.core.TemporalContext(a.position_ids[:,0].float()[None]/40,s[None],t[None],torch.ones_like(s[None]),2,.2,.25,6)
+
+def test_attention_adapter_matches_dense_score_policy_and_shapes():
+ ctx=context();torch.manual_seed(12)
+ q,k,v=[torch.randn(1,2,ctx.times.shape[1],8) for _ in range(3)]
+ def forbidden(*args,**kwargs):raise AssertionError('wrong backend')
+ out=node.attention_override(ctx)(forbidden,q,k,v,2,skip_reshape=True)
+ expected=node.core.dense_reference(q,k,v,ctx).transpose(1,2).flatten(2)
+ torch.testing.assert_close(out,expected)
+ raw=[x.transpose(1,2).flatten(2) for x in (q,k,v)]
+ torch.testing.assert_close(node.attention_override(ctx)(forbidden,*raw,2),expected)
+
+def test_token_refiner_uses_existing_attention():
+ ctx=context();q=torch.randn(1,2,3,8)
+ def original(*args,**kwargs):return 'original'
+ assert node.attention_override(ctx)(original,q,q,q,2,skip_reshape=True)=='original'
+
+class Model:
+ def __init__(self):self.model_options={}
+ def clone(self):m=Model();m.model_options=copy.deepcopy(self.model_options);return m
+ def get_model_object(self,key):return type('MiniMaxH3Model',(),{})()
+ def set_model_unet_function_wrapper(self,w):self.model_options['model_function_wrapper']=w
+
+def test_per_model_wrapper_clones_payload_and_applies_attention():
+ m=Model();patched=node.patch_temporal_model(m);assert not m.model_options
+ l,r=example();payload=dict(layout=l,refs=r)
+ def fn(x,t,**c):
+  assert c['minimax_payload']['layout'] is not l
+  assert 'optimized_attention_override' in c['transformer_options']
+  return 'ok'
+ args=dict(input=torch.zeros(1),timestep=torch.ones(1),c=dict(minimax_payload=payload))
+ assert patched.model_options['model_function_wrapper'](fn,args)=='ok'
+ assert payload['layout'] is l and 'transformer_options' not in args['c']
+ with pytest.raises(ValueError,match='already installed'):node.patch_temporal_model(patched)
+
+def test_incompatible_attention_rejected():
+ m=Model();m.model_options={'transformer_options':{'optimized_attention_override':object()}}
+ with pytest.raises(ValueError,match='VSA'):node.patch_temporal_model(m)
+ with pytest.raises(ValueError,match='MODEL input'):node.patch_temporal_model(None)
+
+
+def test_second_video_selection_preserves_first_video_and_images():
+ refs=[dict(kind='image',latent_h=4,latent_w=8),
+       dict(kind='video',latent_h=4,latent_w=4,latent_t=3,ref_audio_t=0),
+       dict(kind='video',latent_h=8,latent_w=4,latent_t=7,ref_audio_t=0)]
+ native=CORE.PackedLayout(3,7,8,8,2,refs=refs)
+ aligned,source,target=node.temporal_layout(native,refs,None,2)
+ assert int(source.sum())==56
+ assert torch.equal(aligned.position_ids[~source],native.position_ids[~source])
+ assert not source[3:23].any()
+
+@pytest.mark.parametrize('install_after_temporal',[False,True])
+@pytest.mark.parametrize('dense_name,expected_backend',[('comfy_kitchen_int8',2),('pytorch',0)])
+def test_model_attention_backend_is_honored_without_manual_selection(monkeypatch,install_after_temporal,dense_name,expected_backend):
+ from types import ModuleType
+ mod=ModuleType('comfy.ldm.modules.attention')
+ def dense(*args,**kwargs):return 'kitchen'
+ mod.get_attention_function=lambda name,default=None:dense if name==dense_name else default
+ monkeypatch.setitem(sys.modules,'comfy.ldm.modules.attention',mod)
+ def setter(optimized_attention):
+  def optimized_attention_override(_, *args, **kwargs):return optimized_attention(*args,**kwargs)
+  optimized_attention_override.__module__='comfy.model_patcher'
+  optimized_attention_override.__qualname__='ModelPatcher.set_model_optimized_attention.<locals>.optimized_attention_override'
+  return optimized_attention_override
+ incoming=setter(dense)
+ m=Model()
+ if not install_after_temporal:m.model_options={'transformer_options':{'optimized_attention_override':incoming}}
+ patched=node.patch_temporal_model(m)
+ selected=[]
+ def capture(ctx,previous=None,backend=None,**kwargs):
+  selected.append((backend,previous));return 'temporal attention'
+ monkeypatch.setattr(node,'attention_override',capture)
+ l,r=example()
+ args=dict(input=torch.zeros(1),timestep=torch.ones(1),c=dict(minimax_payload=dict(layout=l,refs=r),transformer_options={'optimized_attention_override':incoming}))
+ def fn(x,t,**c):
+  assert c['transformer_options']['optimized_attention_override']=='temporal attention'
+  return 'ok'
+ assert patched.model_options['model_function_wrapper'](fn,args)=='ok'
+ assert selected==[(node.BACKENDS[expected_backend],incoming)]
+
+@pytest.mark.parametrize('hidden_by_later_dense_node',[False,True])
+def test_sol_node_owned_producer_is_bypassed_but_temporal_override_remains(monkeypatch,hidden_by_later_dense_node):
+ Patch=type('SparseAttnPatch',(),{'__module__':'comfy_extras.nodes_sparse_attention'})
+ patch=Patch();patch.vsa=False;patch.topk_ratio=0.
+ previous=None
+ def sparse_override(*args):return patch,previous,args
+ sparse_override.__module__='comfy_extras.nodes_sparse_attention'
+ sparse_override.__qualname__='make_attention_override.<locals>.override'
+ def block_patch(*args):return patch,args
+ block_patch.__module__=sparse_override.__module__
+ block_patch.__qualname__='make_h3_block_patch.<locals>.block_patch'
+ m=Model();m.model_options={'transformer_options':{'optimized_attention_override':None if hidden_by_later_dense_node else sparse_override,'patches_replace':{'dit':{'block':block_patch}}}}
+ patched=node.patch_temporal_model(m)
+ selected=[]
+ def capture(ctx,previous=None,backend=None,sparse_patch=None,**kw):
+  selected.append((backend,sparse_patch));return 'temporal sol hook'
+ monkeypatch.setattr(node,'attention_override',capture)
+ l,r=example();opts=patched.model_options['transformer_options']
+ args=dict(input=torch.zeros(1),timestep=torch.ones(1),c=dict(minimax_payload=dict(layout=l,refs=r),transformer_options=opts))
+ def fn(x,t,**c):
+  assert c['transformer_options']['patches_replace']['dit']=={}
+  assert c['transformer_options']['optimized_attention_override']=='temporal sol hook'
+  return 'ok'
+ assert patched.model_options['model_function_wrapper'](fn,args)=='ok'
+ assert selected==[(node.SOL_BACKEND,patch)]
+ assert m.model_options['transformer_options']['patches_replace']['dit']
+
+
+def test_sol_dense_window_preserves_temporal_policy(monkeypatch,capsys):
+ l,r=example();aligned,src,tgt=node.temporal_layout(l,r,None)
+ ctx=node.core.TemporalContext(aligned.position_ids[:,0].float()[None]/40,src[None],tgt[None],torch.ones(1,l.seq_len,dtype=torch.bool),2.,.2,.25,6.)
+ plan=node.fast.make_plan(aligned.position_ids,src,tgt,2.,.2,.25,6.)
+ patch=SimpleNamespace(dense_reason=lambda *a:'outside sigma window',log_once=lambda *a:None)
+ q,k,v=[torch.randn(1,2,l.seq_len,128) for _ in range(3)]
+ monkeypatch.setattr(node.fast,'attention',lambda q,k,v,p:node.core.dense_reference(q,k,v,ctx))
+ expected=node.core.dense_reference(q,k,v,ctx)
+ out=node.attention_override(ctx,backend=node.SOL_BACKEND,plan=plan,sparse_patch=patch,dense_backend=node.BACKENDS[0],log_first=True)(None,q,k,v,2,skip_reshape=True,skip_output_reshape=True)
+ torch.testing.assert_close(out,expected)
+ assert 'dense fallback: outside sigma window' in capsys.readouterr().out
+
+
+def test_sol_sparse_window_runs_temporal_kernel_and_logs_activation(monkeypatch,capsys):
+ from types import ModuleType
+ ck=ModuleType('comfy_kitchen');ck.sol_attn_is_available=lambda device:True
+ monkeypatch.setitem(sys.modules,'comfy_kitchen',ck)
+ l,r=example();aligned,src,tgt=node.temporal_layout(l,r,None)
+ ctx=node.core.TemporalContext(aligned.position_ids[:,0].float()[None]/40,src[None],tgt[None],torch.ones(1,l.seq_len,dtype=torch.bool),2.,.2,.25,6.)
+ plan=node.fast.make_plan(aligned.position_ids,src,tgt,2.,.2,.25,6.)
+ patch=SimpleNamespace(dense_reason=lambda *a:None,sinks=lambda *a:((0,0),(0,0)),tau=1.3,extra_tokens=256)
+ actual_sol=node.sol.attention
+ calls=[]
+ def oracle(q,k,v,key_bias=None,**kw):
+  calls.append(key_bias)
+  mask=None if key_bias is None else key_bias.view(1,1,1,-1)
+  return torch.nn.functional.scaled_dot_product_attention(q.transpose(1,2),k.transpose(1,2),v.transpose(1,2),attn_mask=mask).transpose(1,2)
+ monkeypatch.setattr(node.sol,'attention',lambda *a,**kw:actual_sol(*a,**kw,kernel=oracle))
+ q,k,v=[torch.randn(1,2,l.seq_len,128) for _ in range(3)]
+ expected=node.core.dense_reference(q,k,v,ctx)
+ out=node.attention_override(ctx,backend=node.SOL_BACKEND,plan=plan,sparse_patch=patch,dense_backend=node.BACKENDS[0])(None,q,k,v,2,skip_reshape=True,skip_output_reshape=True)
+ torch.testing.assert_close(out,expected,atol=1e-6,rtol=1e-5)
+ assert len(calls)==plan.frames+1
+ assert 'sol-attn ACTIVE' in capsys.readouterr().out
+
+
+def test_node_prepares_selected_kitchen_before_sampling_but_not_flash(monkeypatch):
+ calls=[]
+ monkeypatch.setattr(node.sol.native_kernel,'prepare_model',lambda model:calls.append(model))
+ m=Model();m.load_device=torch.device('cuda:0')
+ node.patch_temporal_model(m,backend=node.BACKENDS[0])
+ assert calls==[]
+ node.patch_temporal_model(m,backend=node.BACKENDS[2])
+ assert calls==[m]

tests/test_minimax_h3_temporal_bootstrap.py

@@ -0,0 +1,105 @@
+import concurrent.futures,hashlib,io,sys
+from pathlib import Path
+from types import SimpleNamespace
+import pytest
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+import minimax_h3_temporal_bootstrap as boot
+
+
+class NativeFunction:
+ def __init__(self,result):self.result=result
+ def __call__(self):return self.result
+
+@pytest.fixture
+def setup(monkeypatch,tmp_path):
+ monkeypatch.setenv('BFS_H3_TEMPORAL_CACHE_DIR',str(tmp_path))
+ boot._READY.clear();boot._FAILED.clear()
+ source=b"VERSION='h3_native_temporal_kernel_v1'\ndef is_available(device):return True\n"
+ binary=b'fake verified native fixture'
+ assets=[('temporal_attention.py','https://test.invalid/api',hashlib.sha256(source).hexdigest(),1024),('backends/cuda/native_temporal/libck_temporal_sm120.so','https://test.invalid/bin',hashlib.sha256(binary).hexdigest(),1024)]
+ monkeypatch.setattr(boot,'ASSETS',tuple(assets))
+ requests=[]
+ def urlopen(request,timeout):
+  requests.append(request.full_url)
+  return io.BytesIO(source if request.full_url.endswith('/api') else binary)
+ monkeypatch.setattr(boot.urllib.request,'urlopen',urlopen)
+ loads=[]
+ def cdll(path):
+  loads.append(path)
+  return SimpleNamespace(ck_temporal_abi_version=NativeFunction(1),ck_temporal_arch=NativeFunction(120))
+ monkeypatch.setattr(boot.ctypes,'CDLL',cdll)
+ monkeypatch.setattr(boot,'compatible',lambda d:(True,None))
+ yield SimpleNamespace(root=boot.cache_root(),requests=requests,loads=loads,source=source,binary=binary)
+ boot._READY.clear();boot._FAILED.clear()
+
+
+def test_first_use_installs_once_and_preserves_current_environment(setup):
+ with concurrent.futures.ThreadPoolExecutor(max_workers=4) as pool:
+  results=list(pool.map(boot.ensure,['cuda:0']*4))
+ assert all(m is results[0] for m in results)
+ assert len(setup.requests)==2 and len(setup.loads)==1
+ assert (setup.root/'temporal_attention.py').read_bytes()==setup.source
+ assert not list(setup.root.rglob('*.partial'))
+ assert str(results[0].__file__).startswith(str(setup.root))
+
+
+def test_cached_bytes_work_without_network_after_restart(setup,monkeypatch):
+ assert boot.ensure('cuda:0') is not None
+ boot._READY.clear();boot._FAILED.clear()
+ def offline(*a,**kw):raise AssertionError('valid cache must not access the network')
+ monkeypatch.setattr(boot.urllib.request,'urlopen',offline)
+ assert boot.ensure('cuda:0') is not None
+
+
+def test_corrupted_download_never_loads_or_executes_code(setup,monkeypatch):
+ monkeypatch.setattr(boot.urllib.request,'urlopen',lambda *a,**kw:io.BytesIO(b'wrong bytes'))
+ assert boot.ensure('cuda:0') is None
+ assert setup.loads==[]
+ assert not (setup.root/'temporal_attention.py').exists()
+ assert not list(setup.root.rglob('*.partial'))
+
+
+def test_failure_is_not_retried_every_attention_block(setup,monkeypatch):
+ calls=[]
+ def broken(*a,**kw):calls.append(1);raise OSError('offline')
+ monkeypatch.setattr(boot.urllib.request,'urlopen',broken)
+ assert boot.ensure('cuda:0') is None
+ assert boot.ensure('cuda:0') is None
+ assert len(calls)==1
+
+
+def test_bad_existing_cache_is_replaced_only_after_verified_download(setup):
+ setup.root.mkdir(parents=True)
+ (setup.root/'temporal_attention.py').write_text('corrupt')
+ assert boot.ensure('cuda:0') is not None
+ assert (setup.root/'temporal_attention.py').read_bytes()==setup.source
+
+
+def test_size_limit_aborts_and_cleans_partial_file(tmp_path,monkeypatch):
+ monkeypatch.setattr(boot.urllib.request,'urlopen',lambda *a,**kw:io.BytesIO(b'12345'))
+ with pytest.raises(RuntimeError,match='size limit'):boot.download(tmp_path/'asset','https://test.invalid/asset','ignored',2)
+ assert list(tmp_path.iterdir())==[]
+
+
+def test_unsupported_device_keeps_existing_backend_without_download(setup,monkeypatch):
+ monkeypatch.setattr(boot,'compatible',lambda d:(False,'unsupported GPU'))
+ assert boot.ensure('cuda:0') is None
+ assert setup.requests==[]
+
+
+def test_device_failure_does_not_block_supported_device(setup,monkeypatch):
+ monkeypatch.setattr(boot,'compatible',lambda d:(d=='cuda:1','unsupported GPU'))
+ assert boot.ensure('cuda:0') is None
+ assert boot.ensure('cuda:1') is not None
+
+
+def test_opt_out_is_checked_before_gpu_or_network(monkeypatch):
+ monkeypatch.setenv('BFS_H3_TEMPORAL_AUTO_INSTALL','0')
+ assert boot.compatible('cuda:0')==(False,'automatic setup disabled')
+
+
+def test_bad_binary_arch_is_rejected_before_python_api_execution(setup,monkeypatch):
+ before=set(sys.modules)
+ monkeypatch.setattr(boot.ctypes,'CDLL',lambda p:SimpleNamespace(ck_temporal_abi_version=NativeFunction(1),ck_temporal_arch=NativeFunction(89)))
+ assert boot.ensure('cuda:0') is None
+ assert not [n for n in set(sys.modules)-before if n.startswith('_bfs_ck_temporal')]

tests/test_minimax_h3_temporal_fast.py

@@ -0,0 +1,29 @@
+import sys
+from pathlib import Path
+import torch
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+import minimax_h3_temporal_fast as fast
+import minimax_h3_native_temporal_core as core
+
+
+def oracle(q,k,v):
+    scores=q.double()@k.double().transpose(-1,-2)/q.shape[-1]**.5
+    return (scores.softmax(-1)@v.double()).to(q.dtype),scores.logsumexp(-1).float()
+
+
+def test_exact_partition_matches_original_temporal_policy():
+    torch.manual_seed(5)
+    frames=7;sf=2;tf=3;n=4+frames*sf+2+frames*tf
+    pos=torch.zeros(n,3,dtype=torch.float64)
+    times=torch.tensor([0,1,5,9,13,17,18],dtype=torch.float64)*5/3
+    pos[4:18,0]=times.repeat_interleave(sf)
+    pos[20:,0]=times.repeat_interleave(tf)
+    source=torch.zeros(n,dtype=torch.bool);source[4:18]=True
+    target=torch.zeros(n,dtype=torch.bool);target[20:]=True
+    q,k,v=[torch.randn(1,2,n,8) for _ in range(3)]
+    for strength,cap in [(2.,6.),(0.,6.),(2.,0.),(10.,6.)]:
+        plan=fast.make_plan(pos,source,target,strength,.2,.25,cap)
+        ctx=core.TemporalContext(pos[:,0].float()[None]/40,source[None],target[None],torch.ones(1,n,dtype=torch.bool),strength,.2,.25,cap)
+        actual=fast.attention(q,k,v,plan,flash=oracle,frame_chunk=3)
+        expected=core.dense_reference(q,k,v,ctx)
+        torch.testing.assert_close(actual,expected,atol=1e-6,rtol=1e-5)

tests/test_minimax_h3_temporal_kernel.py

@@ -0,0 +1,55 @@
+import sys
+from pathlib import Path
+from types import ModuleType,SimpleNamespace
+import torch
+import pytest
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+import minimax_h3_temporal_kernel as bridge
+
+
+def fake_api(monkeypatch):
+ api=ModuleType('comfy_kitchen.temporal_attention');api.VERSION='h3_native_temporal_kernel_v1';api.is_available=lambda d:True
+ records=[]
+ api.TemporalPolicy=lambda *args:args
+ def dense(q,k,v,p,**kw):records.append(('dense',p,kw));return torch.ones(1,dtype=torch.bfloat16)
+ def sol(q,k,v,p,**kw):records.append(('sol',p,kw));return torch.ones(1,dtype=torch.bfloat16)
+ api.int8_attention=dense;api.sol_attn=sol
+ monkeypatch.setitem(sys.modules,'comfy_kitchen.temporal_attention',api)
+ return api,records
+
+
+def test_bridge_uses_one_native_call_and_preserves_policy(monkeypatch):
+ api,records=fake_api(monkeypatch)
+ q=SimpleNamespace(device=torch.device('cuda'),dtype=torch.bfloat16)
+ plan=SimpleNamespace(source_start=4,source_stop=10,target_start=12,target_stop=21,source_frame_rows=2,target_frame_rows=3,frame_times=(0.,.7,1.4))
+ ctx=SimpleNamespace(strength=2.,window=.2,falloff=.25,cap=6.)
+ bridge.run('dense',q,q,q,plan,ctx)
+ bridge.run('sol',q,q,q,plan,ctx,tau=1.3,token_aug=256,sink_q=(0,1))
+ assert len(records)==2 and [r[0] for r in records]==['dense','sol']
+ assert records[0][1]==(4,10,12,21,2,3,(0.,.7,1.4),2.,.2,.25,6.)
+ assert records[1][2]=={'tau':1.3,'token_aug':256,'sink_q':(0,1)}
+ ctx.strength=3.
+ assert bridge.policy_for(api,plan,ctx)[7]==3.
+
+
+def test_bridge_rejects_wrong_extension_version(monkeypatch):
+ api,_=fake_api(monkeypatch);api.VERSION='unknown'
+ q=SimpleNamespace(device=torch.device('cuda'))
+ with pytest.raises(RuntimeError,match='version'):bridge.api_for(q)
+
+
+def test_cpu_and_missing_binary_keep_compatibility_path(monkeypatch):
+ assert bridge.api_for(torch.zeros(1)) is None
+ api,_=fake_api(monkeypatch);api.is_available=lambda d:False
+ monkeypatch.setattr(bridge.bootstrap,'ensure',lambda d:None)
+ q=SimpleNamespace(device=torch.device('cuda'))
+ assert bridge.api_for(q) is None
+
+
+def test_prepare_model_runs_before_sampling_when_cuda_is_selected(monkeypatch):
+ calls=[]
+ monkeypatch.setattr(bridge,'api_for',lambda q:calls.append(q.device) or 'ready')
+ assert bridge.prepare_model(SimpleNamespace(load_device=torch.device('cuda:0'))) == 'ready'
+ assert calls==[torch.device('cuda:0')]
+ assert bridge.prepare_model(SimpleNamespace(load_device=torch.device('cpu'))) is None
+ assert bridge.prepare_model(SimpleNamespace()) is None

tests/test_minimax_h3_temporal_kitchen.py

@@ -0,0 +1,55 @@
+import sys
+from pathlib import Path
+from types import ModuleType
+import torch
+import pytest
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+import minimax_h3_temporal_kitchen as kitchen
+import minimax_h3_temporal_fast as fast
+import minimax_h3_native_temporal_core as core
+import minimax_h3_native_temporal as node
+
+@pytest.mark.parametrize('strength,cap',[(2.,6.),(0.,6.),(2.,0.),(10.,6.)])
+def test_frame_bias_matches_full_temporal_rule(strength,cap):
+    torch.manual_seed(42)
+    frames,sf,tf=7,2,3
+    n=4+frames*sf+2+frames*tf+3
+    pos=torch.zeros(n,3,dtype=torch.float64)
+    times=torch.tensor([0,1,5,9,13,17,18],dtype=torch.float64)*5/3
+    pos[4:18,0]=times.repeat_interleave(sf);pos[20:41,0]=times.repeat_interleave(tf)
+    source=torch.zeros(n,dtype=torch.bool);source[4:18]=True
+    target=torch.zeros(n,dtype=torch.bool);target[20:41]=True
+    plan=fast.make_plan(pos,source,target,strength,.2,.25,cap)
+    ctx=core.TemporalContext(pos[:,0].float()[None]/40,source[None],target[None],torch.ones(1,n,dtype=torch.bool),strength,.2,.25,cap)
+    q,k,v=[torch.randn(1,2,n,8) for _ in range(3)]
+    calls=[]
+    def oracle(q,k,v,attn_mask=None):
+        if attn_mask is not None:
+            assert attn_mask.shape==(1,1,1,n)
+            calls.append(attn_mask)
+        return torch.nn.functional.scaled_dot_product_attention(q,k,v,attn_mask=attn_mask)
+    actual=kitchen.attention(q,k,v,plan,ctx,kernel=oracle)
+    expected=core.dense_reference(q,k,v,ctx)
+    torch.testing.assert_close(actual,expected,atol=1e-6,rtol=1e-5)
+    torch.testing.assert_close(kitchen.attention(q,k,v,plan,ctx,kernel=oracle),actual)
+    if strength and cap:
+        assert len(calls)==2*frames
+        assert plan._kitchen_biases.numel()==frames*n
+        assert torch.all(plan._kitchen_biases[:,~source]==0)
+
+
+def test_detects_only_official_model_attention_backend(monkeypatch):
+    mod=ModuleType('comfy.ldm.modules.attention')
+    kernels={'pytorch':lambda:None,'comfy_kitchen_int8':lambda:None}
+    mod.get_attention_function=lambda name,default=None:kernels.get(name,default)
+    monkeypatch.setitem(sys.modules,'comfy.ldm.modules.attention',mod)
+    def setter(optimized_attention):
+        def optimized_attention_override(_, *args, **kwargs):
+            return optimized_attention(*args,**kwargs)
+        optimized_attention_override.__module__='comfy.model_patcher'
+        optimized_attention_override.__qualname__='ModelPatcher.set_model_optimized_attention.<locals>.optimized_attention_override'
+        return optimized_attention_override
+    for key,fn in kernels.items():assert node.dense_backend_selection(setter(fn))==key
+    assert node.dense_backend_selection(None) is None
+    with pytest.raises(ValueError):node.dense_backend_selection(lambda *args:None)
+    with pytest.raises(ValueError):node.dense_backend_selection(setter(lambda:None))

tests/test_minimax_h3_temporal_sol.py

@@ -0,0 +1,107 @@
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+import torch
+import pytest
+sys.path.insert(0,str(Path(__file__).resolve().parents[1]))
+import minimax_h3_temporal_sol as sol
+import minimax_h3_temporal_fast as fast
+import minimax_h3_native_temporal_core as core
+import minimax_h3_sparse_compat as compat
+
+@pytest.mark.parametrize('strength,cap',[(2.,6.),(0.,6.),(2.,0.),(10.,6.)])
+def test_temporal_sol_bias_matches_dense_policy(strength,cap):
+ torch.manual_seed(14);frames,sf,tf=7,2,3;n=44
+ pos=torch.zeros(n,3,dtype=torch.float64)
+ times=torch.tensor([0,1,5,9,13,17,18],dtype=torch.float64)*5/3
+ pos[4:18,0]=times.repeat_interleave(sf);pos[20:41,0]=times.repeat_interleave(tf)
+ source=torch.zeros(n,dtype=torch.bool);source[4:18]=True
+ target=torch.zeros(n,dtype=torch.bool);target[20:41]=True
+ plan=fast.make_plan(pos,source,target,strength,.2,.25,cap)
+ ctx=core.TemporalContext(pos[:,0].float()[None]/40,source[None],target[None],torch.ones(1,n,dtype=torch.bool),strength,.2,.25,cap)
+ q,k,v=[torch.randn(1,2,n,128) for _ in range(3)]
+ calls=[]
+ def oracle(q,k,v,key_bias=None,**kw):
+  assert kw['sink_blocks'][1]*64>=plan.source_stop
+  assert kw['topk_ratio']==0. and kw['tail'] is True
+  if key_bias is not None:
+   assert key_bias.shape==(n,)
+   assert torch.all(key_bias[~source]==0)
+  calls.append(key_bias)
+  return torch.nn.functional.scaled_dot_product_attention(q.transpose(1,2),k.transpose(1,2),v.transpose(1,2),attn_mask=None if key_bias is None else key_bias.view(1,1,1,-1)).transpose(1,2)
+ actual=sol.attention(q,k,v,plan,ctx,tau=1.3,kernel=oracle)
+ expected=core.dense_reference(q,k,v,ctx)
+ torch.testing.assert_close(actual,expected,atol=1e-6,rtol=1e-5)
+ assert len(calls)==(frames+1 if strength and cap else 1)
+ # Repeat uses the compact cached biases and does not change the policy.
+ torch.testing.assert_close(sol.attention(q,k,v,plan,ctx,tau=1.3,kernel=oracle),actual)
+
+
+def make_sparse(vsa=False,topk=0.,module='comfy_extras.nodes_sparse_attention'):
+ SparseAttnPatch=type('SparseAttnPatch',(),{'__module__':module})
+ patch=SparseAttnPatch();patch.vsa=vsa;patch.topk_ratio=topk
+ previous=None
+ def override(*args):return patch,previous,args
+ override.__module__=module
+ override.__qualname__='make_attention_override.<locals>.override'
+ def block_patch(*args):return patch,args
+ block_patch.__module__=override.__module__
+ block_patch.__qualname__='make_h3_block_patch.<locals>.block_patch'
+ return patch,override,block_patch
+
+
+def test_only_sol_owned_hooks_are_bypassed_without_mutating_model():
+ patch,override,block=make_sparse()
+ options={'optimized_attention_override':override,'patches_replace':{'dit':{('double_block',0):block},'other':{'keep':1}}}
+ dense,actual=compat.selection(options)
+ assert dense is None and actual is patch
+ copied=compat.without_owned_blocks(options,patch)
+ assert copied['patches_replace']['dit']=={}
+ assert copied['patches_replace']['other']=={'keep':1}
+ assert options['patches_replace']['dit']
+ # A dense selection applied after Sparse can hide the sparse closure; recover
+ # the exact same state from the node's owned H3 block hook.
+ hidden={**options,'optimized_attention_override':None}
+ assert compat.selection(hidden)==(None,patch)
+ unknown={**options,'patches_replace':{'dit':{'unknown':lambda:None}}}
+ with pytest.raises(ValueError,match='another attention/block'):compat.selection(unknown)
+
+@pytest.mark.parametrize('vsa,topk',[(True,0.),(False,.1)])
+def test_vsa_and_sla_still_rejected(vsa,topk):
+ _,override,block=make_sparse(vsa,topk)
+ with pytest.raises(ValueError,match='only sol-attn'):compat.selection({'optimized_attention_override':override})
+ with pytest.raises(ValueError,match='only sol-attn'):compat.selection({'patches_replace':{'dit':{'block':block}}})
+
+
+@pytest.mark.parametrize('module',[
+ '/home/alissonerdx/tools/ComfyUI/comfy_extras/nodes_sparse_attention',
+ 'C:\\ComfyUI\\comfy_extras\\nodes_sparse_attention',
+ 'nodes_sparse_attention',
+ 'comfy_extras.nodes_sparse_attention',
+])
+def test_comfy_file_loader_module_names_are_recognized(module):
+ patch,override,block=make_sparse(module=module)
+ options={'optimized_attention_override':override,'patches_replace':{'dit':{'block':block}}}
+ assert compat.selection(options)==(None,patch)
+ hidden={**options,'optimized_attention_override':None}
+ assert compat.selection(hidden)==(None,patch)
+ assert compat.without_owned_blocks(options,patch)['patches_replace']['dit']=={}
+
+@pytest.mark.parametrize('module',['other_sparse_attention','nodes_sparse_attention_fake','/tmp/not_nodes_sparse_attention'])
+def test_similar_unrelated_module_names_are_rejected(module):
+ _,override,block=make_sparse(module=module)
+ with pytest.raises(ValueError):compat.selection({'optimized_attention_override':override,'patches_replace':{'dit':{'block':block}}})
+
+
+def test_actual_comfy_sparse_factories_under_file_loader_name():
+ module='/home/alissonerdx/tools/ComfyUI/comfy_extras/nodes_sparse_attention'
+ fixture=Path(__file__).parent/'fixtures/comfy_sparse_node_factories.py'
+ namespace={'__name__':module}
+ exec(compile(fixture.read_text(),str(fixture),'exec'),namespace)
+ patch=namespace['SparseAttnPatch'](tau=1.3,topk_ratio=0.,vsa=False,sigma_start=1.,sigma_end=0.,min_tokens=12288,dense_blocks=set(),sink_conditioning='exact_kv_and_rows',extra_tokens=256,verbose=False)
+ override=namespace['make_attention_override'](patch,None)
+ block=namespace['make_h3_block_patch'](SimpleNamespace(),0,patch)
+ options={'optimized_attention_override':override,'patches_replace':{'dit':{('double_block',0):block}}}
+ assert compat.selection(options)==(None,patch)
+ assert compat.without_owned_blocks(options,patch)['patches_replace']['dit']=={}
+ assert options['patches_replace']['dit']

web/js/bfs_shot_loop.js

@@ -108,7 +108,7 @@ function Panel(io) {
     plan.auto_prompts = { ...(p.auto_prompts || {}) };
     if (!Array.isArray(plan.bounds)) plan.bounds = [];
     if (!Array.isArray(plan.segs)) plan.segs = [];
-    // plans before 1.113 had two references (ref / ref2): now a list of up to 12 pictures
+    // plans before 1.113 had two references (ref / ref2): now a list of up to MAX_REFS pictures
     plan.global_refs = refList(p.global_refs, p.global_ref, p.global_ref2);
     plan.segs.forEach(m => { if (m) m.refs = refList(m.refs, m.ref, m.ref2); });
     // Colored Mark: a color per picture (was: two marker colors), regions tied to a picture (was: region 2)
@@ -527,7 +527,7 @@ const meta = i => { while (plan.segs.length <= i) plan.segs.push({}); return pla
     const marks = s.color_guide ? [[s.color_guide_picture, pictureColor(plan, s.color_guide_picture)],
       ...s.regions.filter(r => String(r.text || "").trim() || (r.points || []).length).map(r => [r.picture, pictureColor(plan, r.picture)])] : [];
     return JSON.stringify([s.prompt || plan.global_prompt || "", s.names, marks, modeOf(s), s.mask?.text || "", s.target || "",
-      plan.auto_prompt?.template || "", plan.auto_prompt?.style || "h3"]);
+      plan.auto_prompt?.template || "", plan.auto_prompt?.style || "h3", plan.auto_prompt?.trigger ?? "motion_transfer", "motion-trigger-bindings-v7"]);
   };
   const writtenOf = s => plan.auto_prompts?.[segKey(s)] || null;
   const setAuto = (k, v) => { plan.auto_prompt = { ...plan.auto_prompt, [k]: v }; save(); };
@@ -639,9 +639,9 @@ const meta = i => { while (plan.segs.length <= i) plan.segs.push({}); return pla
     } catch { /* ignore */ }
   }
   // ---- stored shots: redo some, or combine what is stored into an MP4 (low memory, on the server)
-  async function redo(shots, cascade = true) {
+  async function redo(shots, cascade = true, cancel = false) {
     try {
-      const r = await api.fetchApi("/bfs/shotloop/redo", { method: "POST", body: JSON.stringify({ plan: { ...plan }, shots, cascade }) });
+      const r = await api.fetchApi("/bfs/shotloop/redo", { method: "POST", body: JSON.stringify({ plan: { ...plan }, shots, cascade, cancel }) });
       const j = await r.json(); if (j.error) error.value = j.error; else { Object.assign(cache, j); prog.done = j.done; }
     } catch (e) { error.value = `${e}`; }
   }
@@ -953,7 +953,15 @@ function upgradeConditioningWidgets(node, info) {
 
 function upgradeLegacySlots(node, info) {
   const cls = node.comfyClass;
-  if (cls === "BFSShotH3Conditioning" || cls === "BFSShotH3ConditioningSimple") { upgradeConditioningWidgets(node, info); return; }
+  if (cls === "BFSShotH3Conditioning" || cls === "BFSShotH3ConditioningSimple") {
+    upgradeConditioningWidgets(node, info);
+    // 1.115.9 appends MODEL routing to Simple; preserve all existing slot indices/links.
+    if (cls === "BFSShotH3ConditioningSimple") {
+      if (!node.inputs?.some(x => x.name === "model")) node.addInput("model", "MODEL");
+      if (!node.outputs?.some(x => x.name === "model")) node.addOutput("model", "MODEL");
+    }
+    return;
+  }
   // ref_image / ref_image_2 outputs at slots `at`, at+1 -> refs (list) / ref_count
   const listOutputs = at => {
     const a = node.outputs?.[at], b = node.outputs?.[at + 1];

web/js/bfs_video_compare.js

@@ -125,6 +125,7 @@ function buildPlayer(host) {
     const r = strip.getBoundingClientRect(), n = st.data?.frames || 0;
     if (n) seek(Math.min(n - 1, Math.max(0, Math.floor((e.clientX - r.left) / r.width * n))));
   });
+  const notes = el("div", {style:"white-space:pre-wrap;font-size:12px;padding:6px;color:#e6b566"});
   host.append(el("div", { class: "bfsvc" }, [
     el("div", { class: "row" }, [el("span", {}, "top"), selTop, btnSwap, el("span", {}, "bottom"), selBot, selMode, btnFull]),
     stage,
@@ -132,6 +133,7 @@ function buildPlayer(host) {
     el("div", { class: "row" }, [btnPrev, btnPlay, btnNext, scrub, btnMute, info]),
     el("div", { class: "row" }, [btnSave, btnRec, selFormat, el("label", { title: "Draw the names on the saved / recorded video" }, [chkLabels, "labels"]), lblRefs]),
     strip,
+    notes,
   ]));
 
   const fps = () => st.data?.fps || 24;
@@ -456,6 +458,7 @@ function buildPlayer(host) {
       lblRefs.style.display = d.refs?.length ? "flex" : "none";
       showRefs();
       st.data = d;
+      notes.textContent = (d.notes || []).join("\n");
       empty.style.display = "none";
       scrub.max = String(Math.max(0, d.frames - 1));
       refreshSelects();

web/js/docs/BFSShotPlanner.md

@@ -61,14 +61,14 @@ Three run modes:
   stored on disk; then BFS Shot Join expands the part of the graph between the planner and itself (ComfyUI node
   expansion) for the next shot, so that shot is conditioned *after* the previous one exists and continuity
   (*head* / *history*) works. No browser re-queue (the tab can be closed), Save Video fires once with the full video,
-  and after a failure a new Queue resumes from the stored shots (*Reset loop* starts over). Previews connected inside
+  and after a failure a new Queue resumes from the stored shots (*Redo all* regenerates every shot). Previews connected inside
   the shot's part of the graph show every shot. Needs exactly one BFS Shot Planner before the join. Same idea as the
   recursive loop of Ethanfel's MiniMax H3 Context Loop.
 - **Queue loop (one shot per run)**: each run generates the next pending shot and stores it on
   disk. The join blocks every node after it until the last shot, then outputs the full video,
   so a Save Video node only fires once. With *Auto-queue the next shot* on (default), the panel re-queues
   the workflow after each shot, so one click still renders the whole video (keep the ComfyUI tab open: the panel
-  does the re-queue). *Reset loop* starts over. Changing the plan starts a new run. It resumes from the stored
+  does the re-queue). *Redo all* regenerates every shot. Changing the plan starts a new run. It resumes from the stored
   shots after a failure, and continuity *head* / *history* needs it: each shot reads the result of the one before.
 - **Auto loop (all shots at once)**: every shot in one queue run, each node runs for every shot before the next node.
   Holds all decoded shots in memory. Continuity *head* / *history* does not work here (all shots are conditioned
@@ -392,6 +392,31 @@ explicitly in any template.
 
 ### 🤖 Auto prompt (1.113.0)
 
+**Visual edits and color bindings (1.115.4):** the guide fixes the assigned performers' motion, action timing,
+head movement, gaze and expressions. Explicitly request the target style, background, lighting, weather, outfit
+or camera changes. For example: `Replace the man with the goblin from <Picture 1>, render the whole scene in anime
+style, and use the hall from <Picture 2> as the setting. Keep the guide performance and timing.` A setting can also
+be described in text. Use **full frame** generation for changes to the entire scene; masked generation only changes
+the generated region.
+
+References can have different roles. In the example above, Picture 1 defines the character and Picture 2 defines
+the environment. To describe the latter, choose **scene / setting** in that picture's description preset. For several
+characters, bind colors only to their character references: `The subject marked in green uses <Picture 1>; the subject
+marked in purple uses <Picture 2>; <Picture 3> defines the setting.` Picture indices stay unchanged. A marking color
+identifies the same performer across position changes, occlusion and cuts; it is not a target skin/clothing color.
+Facial highlights are performance annotations. Prompt wording is generic and does not require a specific 2D/3D marking
+implementation. This update changes prompt writing, not the existing marking renderer or tracking behavior.
+
+After updating, restart ComfyUI, hard-refresh the browser and click **Write all prompts** to regenerate previously
+written prompts with these rules. Existing written prompts are preserved until rewritten.
+
+**1.115.5:** a setting replacement alone does not request a new rendering style. The writer preserves the source
+medium unless the actual request or an explicitly designated style reference changes it; instructional examples
+must not supply unrequested target attributes. For background replacement, use full-frame generation, disable
+paste/crop-only composition and inpainting outside a subject mask, set **first_frame** to **none**, and leave
+**setting_ref** off (that option adds the original source location). The full aligned guide still contains source
+scene geometry, so a correct prompt alone does not guarantee complete background replacement by a checkpoint.
+
 In **Prompts & refs → 🤖 Auto prompt**, *Write the prompts with the VLM* turns the prompt boxes (global and per shot)
 into **requests** ("replace the bottle with the red can from Picture 2", "Picture 1 dances in the rain"), and the VLM
 writes each shot's MiniMax H3 prompt in the official full-reference format (MiniMaxAI/MiniMax-H3
@@ -721,3 +746,19 @@ or *Crop + mask*.
 | Only the face or the hair is described | the click was on the head: click the body |
 | The other person is described | add an exclude point (right-click) on the other person, **👁 Segment**, save, describe again |
 | The swap still hits the wrong person | make the description more specific (position: "on the left", clothing colours) |
+
+### Cancel a pending redo
+
+After marking a stored shot for redo, click **Undo redo** to keep its previous take.
+**Clear redo selection** cancels all pending redo selections and preserves stored files and loop progress.
+Once a shot has been regenerated, its previous take cannot be restored through this control.
+**Refresh status** rereads the cache status; it cannot recover files deleted by older versions of Reset loop.
+Changing the seed, model or other render settings still invalidates stored shots even after cancelling redo.
+
+### Experimental native temporal motion transfer
+
+`BFS Shot H3 Conditioning` now has **native temporal alignment (experimental)** in `guide_mode`. Connect MODEL in and route its MODEL output to the guider. This uses the matching experimental LoRA, the shot as native `<Video 1>`, temporal RoPE alignment and soft temporal attention; it does not use Add Guide. Use 24 fps and disable continuity/first-frame/mask guides and VSA. Simple appends a MODEL output without moving existing outputs. The standalone **BFS MiniMax H3 Native Temporal Alignment (Experimental)** node provides the same patch for native workflows. See `H3_NATIVE_TEMPORAL.md` for the full contract and validation limits.
+
+Native temporal performance (1.115.11): `native_temporal_backend` defaults to `flash partitions (no compile)`, preserving the same temporal penalty. `native_guide_scale` can reduce only guide resolution (try 0.5); output dimensions and frame count stay unchanged. The original Flex backend remains available for comparison.
+
+Prompt bindings (1.115.12): completed short captions are no longer expanded into legacy two-subject templates. Auto-written prompts whose request/reference/mark signature is outdated are regenerated with a connected Planner VLM; otherwise use Write again before rendering. The writer rejects picture indices beyond the connected references after its bounded retry. Character outfits come from identity references, while requested setting changes override conflicting source scenery/lighting.

web/js/shotloop/common.js

@@ -472,7 +472,7 @@ export const secondPassSize = (w, h, sp) => {
 };
 
 // ---- reference pictures (<Picture 1..12>)
-export const MAX_REFS = 12;
+export const MAX_REFS = 100;
 // a reference list of the plan: `refs` (plans from 1.113 on), else the old ref / ref2 pair; empty slots at the end dropped
 export const refList = (refs, a = "", b = "") => {
   const out = (Array.isArray(refs) ? refs : [a, b]).map(x => (typeof x === "string" ? x : "")).slice(0, MAX_REFS);

web/js/shotloop/view_prompts.js

@@ -1,7 +1,7 @@
 // 📝 Prompts & refs tab: the global reference / prompt, the VLM (shot suggestions) and the reference descriptions ({details}).
 import { h, fld, check, section, pill, hint, select, refBox, promptChips, MAX_REFS, MARK_PALETTE, pictureColor } from "./common.js";
 
-const PRESETS = ["short", "object", "full body", "head / face", "face attributes", "outfit", "custom"];
+const PRESETS = ["short", "object", "full body", "head / face", "face attributes", "outfit", "scene / setting", "custom"];
 
 function globalSection(c) {
   const { plan } = c;
@@ -116,8 +116,10 @@ function autoSection(c) {
     h("div", { class: "row", style: "gap:12px;align-items:center" }, [
       check(A.enabled, v => c.setAuto("enabled", v), "Write the prompts with the VLM",
         "On: the prompt boxes (global and per shot) hold your request ('change the bottle into a red can', 'Picture 1 dances in the rain'); the VLM writes each shot's MiniMax H3 prompt from key frames of the shot, the pictures and their descriptions, the request and the shot's Colored Marks, selection and mode, in H3's official full-reference format."),
-      fld("Style", select(A.style || "h3", [["h3", "H3 official (six sections)"], ["captions", "OpenJutsu captions (short)"]],
-        v => c.setAuto("style", v), { style: "width:auto" }), "captions: 'Replace the ... in the video with ... from <Picture 1>. Keep ...', the training captions' style (OpenJutsu); a template on the prompt input wins over both"),
+      fld("Style", select(A.style || "h3", [["h3", "H3 official (six sections)"], ["captions", "Short edit request"]],
+        v => c.setAuto("style", v), { style: "width:auto" }), "Choose prompt format. The guide performance stays fixed; request style, setting, lighting, outfit or camera changes explicitly. Identify setting pictures in the request; bind colors only to character references."),
+      fld("Trigger", select(A.trigger ?? "motion_transfer", [["motion_transfer", "motion_transfer:"], ["body_swap", "body_swap:"], ["", "None (plain prompt)"]],
+        v => c.setAuto("trigger", v), { style: "width:auto" }), "Match the LoRA caption prefix. Choose None for a model trained without a trigger."),
       fld("Key frames", select(String(A.frames || 0), [["0", "auto (by length)"], ...[3, 4, 5, 6, 8, 10, 12].map(n => [String(n), String(n)])],
         v => c.setAuto("frames", parseInt(v) || 0), { style: "width:auto" }), "auto: about 1.5 per second, 3 to 8"),
       fld("Source video in the prompt", select(A.source || "guide", [["guide", "guide video (aligned guide)"], ["video", "<Video 1> (native reference video)"]],

web/js/shotloop/view_run.js

@@ -40,7 +40,7 @@ function loopSection(c) {
       h("div", { class: "prog" }, [h("div", { style: `width:${(100 * prog.done / Math.max(1, prog.count || active.value.length)).toFixed(1)}%` })]),
       h("div", { class: "row", style: "margin-top:6px" }, [
         plan.run === "queue" ? check(plan.auto_continue, v => c.setPlan("auto_continue", v), "Auto-queue the next shot") : null,
-        h("button", { onClick: () => c.progress(false) }, "↻ Status"), h("button", { class: "dng", onClick: () => c.progress(true) }, "⟲ Reset loop"),
+        h("button", { onClick: () => c.progress(false) }, "↻ Refresh status"), h("button", { title: "Cancel pending redo selections. Stored shots and loop progress are preserved.", onClick: () => c.progress(true) }, "Clear redo selection"),
       ]),
       hint(plan.run === "queue"
         ? "Each run generates one shot and stores it. Nodes after BFS Shot Join only run on the last shot, with the full video."
@@ -91,7 +91,8 @@ function cacheSection(c) {
       const i = c.segs.value.indexOf(active.value[k]);
       return h("span", { class: ["pill", g ? "ok" : ""], title: g ? "stored and up to date" : "runs on the next Queue" }, [
         `#${(i >= 0 ? i : k) + 1} ${g ? "💾" : "⏳"}`,
-        g ? h("button", { class: "chip", style: "margin-left:4px", title: "generate this shot again on the next Queue"
+        (cache.redo || []).includes(k) ? h("button", { class: "chip", title: "Cancel redo and keep the stored take",
+          onClick: () => c.redo([k], true, true) }, "Undo redo") : g ? h("button", { class: "chip", style: "margin-left:4px", title: "generate this shot again on the next Queue"
           + (plan.redo_only ? " (only this shot)" : " (and the shots that continue from it)"),
           onClick: () => c.redo([k], !plan.redo_only) }, "↻") : null]);
     }) : [hint(n ? "Nothing stored yet for this video: Queue the workflow." : "")]),