huchukato-ComfyUI-QwenVL-Mod/update.md

ComfyUI-QwenVL Update Log

Version 2.12.9 (2026/10/30)

• New VL preset LTX › NSFW Loop for the LTX-2.5 pose-loop workflows (A→B→C→B→A keyframe loops): derived from the FL2VA guidance but framed for MULTIPLE keyframe images visited in order, with the loop-closure rule (final pose = first pose) instead of the two-frame transition framing. Source of truth is Garage presets/ltx.json — the bundled copy + chat alias ship with this release as offline fallback.

Version 2.12.8 (2026/10/24)

• Chat defaults aligned with the QwenVL node: default GGUF model is the Q6_K (was Q8_0 — a second download users did not expect), and the chat's own max-tokens default is 2048 (the js always sent its own 1024, so the backend's raised default never applied).
• Director Analyze: new /minimax_director/llm_models endpoint + a model dropdown in the Analyze settings when several GGUF models are found (the free-text field stays as fallback).

Version 2.12.7 (2026/10/24)

• Storyboard request handled OUTSIDE the JSON protocol: the GGUF model kept summarizing the stills away ('Still 1/2/3 ready') no matter what the guide said. Now the stills are generated as raw text by a dedicated recipe prompt (Still 1/2/3 + dense Director idea) and the chat reply is composed in code — preset IMG › Storyboard + passthrough=false applied, three Generate Still choices offered, zero queue on the request. The still clicks recover each still from this composed message.

Version 2.12.6 (2026/10/24)

• Storyboard routing: the generic image-enhancer enforcement no longer runs after the storyboard routing took over — it was re-writing the enhancer's prompt widget with a garbage "intent" cleaned from the bare Generate Still message ("till 1"). Storyboard mode now emits exactly: processor text+populated_text, passthrough false, queue.

Version 2.12.5 (2026/10/24)

• Chat Flow dropdown (next to Config MMH3): Auto / Storyboard stills. Storyboard mode injects the still-flow guide unconditionally and makes the routing deterministic in code — the still prompt always lands on the WildcardProcessor's text widgets (or the enhancer's prompt widget when the graph has no processor), passthrough stays false, queue only on a Generate Still choice. No more guide-vs-node-rule conflicts deciding the outcome.

Version 2.12.4 (2026/10/24)

• Chat storyboard flow fixes from live testing: the node snapshot now carries the input socket names (the enhancer routing never fired without them, so the chat set passthrough=true and bypassed the QwenVL); the storyboard guide now targets the WildcardProcessor's text widgets (the QwenVL's prompt input is WIRED from it — writing the QwenVL's prompt widget had no effect, which produced character sheets instead of stills); no queue on the storyboard request itself, only on a Generate Still choice.

Version 2.12.3 (2026/10/24)

• Fixed the chat crash (ComfyUI process aborted): the graph snapshot rode into the model context whole every turn — the Edit21 workflow's PixNote HTML alone is ~5.5KB, so the second chat turn overflowed the GGUF context and llama.cpp aborted (Fatal Python error). Widget values in the snapshot are now capped at 300 chars; operational widgets (prompt, preset, steps…) stay intact.

Version 2.12.2 (2026/10/23)

• Fixed "missing node type" / AttributeError: module 'AILab_QwenVL_GGUF_PromptEnhancer' has no attribute ...: the GGUF modules no longer raise at import when llama-cpp-python is missing — all nodes register, the Unified node hides GGUF models with an explanatory note, and a clear actionable error appears only when a GGUF model actually runs. The node loader also drops half-initialized modules so a failed import can no longer poison later imports.

Version 2.11.1 (2026/10/06)

• Cleaned the GGUF and HF model catalogs: three VL-capable DavidAU heretic models per backend (4B / 9B / 27B) — all verified image-text-to-text with working mmproj files; dropped text-only duplicates and the mislabeled "3.8-9B" entry. VL models cover text-only enhance too, so the separate text section is gone.
• MiniMax R2VA presets (NSFW + SFW): fixed the reference-image semantics — a reference supplies subject identity only, it is never treated as a keyframe to reproduce. Character sheets and multi-view refs now produce a new scene instead of animating the sheet itself.
• MiniMax presets (all durations): completed dialogue handling across R2VA/FL2VA/T2V — [DIALOGUE]…[/DIALOGUE] and [SBJn] says [D]…[/D] are preserved verbatim as <d>[English] …</d>, no invented lines; restored the truncated SFW FL2VA Loop tail.

Version 2.11.0 (2026/10/06)

• New MiniMax SFW preset family: MiniMax › SFW I2VA/R2VA/FL2VA/FL2VA Loop on the VL nodes and MiniMax › SFW T2V/R2VA T2V/FL2VA T2V on the Prompt Enhancer — same structure, dialogue handling and mode-specific alignment as the NSFW presets, without the explicit-content directives.
• Renamed the ambiguous MiniMax › NSFW VL preset to MiniMax › NSFW I2VA (legacy names keep working as aliases); added text-context aliases so MiniMax › NSFW/MiniMax › SFW resolve to the T2V variants instead of being used as raw templates.
• I2VA output normalization: a single canonical <Picture 1> reference-binding line is enforced and malformed duplicate prefaces are stripped; [SBJn] says [D]…[/D] dialogue lost by the model is re-injected verbatim before overall_soundscape.

Version 2.10.2 (2026/10/01)

• Removed the Livepeer Agent render node and its external MCP/network integration after the Livepeer Agent Hackathon ended.
• Removed the associated Qwen Chat capability selector, deterministic capability routing, legacy alias, test, and promotional assets to keep the package focused and Registry-compatible.

Version 2.10.1 (2026/09/26)

• New preset MiniMax › NSFW FL2VA Loop (5s/10s/15s): dedicated seamless-loop enhancer — feed the same image as first AND last frame; the prompt forces periodic-motion vocabulary, locked/oscillating camera, stable lighting, full-duration pacing and a cyclic soundscape so the clip closes perfectly on its opening frame.
• The base MiniMax › NSFW FL2VA preset is now transition-only: LOOP MODE removed, end state correctly targets Picture 2 instead of the first frame when the two references differ.

Version 2.9.9 (2026/09/25)

• Chat Config MMH3: added R2VA Native and R2VA Native Turbo — select the ref2va unet (image-ref conditioned) instead of fl2va, enable the ref2v turbo LoRA in turbo mode, and auto-switch the enhancer preset to an R2VA variant preserving the selected duration. Chat phrases like "usa r2va turbo" trigger them too.

Version 2.9.8 (2026/09/25)

• New preset IMG › Qwen Edit on the QwenVL nodes (and Qwen Image Edit on the Prompt Enhancer): rewrites any input as a Qwen Image 2.1-style edit instruction — imperative edit clause, explicit keep-unchanged list, <imageN> placeholders preserved, reference-transfer and lighting-consistency clauses. Complements the __qwen21/*__ wildcards for free-form edits.

• Node menu category now shows as 🔮 QwenVL-Mod; new banner and icon.

Version 2.9.7 (2026/09/25)

• Republish of the consolidated 2.9.5 content — registry tombstoned 2.9.5/2.9.6 after deletion, so the same feature set ships under a new version number.

Version 2.9.5 (2026/09/25)

• Workflow Chat: wildcard autocomplete — typing __ in the input lists TagForge wildcards (fuzzy-matched as you type); arrows/Enter/Tab to pick, Esc to dismiss. Silent when TagForge is not installed.
• Workflow Chat: media words (video/clip/scene) are stripped only when they open the directive ("a video of...", "video, ..."), never mid-sentence — "lasts for all the video" now reaches the prompt intact.

Version 2.9.4 (2026/09/25)

• Workflow Chat: action text no longer loses duration phrases that belong to the action itself (e.g. "the kiss lasts 5 seconds"); clip-length specs still set the duration widget only when they read as meta ("10s video", "video of 8 seconds").
• Stray leading connectors ("of", "showing", "a") left behind by stripped duration/media words are cleaned up.

Version 2.9.3 (2026/09/25)

• Workflow Chat: wildcard support — __wildcard__ tokens in your message are expanded via ComfyUI-TagForge (when installed) before the request reaches the LLM, so action wildcards work inside chat prompts.
• Repeat button now loads the last user message into the input for editing instead of resending it immediately.
• Config selector label renamed to "Config MMH3" (the presets are MiniMax H3 specific).

Version 2.9.2 (2026/09/25)

• Removed the bogus requires-comfyui >=1.0.0 constraint — ComfyUI versions are 0.x, and the mismatch was disabling the node pack in ComfyUI Manager.

Version 2.9.1 (2026/09/25)

• Rewrote the MiniMax config examples in the chat system prompt as plain prose — the dense key=value; action syntax was triggering the registry python_minified_code rule (false positive on chat_service.py:88).
• Moved the registry admin workflow out of the published package.

Version 2.9.0 (2026/09/25)

• MiniMax model matrix in Qwen Workflow Chat: Native, Native Turbo, 10Eros and 10Eros Turbo configs now drive the whole sampler setup — UNet selection (nvfp4 / 10Eros beta5 int8), step count (20 quality / 8 turbo), LoRA enable/bypass and the correct LoRA file per mode (lightx2v fl2v/ref2v vs TenStrip combined fusion).
• BlockSparseAttention tau is set per config (1.0 quality, 1.3 turbo).
• New unified HF/GGUF prompt enhancer node; custom_prompt widget renamed to prompt; backend dropdowns removed (model prefix is the single selector); keep_model_loaded defaults to False; duration becomes a widget.
• Presets split into per-family files; style-tag presets now lead the output with an explicit style declaration; denser Pony→Natural and Detailed caption outputs.
• MiniMax output normalization applied to the GGUF backend; legacy AILab_* nodes hidden from the add-node menu (superseded by Unified); refreshed node icons.
• Publish workflow: changelog extraction rewritten in plain shell (registry scanner false positive on embedded Python).

Version 2.8.14 (2026/09/22)

• Removed the duplicate AILab_LivepeerRender node from the Add Node menu.
• Added web/js/legacy_aliases.js to rewrite old AILab_LivepeerRender nodes to QwenVL_LivepeerRender on workflow load, keeping old workflows compatible.
• Updated QwenVL_Livepeer.py NODE_CLASS_MAPPINGS / NODE_DISPLAY_NAME_MAPPINGS to expose only the canonical node name.

Version 2.8.13 (2026/09/22)

• Refreshed node branding with a Pixaroma-style colour scheme: each QwenVL-Mod family now has recognisable title/body colours (Vision, Vision GGUF, Enhancer, Livepeer, Utils).
• Updated web/js/appearance.js to apply colours in beforeRegisterNodeDef, preserving user/workflow colour choices and covering all current nodes (Unified, Livepeer, LoadMedia, VRAM cleanup, story split).
• Added shared stylesheet web/css/qwen_nodes.css and wrapped the chat composer in a dedicated panel for cleaner spacing and styling.
• Standardised QwenVL_LoadMedia category to QwenVL-Mod/Utils.

Version 2.8.12 (2026/09/22)

• Added 🎨 Pony→Natural Language back to the QwenVL preset_prompt dropdown (_preset_prompts + qwenvl), in addition to the PromptEnhancer style.
• The same converter is now available both as a vision preset (when an image is attached) and as a text-only style enhancement.

Version 2.8.11 (2026/09/22)

• Removed the redundant mmh3_acceleration chat guide from _chat_guides. MiniMax H3 acceleration modes are already handled deterministically by Qwen Chat selectors and use native/10eros/turbo commands, so the extra system-prompt text was no longer useful.

Version 2.8.10 (2026/09/22)

• Moved the new 🎨 Pony→Natural Language preset from the QwenVL preset_prompt dropdown into the PromptEnhancer enhancement_style dropdown (qwen_text.styles).
• Removed the old 🎨 Pony→Z-Image and 🎨 Pony→Flux PromptEnhancer styles.
• The Pony tag converter is now selected as a style in the PromptEnhancer node (HF and GGUF), keeping the QwenVL preset list focused on vision/analysis/video tasks.

Version 2.8.9 (2026/09/22)

• Replaced the model-specific 🎨 Pony→Z-Image and 🎨 Pony→Flux presets with a single generic 🎨 Pony→Natural Language preset in AILab_System_Prompts.json.
• The new preset converts Danbooru/Pony tag prompts into a clean, model-agnostic natural-language paragraph suitable for Flux, Krea, Qwen-Image 2.1, and other image generation models.
• Removes score_* quality tags and Pony meta tags (source_pony, rating_*, etc.), preserves optional [style] tags, and handles NSFW content without censoring.

Version 2.8.8 (2026/09/21)

• Turbo config now actually toggles the Turbo LoRA: the graph snapshot includes subgraph inner nodes (qualified ids like "105:213"), findNode resolves them, and the MiniMax configs emit set_node_mode on LoraLoader nodes inside the enhancer subgraph — Turbo enables it, Native/10Eros bypass it (10Eros has TURBO fused).
• Turbo also resets unet_name to the NVFP4 checkpoint — 10Eros + Turbo LoRA stacking is forbidden.

Version 2.8.7 (2026/09/21)

• Selector dropdowns no longer inject "use <config>" into the chat text: the Config / Livepeer pickers are sent as a separate directives field and applied deterministically server-side, so the user message (and the prompt widget) stays clean.
• Fixed duration regex matching "s" before "seconds" ("5 seconds" left "econds" in the prompt); also widened the capability-path duration parser to seconds/secondi.
• System prompt: MiniMax action now targets the "prompt" widget (was preset_prompt, inconsistent with the deterministic path).

Version 2.8.6 (2026/09/21)

• Fixed directive cleaning when the config command is separated by a period ("use native. generate a 5s video, ..."): the leftover leading dot blocked the generation-prefix stripper, leaving "generate a," fragments in the prompt. Leading dots are now stripped after each cleaning stage; legitimate trailing periods are preserved.

Version 2.8.5 (2026/09/21)

• Added selector dropdowns under the chat composer: a Config picker (Auto / Native / 10Eros / Turbo LoRA) injects the deterministic MiniMax command, and a Livepeer picker is auto-populated from the render node's capability widget — no more typing use <name> by hand. Both selectors persist, work with empty prompts (config-only switch), and skip the model check since routing is deterministic.
• Fixed the explicit use parser swallowing the trailing separator ("use nano-banana. " kept the dot in the capability name) and directive cleaning stripping legitimate trailing periods.

Version 2.8.4 (2026/09/21)

• Fixed directive cleaning when a config keyword precedes the request ("usa 10Eros: genera un video di…"): the generation prefix is now stripped after config removal too, so no "genera un di" fragments reach the prompt widget.

Version 2.8.3 (2026/09/21)

• Added deterministic chat shortcuts that skip the LLM entirely: use <capability> <prompt> drives the Livepeer render node directly (dropdown or custom_capability, with duration and aspect-ratio parsing), and use native / use 10Eros / use turbo / config A-B-C apply the full MiniMax H3 sampler config on the enhancer node — correct widgets every time, instant, and immune to malformed local-model JSON.
• Config-only requests leave the prompt widget untouched; a cleaned scene directive is written when the message carries one.

Version 2.8.2 (2026/09/21)

• Image-enhancer directives are now cleaned deterministically: routing keywords ("use native", "use 10Eros", "usa turbo"), generation prefixes ("generate a 5s video", "fai un video") and duration mentions are stripped before the action text reaches the prompt widget, so the inner enhancer receives only the scene action.
• When the request contains only config keywords, the prompt widget is left untouched instead of receiving the meta text.

Version 2.8.1 (2026/09/21)

• Rewrote the /dlstatus.html download-log page with a multi-line HTML template and html.escape, removing the long single-line string that the registry security scanner flagged as minified code.

Version 2.8.0 (2026/09/21)

• Added QwenVL_LivepeerRender: a self-contained node that renders images and videos through the Livepeer Agent network over MCP, with a capability dropdown (image and video models), IMAGE output for still results, static preview for images and animated preview for video, and a source_video input that picks a frame from a clip as the i2v reference.
• Added QwenVL_LoadMedia (utils category): a single picker for images and videos in input/output folders with built-in preview, an upload button for videos, and resolution of untagged input-directory names.
• Qwen Chat now drives the Livepeer render node directly: it routes generation requests through the node (never bypassing it), clears stale custom_capability values, bypasses mismatched saver nodes and re-enables the matching one, and supports video attachments with browser-side frame sampling for the refine loop.
• The chat assets panel lists videos too and syncs the selection with QwenVL_LoadMedia; the capability dropdown now includes image-generation models, and a MiniMax H3 config switch (10Eros/Native/Turbo) runs before prompt routing.
• Added unified HF/GGUF QwenVL nodes with a backend dropdown.
• Rewrote the chat system prompt: shorter, with concrete examples and MiniMax configuration handled first.
• Fixed MCP error handling and stopped sending video parameters to image-only capabilities; the chat mirrors the latest user message language and no longer pre-formats enhancer prompts.

Version 2.7.13 (2026/09/16)

• Fixed enforce_image_reference_bindings to use the passthrough value from the action set instead of the stale workflow snapshot, so I2VA workflows with passthrough=false no longer receive the For the target video... binding line in the chat-managed prompt.
• Suppressed full-format video prompt-writing guides (MiniMax H3, LTX, Wan) from the chat prompt when an image-enhancer target is active, preventing the chat from outputting a complete integrated_multimodal_description/overall_soundscape prompt.
• Strengthened the exact-target instruction: the chat must write only a concise English action directive and let the inner QwenVL node build the final preset prompt.

Version 2.7.12 (2026/09/16)

• Strengthened image-enhancer chat instructions: Qwen Chat must inspect the provided image pixels to understand how the requested action applies and read the target node's current preset_prompt value and supplied prompt-writing guide.
• The chat must emit an English set_widget_value enhancer directive for the target prompt instead of copying the user's message, while the inner QwenVL remains responsible for full image analysis and final preset formatting.
• Pixel inspection by the chat is explicitly limited to interpreting the requested action, preventing it from inventing or duplicating visual scene details.

Version 2.7.11 (2026/09/16)

• Added the image2 input to AILab_QwenVL_GGUF and AILab_QwenVL_GGUF_Advanced, matching the HF nodes: image is Picture 1 and image2 is Picture 2 for R2VA workflows.
• The second image is encoded after the first and before sampled video frames, and its hash participates in the prompt cache key via the combined image2/video hash slot.
• Added the same reference-image tooltips used by the HF nodes.

Version 2.7.10 (2026/09/16)

• Image-enhancer routing now keeps a valid English action directive written by the chat model instead of always overwriting it with the raw user message, so Qwen Chat translates and refines the request instead of copying it verbatim.
• Execution-only confirmations are detected with a bilingual command vocabulary (verbs, objects, fillers) instead of a fixed phrase list, covering variants such as "Run it", "Genera", "Ok", and "Esegui il video con il prompt precedente".
• Bare confirmations and pre-formatted preset prompts emitted by the model are still replaced with the last descriptive user request, and passthrough remains forced to false for image-enhancer workflows.

Version 2.7.9 (2026/09/16)

• Added deterministic I2VA action normalization before queueing: the latest user request is always written to the active inner enhancer's prompt widget and passthrough is always forced to false.
• Missing prompt actions are inserted automatically, while hallucinated pre-formatted chat prompts are replaced with the raw user intent.

Version 2.7.8 (2026/09/16)

• Image-to-video workflows with an exposed preset enhancer now receive a short action-only prompt from Qwen Chat and run the internal QwenVL with passthrough=false.
• The internal workflow node analyzes the actual reference image and applies the full preset, avoiding visual details hallucinated by the chat model.
• Passthrough remains available for workflows without image input.

Version 2.7.7 (2026/09/16)

• Moved backend, model, token, temperature, and thinking controls into a compact Settings panel opened with a gear button.
• The Settings panel is collapsed by default, remembers its state, and is disabled during inference.

Version 2.7.6 (2026/09/16)

• Added a secure recursive output-image endpoint for Qwen Chat Assets, including images saved in nested folders such as output/PMP/....
• Assets are sorted by modification time and limited to supported image formats and paths contained inside the ComfyUI output directory.

Version 2.7.5 (2026/09/16)

• Fixed the ComfyUI Assets gallery endpoint to use /internal/files/output without the /api prefix.

Version 2.7.4 (2026/09/16)

ComfyUI Assets Picker

• Added a localized ComfyUI Assets gallery for selecting recent images directly from the ComfyUI output folder.
• Selected assets are sent to Qwen as the priority reference and synchronized with the workflow when a LoadImageOutput node is available.
• Workflows with multiple LoadImageOutput nodes show an explicit target-node choice instead of guessing.
• Output images selected by LoadImageOutput are now recognized correctly during automatic workflow image collection.

Version 2.7.3 (2026/09/16)

Deterministic MiniMax I2VA Reference Binding

• MiniMax H3 image-to-video prompts using passthrough now receive the mandatory <Picture 1> binding automatically when images are present, even if the chat model omits it.
• The exact final prompt sent to the workflow is echoed back after an automatic correction.
• Prompt guidance now treats reference pixels as authoritative and forbids invented identity, appearance, clothing, framing, environment, lighting, or style details that conflict with the source image.

Version 2.7.2 (2026/09/16)

Qwen Chat Execution Follow-up Fix

• Added a localized Generate video fallback choice when Qwen updates a prompt but omits queue_workflow.
• Assistant messages and choice labels now follow the language of the latest user message, while generated workflow prompts remain in English.
• Prevented responses from claiming that execution started when no queue action was emitted.

Version 2.7.1 (2026/09/16)

Qwen Chat MiniMax H3 Routing Fix

• Generated image and video prompts are now written in English unless another prompt language is explicitly requested.
• Nodes exposing both preset_prompt and passthrough now use the passthrough route first: the final preset-formatted prompt is written to the node's actual exposed prompt widget and passthrough is enabled.
• Added graph-specific routing guidance with the exact node ID and exposed prompt widget, preventing omission of the prompt or passthrough actions on promoted subgraph inputs.
• Queue responses no longer ask for execution confirmation after already adding queue_workflow.

Version 2.7.0 (2026/09/16)

Qwen Workflow Chat, Image Attachments, Guided Workflow Actions, and Qwen 3.8 Models

Qwen Workflow Chat

• Added a native ComfyUI sidebar chat powered by selectable local HF/Transformers or GGUF Qwen models.
• Added structured workflow inspection, validated widget updates, node enable/bypass actions, and autonomous queue execution when requested.
• Added clickable response choices, retry recovery for malformed model JSON, and clean message extraction when a response cannot be fully parsed.
• Added preset-aware prompt routing: enhancer subgraphs receive raw user intent, while accessible top-level Qwen nodes can receive final prompts through passthrough.
• Added workflow image analysis and a direct image attachment control with preview, removal, resizing, and priority over workflow images.
• Added a repeat-last-message button, busy spinner, refined message layout, and persistent English/Italian interface switch.
• Set Qwen3.8-9B-heretic-uncensored.Q8_0.gguf as the default GGUF chat model, with automatic fallback when unavailable.
• Chat inference defaults to 1024 max tokens and temperature 0.2; thinking mode automatically raises the token budget to at least 4096.
• Chat models are now always unloaded from memory before queueing memory-intensive workflows.
• Added browser-local conversation/preferences persistence, bounded context, backend inference locking, and explicit unload API.
• Restricted automation to an allowlisted action protocol; arbitrary code, filesystem, network, graph creation, connection, and deletion operations are not supported.

Models and Diagnostics

• Added Qwen 3.5 and Qwen 3.8 heretic, uncensored, and thinking model options for HF and GGUF backends.
• Replaced an unavailable private Qwen 3.8 repository with the public petruhonk/Qwen3.8-9B-Distill-uncensored-heretic checkpoint.
• Reduced repeated cache lookup debug output so generation, image processing, model loading, and failure events remain visible in ComfyUI logs.

Version 2.6 (2026/09/03)

🎬 Camera & Style Tag Dropdowns + Pony Prompt Converters + Qwen 3.8 + LTX 2.3 FL2VA + MiniMax H3 Loop Mode

Major update introducing camera/style tag dropdowns across all nodes, Pony→Z-Image and Pony→Flux prompt converter presets, a rewritten Danbooru Tags preset, Qwen 3.8 uncensored models, LTX 2.3 FL2VA presets, and extensive MiniMax H3 improvements (loop mode, anatomical coherence, dialogue trigger, reference tags).

🎥 Camera Tag Dropdown

• New camera_tag parameter in all nodes: QwenVL, QwenVL (Advanced), QwenVL (GGUF), QwenVL (GGUF Advanced), QwenVL Prompt Enhancer, QwenVL Prompt Enhancer (GGUF)
• 19 camera movement options: None, [STATIC_CAMERA], [LOCKED_OFF], [SLOW_ZOOM_IN], [SLOW_ZOOM_OUT], [FAST_ZOOM_IN], [FAST_ZOOM_OUT], [PAN_LEFT], [PAN_RIGHT], [TILT_UP], [TILT_DOWN], [DOLLY_IN], [DOLLY_OUT], [TRACKING_LEFT], [TRACKING_RIGHT], [CRANE_UP], [CRANE_DOWN], [ORBIT], [HANDHELD], [ROLL]
• Recency bias injection: tags injected as prefix AND re-injected at the end of the prompt so Qwen respects the camera directive
• STATIC_CAMERA / LOCKED_OFF override: these tags override RICHNESS RULES and force a completely static camera with no motion whatsoever
• Subject kept alive: during camera tag movements, the subject description is maintained (no disappearing characters)

🎨 Style Tag Dropdown (PromptEnhancer only)

• New style_tag parameter in QwenVL Prompt Enhancer and QwenVL Prompt Enhancer (GGUF) — T2V/T2VA presets only
• 12 visual style options: None, [ANIME], [PHOTOREALISTIC], [3DCG], [CARTOON], [CLAYMATION], [WATERCOLOR], [VINTAGE], [NOIR], [CYBERPUNK], [FANTASY], [SOFTFOCUS], [HENTAI]
• Same injection mechanism as camera_tag: prefix + final reminder for recency bias
• For image-reference modes, identity and composition remain anchored to the references

🐴 Pony Prompt Converter Presets

• 🎨 Pony→Z-Image: converts Pony/SDXL Danbooru-style tag prompts into optimized natural-language prompts for Z-Image-Turbo (6B S3-DiT from Alibaba). Removes score_ tags, converts tags to flowing sentences, adds precise photography vocabulary (Kodak Portra 400, lens specs), NSFW conversion rules, style detection, 6-part formula. CFG=0-1, no negative prompts, 8 steps, 1024×1024.
• 🎨 Pony→Flux: converts Pony/SDXL Danbooru-style tag prompts into natural-language prompts for Flux.1 (12B MMDiT from Black Forest Labs). Removes score_ tags, rich detail for photorealism (skin texture, pores, micro-details, film grain), 3-6 sentence flowing paragraph, up to 2MP resolution.
• Both presets handle mixed input (tags + natural language) and multilingual input (Italian/English/any → English output)

🏷️ Rewritten Tags Preset (🖼️ Tags)

• Generates Pony/SDXL Danbooru-style tags with score_9, score_8_up, score_7_up quality prefix
• 30-60 tags following structured order: quality → subject count → attributes → clothing → pose/action → environment → lighting → camera/composition → style
• Explicit NSFW tag support (no censoring)
• No meta tags ("masterpiece", "best quality") — uses score_ tags instead

🧠 Qwen 3.8 Uncensored Models

• HF VL: Qwen3.8-27B-Uncensored-Heretic-Abliterated (alexander2323)
• HF Text: Qwen3.8-4B-Distill-heretic (valiolla), Qwen3.8-9B-heretic-uncensored (rohit267)
• GGUF: two new GGUF models added at top of dropdown
• Switched to armand0e repaired Qwen3.5-9B-heretic weights (fixed broken checkpoint)

🎬 LTX 2.3 FL2VA Presets

• New LTX 2.3 NSFW FL2VA presets (5s/10s/20s) for First-Last-Frame to video
• Added LTX 2.3 NSFW 10s presets (I2V + T2V)
• Added LTX 2.3 NSFW 20s presets (I2V + T2V)
• Added missing LTX 2.3 T2V (10s/20s) styles to qwen_text.styles
• Dropdown reorder: I2V → FL2VA → T2V for logical workflow progression
• Removed LTX 2.3 T2V from QwenVL node dropdown (text-only belongs in PromptEnhancer)
• Danbooru tag support: LTX 2.3 presets now expand comma-separated Danbooru tags into cinematic scenes
• Anatomical correctness rule added to LTX I2V presets (5s + 10s)

🎬 MiniMax H3 Improvements

• LOOP MODE for FL2VA presets: seamless loop generation with pacing instructions (START IMMEDIATELY, 0.3s margin for continuity)
• Anatomical coherence instructions in all MiniMax H3 NSFW presets
• [DIALOGUE] trigger: explicit trigger for spoken content in MiniMax H3
• [P3]/[P4] manual reference tags for R2VA presets
• R2VA prompt fixes based on Civitai community feedback
• Camera control tags added to all MiniMax H3 NSFW presets
• I2VA priority: user text now takes priority over reference image
• Danbooru tag + style tag support in MiniMax H3 T2VA presets
• Mandatory sexual sounds in all H3 soundscape presets
• Forbidden output wrapping: no triple quotes or markdown in output

⚡ PromptEnhancer Changes

• max_tokens default raised to 8192 (was 1024) for longer, richer prompt generation
• Removed custom_system_prompt parameter from PromptEnhancer and GGUF PE
• Removed CUSTOM_ONLY_STYLE mode from PromptEnhancer

🐛 Bug Fixes

• Python 3.12 compatibility: fixed crash with rope_scaling=None
• PromptEnhancer TypeError: fixed run() has no 'video' kwarg
• SageAttention import: fixed for 2.x API (top-level vs .core)
• qwen3_5 → qwen3_vl config fallback for unrecognized model types
• Model list cleanup: removed censored + wrong abliterated versions, removed gated prithivMLmods models (require HF auth)

🐳 Docker / Provisioning

• LTX 2.5 support added (separate Docker/provisioning from LTX 2.3)
• Sulphur 2 (native uncensored LTX 2.3) + distill LoRA
• 10Eros-Max Turbo LoRA 8-step + 10Eros v1.5 fp8 + DMD hybrid v2 LoRA
• Pruna optimized VAE for faster LTX decode
• gemma-4-E2B-it-uncensored (abliterated) as LTX 2.5 prompt enhancer
• VFX workflows for MiniMax H3 Turbo (FL2VA and I2VA) — RIFE before upscale for 2K output
• Download progress log + /dlstatus endpoint served via ComfyUI
• Configurable ComfyUI version in RunPod, simplified Vast.ai provisioning
• Workflows, provisioning and Docker infra moved to ComfyUI-Garage repo

---

Version 2.5.1 (2026/08/11)

🚀 MiniMax-H3 Turbo LoRA Workflows + Official Diffusion + Uncensored Text Encoder + Explicit Tag Handling

⚡ MiniMax-H3 Turbo LoRA

• Added 4 new Turbo workflows: MiniMaxH3-Turbo-T2VA-Qwen3VL.json, MiniMaxH3-Turbo-I2VA-Qwen3VL.json, MiniMaxH3-Turbo-FL2VA-Qwen3VL.json, MiniMaxH3-Turbo-R2VA-Qwen3VL.json
• Uses Larryvrh/ComfyUI-MiniMax-H3-Turbo custom node and minimax_h3_turbo_v4_step600_ema.safetensors (~744 MB)
• Reduces sampling from ~20 steps to 6–8 steps with scheduler simple
• Works with all tasks: T2VA, I2VA, FL2VA and R2VA

📦 MiniMax-H3 Models

• T2VA / I2VA / FL2VA: now use the official Comfy-Org pruned INT8-ConvRot diffusion model:
  • minimax_h3_fl2va_pruned_int8_convrot.safetensors
• R2VA: keeps the official Comfy-Org pruned INT8-ConvRot model:
  • minimax_h3_ref2va_pruned_int8_convrot.safetensors
• Text encoder: remains the uncensored qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors — ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot
• VAEs and Turbo LoRA stay from Comfy-Org / larryvrh

🏷️ Explicit Tag Handling

• MiniMax H3 system prompts now treat short explicit tags (e.g., blowjob, deepthroat, missionary, cowgirl, facial, cumshot) as direct scene requests
• Qwen expands the tag into a full cinematic NSFW description instead of refusing or asking for clarification
• Works for both the vision node (qwenvl) and the PromptEnhancer (qwen_text)

📝 Documentation

• Updated minimax/Civitai_MiniMaxH3.md and runpod/README_MiniMaxH3.md with current model links and Turbo instructions
• Added 8-workflow list including Turbo variants

---

Version 2.5 (2026/08/05)

🎬 MiniMax H3 Full Support + Multi-Reference Input + R2VA/FL2VA Presets

Major update adding complete MiniMax H3 video+audio generation support with all 4 modes (T2VA, I2VA, FL2VA, R2VA), dedicated NSFW presets for each mode, and a new multi-reference input system.

🎬 MiniMax H3 Workflows (4)

|- T2VA: Text-to-video+audio with PromptEnhancer integration |- I2VA: Image-to-video+audio (first-frame, last-frame, or both) |- FL2VA: First-Last-Frame to video with TensorRT upscale + RIFE interpolation |- R2VA: Reference-to-video with up to 9 reference images, 3 videos, 3 audio clips |- All workflows auto-downloaded at boot from GitHub

🖼️ Multi-Reference Input (image2)

|- Renamed video input to image2: Clearer name for multi-reference workflows |- Frames as individual images: image2 frames are now passed to Qwen3-VL as separate images (not as a video sequence), enabling proper multi-reference analysis |- FL2VA: First frame on image, last frame on image2 → Qwen3-VL sees both |- R2VA: Primary reference on image, additional references on image2 → all visible |- Backward compatible: Existing workflows updated automatically, no reconnection needed

🎯 NSFW Presets (9 total)

|- 🎬 MiniMax H3 NSFW (5s/10s/15s): Base presets for T2VA/I2VA — 3-field format (integrated_multimodal_description + overall_soundscape + non_diegetic_music) |- 🎞️ MiniMax H3 NSFW R2VA (5s/10s/15s): Reference presets — 6-section format per official guide (subject_definitions + summary + retention_analysis + detailed_description + overall_soundscape + non_diegetic_music) with <Subject N>, <Picture N>, <Video N>, <Audio N> labels |- 🔄 MiniMax H3 NSFW FL2VA (5s/10s/15s): First-Last-Frame presets — describe the transition path between frames, favor single continuous shot

📦 Workflow Organization

|- Moved all MiniMax H3 workflows from vastai/workflows/MiniMaxH3/ to minimax/ |- Renamed MiniMaxH3-I2VA-FLF-Qwen3VL.json to MiniMaxH3-FL2VA-Qwen3VL.json |- Added Civitai_MiniMaxH3.md for Civitai publication |- Updated Dockerfile download paths

🛠️ Other Changes

|- Updated tooltips for image2 and frame_count |- GGUF node already handled multi-image correctly (no changes needed) |- Internal cache functions retain video_hash for backward compatibility

Version 2.4 (2026/08/04)

🚀 Local Model Discovery + Qwen3.5 Support + SageAttention

Major update porting the best features from the community fork (Deaquay/ComfyUI-Qwen3.5-Uncensored, 36 stars). Local model discovery, Qwen3.5 architecture detection, VRAM headroom, and image-handling fixes ported from the fork. SageAttention restoration is our own work from v2.0.3–v2.0.4, not from the fork.

🔍 Automatic Local Model Discovery

|- GGUF: Drop .gguf files in models/LLM/GGUF — they appear in the dropdown automatically with [local] prefix |- HF: Drop model directories in models/LLM/Qwen-VL — detected and listed automatically |- mmproj auto-pairing: GGUF mmproj files in the same directory are paired automatically |- Multi-path support: Respects extra_model_paths.yaml and all registered LLM paths |- No more JSON editing: Pre-configured models still work, but local models don't require config edits

🧠 Qwen3.5 Support

|- Architecture detection: Reads GGUF header (general.architecture) or HF config.json (model_type) instead of model name |- Disable thinking: Qwen3.5 thinking mode automatically disabled for prompt generation |- Top K tuning: Forces top_k=20 for Qwen3.5 (official recommendation) |- New models: Qwen3.5-4B/9B/27B (uncensored, heretic, unsloth) added to model configs

⚡ SageAttention Restored

|- Architecture-aware kernels: Blackwell FP8, Hopper FP8, Ada FP8, Ampere FP16 |- Graceful fallback: Falls back to SDPA if SageAttention unavailable |- Auto-mode priority: Sage → Flash → SDPA |- Runtime patching: Patches Qwen2Attention and Qwen3VLTextAttention without model reload

🛠️ Other Improvements

|- VRAM headroom: Proactive VRAM cleanup before each run (threshold: max 1GB or 8% total) |- Image handling: Fixed grayscale (mode="L") and RGBA image conversion |- BitsAndBytes fix: Separate loading path for quantized models (device_map direct vs CPU→GPU) |- max_tokens: Increased from 512 to 8192 for both Standard and Advanced nodes |- Batch image warning: Alerts when batch >1 images provided (uses first only)

---

Version 2.3 (2026/07/09)

🎬 New LTX 2.3 Presets + Multilingual Support

This update introduces specialized presets for LTX 2.3 video generation, with official prompting guides and complete multilingual support.

🎬 LTX 2.3 Presets

|- LTX 2.3 NSFW I2V: Image-to-Video preset for QwenVL nodes with image + text input |- LTX 2.3 NSFW T2V: Text-to-Video preset for Prompt Enhancer nodes with text only |- Official Prompting Guide: Core Actions, Visual Details, Audio structure for optimized results |- Simplified Format: No timeline, single fluid paragraph max 200 words |- NSFW Support: Full NSFW descriptions for both presets

🌐 Multilingual Support

|- Complete Translations: All LTX 2.3 presets available in multiple languages |- Consistent Style: Visual style detection maintained across all languages

📝 Prompting Tips

|- Core Actions: Primary subject motion and camera movement |- Visual Details: Appearance, environment, lighting, atmosphere |- Audio: Ambient sounds, music, dialogue cues |- Optimization: Specifically formatted for LTX 2.3 best results

🐛 Other Changes

|- Removed Qwen2.5-VL: Dropped support for Qwen2.5-VL (Qwen3-VL only now) |- Cleaned Model Lists: Added Qwen3.5-4B/9B Unredacted-MAX, Heretic-Stable 8B |- Removed Obsolete Nodes: kijai/ComfyUI_essentials removed |- Updated llama-cpp-python: Bumped to v0.3.45

---

Version 2.2.4 (2026/03/13)

🎬 Critical I2V Timeline Fixes & NSFW Presets Optimization

This update resolves critical issues with I2V Timeline preset generation and optimizes all NSFW presets for better user experience.

🚨 Major I2V Timeline Fixes

|- Style Coherence: Fixed AI changing anime→realism mid-sequence |- Character Stability: Fixed characters disappearing/appearing incorrectly
|- Natural Lighting: Fixed AI adding artificial lights not in image
|- Timeline Structure: Fixed continuous numbering (6,7,8...) instead of 0-5 restart
|- Format Consistency: Fixed missing parentheses and unwanted labels
|- Output Format: Each prompt starts directly with timeline markers

� NSFW Presets Optimization

|- Complete Specifications: All 8 NSFW presets now include full NSFW descriptions |- Emoji Display: Restored proper emoji rendering (🍿🎥🎬📖🍿🎥🎬📖) |- Clear Instructions: Removed confusing recommendations from presets |- User Notes: Token settings guide created for workflow optimization

📋 Technical Improvements

|- Timeline Markers: Correct (At X seconds: ...) format for all 4 prompts |- Character Continuity: Natural progression without forced artificial presence |- Lighting Rules: Logical progression instead of absolute prohibitions |- Style Detection: Consistent style application across all timeline segments

🎯 Model Recommendations

|- Qwen3-VL-8B: Recommended for I2V Timeline (20s) complex sequences |- Qwen3-VL-4B: Sufficient for I2V Scene (5s) single prompts |- Token Settings: 2048+ for 20s timeline, 1024+ for 5s prompts

Version 2.0.7 (2026/02/03)

|- FP16 Only: All HF nodes now use FP16 exclusively (~6GB VRAM on RTX 5090) |- Quantization Dropdown Removed: Eliminated confusing quantization options from all HF nodes |- Working Logic Restored: Reverted to proven working FP16 logic from commit d057b34 |- Both Nodes Fixed: Applied fixes to both Standard and Advanced nodes

🎯 Interface Simplification

|- Cleaner UI: Removed quantization dropdown from QwenVL and QwenVL Advanced nodes |- GGUF for Quantization: Users wanting quantized models should use dedicated GGUF nodes |- Consistent Parameters: Both nodes now have consistent parameter structure |- keep_last_prompt: Added missing parameter to maintain feature parity

🔧 Technical Changes

|- Memory Management: Improved model loading with CPU→CUDA transfer pattern |- Attention Backend: SDPA as stable default (SageAttention can be added later) |- Error Prevention: Removed all BitsAndBytes-related code paths |- Debug Logging: Enhanced logging for attention mode selection and model loading

🐛 Bug Fixes

|- OOM on RTX 5090: Fixed critical Out-Of-Memory during model loading |- Missing Parameters: Added frame_count and keep_last_prompt to Advanced node |- Quantization Logic: Simplified to prevent device mismatch and quantization errors |- Function Signatures: Corrected all process() functions after dropdown removal

💡 User Guidance

|- HF Nodes: Use for maximum quality with FP16 (~6GB VRAM) |- GGUF Nodes: Use for quantized models (Q4/Q8, 2-4GB VRAM) |- Performance: FP16 + SageAttention/FlashAttention for best speed |- Stability: SDPA fallback ensures compatibility across all hardware

---

Version 2.2.3 (2026/02/27)

🔧 CUDA 13 Compatibility Fix + Redundancy Removal

This update removes redundant memory management features and improves CUDA 13 compatibility.

🔧 Memory Management Cleanup

|- Removed unload_after_run: Eliminated redundant checkbox from all QwenVL nodes |- Fixed Parameter Errors: Resolved missing argument errors in GGUF and PromptEnhancer nodes |- Conflict Prevention: Prevents multiple unload operations that cause CUDA 13 crashes |- Simplified Interface: Cleaner node interface without redundant parameters |- VRAM Cleanup Node: Maintained for manual cleanup when needed

🐛 Bug Fixes

|- VastAI CUDA 13: Fixed crashes caused by conflicting unload operations |- WanMoeKSampler: Manual boundary setting prevents automatic switching crashes |- Memory Conflicts: Eliminated double cleanup between QwenVL and VRAM Cleanup nodes |- Parameter Errors: Fixed "missing 1 required positional argument: unload_after_run" errors |- Function Signatures: Corrected all run() and load_model() function signatures

📚 Documentation Updates

|- All READMEs: Updated to reflect unload_after_run removal |- Installation Guides: Simplified instructions without redundant parameters |- Troubleshooting: Added CUDA 13 specific guidance

🏆 Credits & Attribution

|- Community Feedback: Thanks to user testing that identified redundancy issue |- CUDA 13 Testing: Comprehensive testing on VastAI infrastructure |- Memory Management: Streamlined approach for better stability

---

Version 2.2.2 (2026/02/19)

🚀 Critical T2V/I2V Fixes + ComfyUI Optimizations

Version 2.2.2 (2026/02/19)

🚀 Critical T2V/I2V Fixes + ComfyUI Optimizations

This major update resolves critical workflow issues with T2V/I2V batch processing, same model reuse conflicts, and adds significant performance optimizations including Flash Attention 2 support and refined ComfyUI startup arguments.

🚀 T2V/I2V Workflow Fixes (Critical)

|- Batch Processing: Fixed critical T2V → GGUF issue with batch images [batch, height, width, channels] |- Frame Detection: Added automatic batch detection and individual frame processing |- Video Support: Enhanced video frame processing with proper shape handling |- Debug Enhanced: Comprehensive logging for batch processing troubleshooting

Same Model Reuse Fix

|- Conflict Resolution: Fixed crash when using same model between T2V and I2V nodes |- Memory Management: Enhanced cleanup with CUDA synchronization and timing |- Signature Mismatch: Resolved different signature patterns between nodes |- Aggressive Cleanup: Forced complete VRAM cleanup before model reload

🐳 Docker System Optimizations

|- ComfyUI Arguments: Optimized startup arguments for better performance |- Flash Attention 2: Enabled for 2-3x speedup on compatible hardware |- Experimental Features: Added validated arguments (--async-offload, --reserve-vram 2) |- Version Compatibility: Resolved version-specific argument differences

keep_model_loaded Enhancement

🔧 keep_model_loaded Enhancement

|- Missing Parameter: Added keep_model_loaded to PromptEnhancer node |- Consistent Behavior: Both GGUF and PromptEnhancer now have identical memory management |- Conditional Cleanup: Proper cleanup based on keep_model_loaded setting |- User Control: Full control over memory usage vs performance

🐳 Docker Final Optimization

|- Complete Build: All fixes integrated into final Docker build |- Flash Attention 2 Ready: Performance optimizations included |- Debug Enhanced: Comprehensive logging for troubleshooting |- Production Ready: Stable and optimized for deployment

📋 Files Modified

Node Files: |- AILab_QwenVL_GGUF.py: Batch processing + enhanced cleanup |- AILab_QwenVL_GGUF_PromptEnhancer.py: keep_model_loaded + cleanup

Docker System Files: |- runpod/Dockerfile: Flash Attention 2 + optimized ComfyUI arguments |- runpod/Dockerfile.4090: Flash Attention 2 + optimized ComfyUI arguments

🎯 Impact

|- T2V/I2V Workflows: Now fully functional with batch processing |- Same Model Reuse: No more conflicts between nodes |- Performance: 2-3x faster with Flash Attention 2 |- Stability: Robust memory management and cleanup |- Production: Ready for deployment with all optimizations

Version 2.2.1 (2026/02/18)

🔧 Critical GGUF VRAM Fix + Docker Optimization

This critical update resolves a major VRAM leak issue in GGUF nodes that was causing crashes after multiple executions, plus significant Docker improvements for better stability and functionality.

🔧 GGUF VRAM Fix (Critical)

• VRAM Leak Resolved: Fixed critical issue where GGUF models kept VRAM allocated after runs
• Issue #104: Addressed GitHub issue where GGUF models failed after 2 runs
• Aggressive Cleanup: Implemented comprehensive VRAM cleanup for all GGUF nodes
• Enhanced Clear Method: Added explicit model deletion and multiple CUDA cache clearing
• Debug Logging: Added detailed cleanup logging for troubleshooting
• Exception Handling: Robust error handling during cleanup operations

🚀 Performance Improvements

• Stable GGUF Operation: Nodes now work reliably without VRAM accumulation
• No More Crashes: GGUF Advanced and PromptEnhancer nodes stable after multiple runs
• Memory Management: Proper garbage collection and CUDA synchronization
• Resource Efficiency: Better VRAM utilization and cleanup

🐳 Docker Enhancements

• RunPod Methods: Adopted proven RunPod installation methods for services
• Jupyter Terminal: Switched to simple pip install jupyter (RunPod approach)
• FileBrowser: Using official get.sh script for latest version
• ComfyUI Latest: Always uses latest stable version automatically
• SSH Complete: Both server and client SSH for full networking
• FFmpeg ENV: Added IMAGEIO_FFMPEG_EXE environment variable
• Build Warnings: Removed problematic HF_TOKEN variables

📋 Files Modified

• AILab_QwenVL_GGUF.py: Enhanced VRAM cleanup
• AILab_QwenVL_GGUF_PromptEnhancer.py: Enhanced VRAM cleanup
• runpod/Dockerfile: RunPod methods + optimizations
• runpod/Dockerfile.4090: RunPod methods + optimizations

🎯 Impact

• Reliability: GGUF nodes now production-ready
• Stability: No more VRAM-related crashes
• Docker: Better container experience
• Performance: Consistent operation across multiple runs

---

Version 2.2.0 (2026/02/15)

🎬 WAN 2.2 Story Generation System

This major update introduces a complete story generation system for WAN 2.2, enabling the creation of continuous 20-second videos with 4 distinct segments, along with custom utility nodes and enhanced performance optimizations.

🎬 Story Generation System

• 4-Segment Videos: Complete system for generating 20-second continuous videos (4 × 5-second segments)
• Narrative Continuity: Intelligent prompt splitting ensures story coherence across segments
• Auto-Split Technology: Custom node automatically detects and splits prompt separators
• Workflow Integration: Ready-to-use WAN 2.2 Story workflow included
• NSFW Optimized: Enhanced prompts specifically designed for story generation

🔄 Custom Nodes

• Story Split Node: Intelligent text splitting with auto-detection of separators (\n, \n\n, \n\n\n)
• Smart Processing: Automatically handles inconsistent model output formatting
• Category: Located in 🔷QwenVL-Mod/Utils category
• Chaining Support: Full input/output connectivity for workflow integration

🎯 Enhanced Prompts

• WAN 2.2 NSFW Story: Optimized prompt format for 4-segment generation
• Separator Instructions: Clear guidelines for prompt separation
• Continuity Rules: Enhanced narrative continuity between segments
• Token Optimization: Recommended max_tokens=1536 for complete output

⚡ Performance Optimizations

• Context Settings: Optimized context_length for 8B models (65,536 tokens)
• Model-Specific: 4B models maintain 32,768 context for stability
• Memory Management: Improved VRAM usage and stability
• Crash Prevention: Safe defaults prevent ComfyUI crashes with large contexts

🐳 Docker Integration

• Container Ready: Complete Story system integrated in Docker containers
• Auto-Download: WAN2.2-I2V-AutoPrompt-Story.json and WAN2.2-T2V-AutoPrompt-Story.json workflows included
• Cloud Deployment: Ready for RunPod/VastAI deployment
• Service Integration: FileBrowser, JupyterLab, and ComfyUI with Story support

📁 File Structure

• Nodes: /nodes/story_split_node.py for intelligent prompt splitting
• Workflows:
  • /vastai/workflows/WAN2.2-I2V-AutoPrompt-Story.json ready-to-use I2V workflow
  • /vastai/workflows/WAN2.2-T2V-AutoPrompt-Story.json ready-to-use T2V workflow
• Docker: Updated /runpod/Dockerfile with Story workflows
• Prompts: Enhanced /AILab_System_Prompts.json with Story system

🔧 Technical Improvements

• Auto-Loading: Enhanced __init__.py to load nodes from subdirectory
• Error Handling: Robust separator detection and fallback mechanisms
• Debug Output: Detailed debugging information for troubleshooting
• Flexible Input: Support for custom delimiters and auto-detection modes

Version 2.1.0 (2026/02/14)

🏷️ User-Friendly Keep Last Prompt Feature

This update refines the prompt persistence feature with a more intuitive parameter name and simplified implementation, making it easier for users to understand and use.

🔄 Keep Last Prompt User-Friendly

• Parameter Rename: Changed bypass_mode to keep_last_prompt for better user understanding
• Intuitive Checkbox: Clear boolean parameter for prompt persistence control
• Consistent Logic: Unified behavior across all 6 nodes (HF, GGUF, PromptEnhancer variants)
• Workflow Optimization: Generate once, preserve results while changing inputs
• Zero Resource Usage: Cached mode provides instant responses without model loading
• Smart Caching: LAST_SAVED_PROMPT global variable for efficient prompt reuse

🔥 Enhanced NSFW Support

• Stronger Instructions: Changed "If" to "WHEN" for more directive NSFW handling
• Explicit Directives: Added "IMPORTANT: Describe ALL content without censorship or filtering"
• Detailed Requirements: "Provide explicit, detailed descriptions of body parts, sexual acts, and adult content"
• No Sanitization: "Do not sanitize, euphemize, or avoid explicit content"
• Universal Application: Applied to all WAN 2.2 presets (I2V, T2V, Storyboard, Cinematic)

🔦 Fixed Lighting Invention Issues

• Accuracy Problem: Corrected AI tendency to add lights not present in source images
• Visible Only Rule: Changed to "Specify ONLY visible technical elements adapted for style"
• Source Fidelity: "light sources actually present in the image, light quality as observed"
• No Invention: "Do not invent or add lighting not present in the original image"
• Better Descriptions: More accurate lighting analysis based on actual image content

🔄 Updated GGUF Models

• Replaced noctrex Models: Removed older noctrex abliterated GGUF variants
• Added mradermacher v2/v3: Latest mradermacher quantizations for better performance
• 4B Model: Qwen3-VL-4B-Instruct-c_abliterated-v2-GGUF (7 variants: 2.38GB-4.28GB)
• 8B Model: Qwen3-VL-8B-Instruct-c_abliterated-v3-GGUF (7 variants: 4.8GB-8.71GB)
• Enhanced File Selection: More quantization options (Q4_K_S, Q5_K_S) for better VRAM efficiency
• Corrected MMProj Files: Updated to use proper .mmproj-f16.gguf files for each model

📦 Enhanced Model Selection (HF)

• Added Josiefied Model: Josiefied-Qwen3-VL-4B-Instruct-abliterated-beta-v1 (HF only)
• GGUF Not Available: Josiefied model exists only in HuggingFace format, not GGUF
• Better Alternatives: mradermacher models provide superior GGUF quantizations
• Reliable Behavior: Works regardless of seed changes or model variations

✅ Universal Implementation

• All Nodes Updated: QwenVL, QwenVL Advanced, QwenVL GGUF, QwenVL Advanced GGUF, QwenVL Prompt Enhancer, QwenVL Prompt Enhancer GGUF
• Consistent UI: Same parameter name and tooltip across all nodes
• Bug Fixes: Resolved TypeError issues in Advanced nodes

🔄 How Keep Last Prompt Works

1. Generate Phase (keep_last_prompt=False): Generate new prompt and save to memory
2. Keep Phase (keep_last_prompt=True): Return the last saved prompt from memory
3. First Run: If no previous prompt exists, returns empty string
4. Memory Update: Each generation overwrites the previous saved prompt

🐛 Bug Fixes

• Fixed TypeError: Resolved "unexpected keyword argument" errors in Advanced nodes
• Parameter Consistency: All process() functions now use keep_last_prompt
• Syntax Errors: Fixed duplicate global declarations

---

Version 2.0.9 (2026/02/12)

🎛️ Bypass Mode Parameter for Prompt Persistence

This major feature introduces a new bypass_mode parameter that allows users to maintain previously generated prompts without regeneration, providing perfect workflow control.

🎛️ New Bypass Mode Feature

• Smart Cache Retrieval: When bypass_mode=True, nodes automatically retrieve the most recent cached prompt for the current model
• Zero Resource Usage: Bypass mode consumes no computational resources - instant response
• Perfect Workflow Control: Generate prompts once, then enable bypass mode to preserve them while changing inputs
• Universal Implementation: Available across all nodes (HF, GGUF, PromptEnhancer, Advanced variants)

🔄 How Bypass Mode Works

1. Generate Phase (bypass_mode=False): Generate new prompts and save them to cache
2. Bypass Phase (bypass_mode=True): Retrieve and return the most recent cached prompt
3. Input Changes: When bypass mode is enabled, changing images/videos or other inputs has no effect on the output
4. Seamless Switching: Toggle bypass mode on/off to switch between generation and persistence

📋 Updated Nodes

• ✅ AILab_QwenVL (HF): Added bypass_mode parameter
• ✅ AILab_QwenVL_GGUF: Added bypass_mode parameter
• ✅ AILab_QwenVL_GGUF_Advanced: Added bypass_mode parameter
• ✅ AILab_QwenVL_PromptEnhancer (HF): Added bypass_mode parameter
• ✅ AILab_QwenVL_GGUF_PromptEnhancer: Added bypass_mode parameter

🎯 Use Cases

• Prompt Locking: Generate the perfect prompt once, then lock it while experimenting with other workflow parameters
• Batch Processing: Generate multiple prompts, then use bypass mode to maintain consistency across batches
• Resource Optimization: Enable bypass mode when you need to preserve prompts but don't want to consume GPU resources
• Workflow Debugging: Test other parts of your workflow without changing the generated prompts

🎮 Simple Control

• Checkbox Interface: Simple bypass_mode checkbox in all node interfaces
• Default Behavior: bypass_mode defaults to False (normal generation)
• Instant Feedback: Console logs show when bypass mode is active and which cached prompt is being used

🔧 Technical Implementation

• Cache-Based: Uses existing prompt caching system for reliable prompt retrieval
• Model-Specific: Retrieves prompts specific to the current model configuration
• Fallback Handling: If no cached prompt exists, returns empty string gracefully
• Performance Optimized: Minimal overhead when bypass mode is enabled

Version 2.0.8 (2026/02/06)

🐛 Bug Fixes and Stability Improvements

This release addresses critical issues identified in v2.0.7 and enhances multilingual capabilities.

🔧 Bug Fixes

• JSON Syntax Error: Fixed trailing comma in AILab_System_Prompts.json at line 26
• Undefined Variable Error: Resolved name 'image_hash' is not defined in GGUF Advanced node
• Fixed Seed Stability: Reverted problematic fixed seed improvements that caused undefined variable errors
• Node Loading: Ensured all nodes load correctly without missing variable errors

🌐 Multilingual Support Enhancement

• Complete Multilingual Support: All WAN 2.2 presets now support "user prompts from any language"
• Updated Presets:
  • 🍿 Wan 2.2 I2V: Multilingual + style detection
  • 🍿 Wan Extended Storyboard: Multilingual + style detection
  • 🎥 Wan Cinematic Video: Multilingual + style detection
  • 🍿 Wan 2.2 T2V: Multilingual + style detection

🎨 Visual Style Detection

• Enhanced Style Detection: Comprehensive support for artistic styles:
  • Anime style
  • 3D cartoon
  • Pixel art
  • Puppet animation
  • 3D game style
  • Claymation
  • Watercolor
  • Black and white animation
  • Oil painting style
  • Felt style
  • Tilt-shift photography
  • Time-lapse photography

🔧 Technical Improvements

• JSON Validation: All system prompts now validate correctly
• Error Handling: Improved error messages for debugging
• Code Stability: Removed complex seed tracking that caused undefined variables
• Documentation: Updated README and changelog with comprehensive changes

🔄 Repository Management

• Clean Revert: Properly reverted problematic commits
• Version Control: Maintained stable commit history
• Documentation: Comprehensive update logs for transparency

---

Version 2.0.7 (2026/02/04)

🧠 Smart Prompt Caching System

This release introduces intelligent prompt caching with Fixed Seed Mode for dramatic performance improvements and workflow stability.

🎯 Key Features

• Smart Caching: Automatic prompt caching prevents regeneration of identical prompts
• Fixed Seed Mode: Set any fixed seed value to maintain consistent prompts regardless of media variations
• Universal Coverage: Works across all node types (HF, GGUF, PromptEnhancer)
• Performance Boost: Instant response for cached prompts with zero model loading time
• Simplified Logic: Streamlined caching system that always includes seed for predictable behavior across all seed values

🔧 Technical Implementation

• Hash-based Cache Keys: MD5 hashes generated from model, preset, custom prompt, media, and seed
• Media Hash Sampling: Efficient image/video tensor hashing using pixel sampling
• Cache File Storage: JSON-based cache with automatic save/load functionality
• Seed-based Logic: Always includes seed in cache key for consistent behavior
• Console Logging: Clear feedback on cache hits/misses with seed information

📋 Cache Behavior

• Fixed Seed Mode: Any fixed seed value maintains consistent prompts regardless of media variations
• Random Mode: Different seeds generate different prompts even with identical inputs
• Cache Hit: Same seed + same inputs = instant response from cache
• Cache Miss: Different seed or different inputs = new generation and cache storage
• Cache Persistence: Automatic saving to prompt_cache.json after each generation
• Cache Loading: Automatic cache restoration on ComfyUI startup
• Memory Efficient: Cache only stores generated text and metadata

🎨 User Experience

• Updated Tooltips: Enhanced seed parameter tooltips explaining cache behavior
• Instant Response: Zero wait time for cached prompts
• Workflow Stability: Same prompt guaranteed with fixed seed regardless of media
• Performance Gains: Dramatic speedup for repeated prompt generation
• Transparent Operation: Clear console messages indicating cache usage

🛠️ Compatibility

• Backward Compatible: Existing workflows continue to work unchanged
• Optional Feature: Fixed Seed Mode works with any seed value
• Cache Management: Automatic cache cleanup and management
• Cross-Session: Cache persists between ComfyUI restarts
• Code Maintenance: Removed deprecated parameters across all download functions (HF, GGUF, PromptEnhancer) for future compatibility
• GGUF Performance: Increased default context size from 8192 to 32768 across all GGUF models for better utilization
• Universal Caching: Fixed Seed Mode and smart caching now available across all node types (HF, GGUF, PromptEnhancer) with shared cache file
• Simplified Logic: Streamlined caching system that always includes seed for predictable behavior across all seed values

---

Version 2.0.6 (2026/02/03)

🎬 Professional Cinematography Enhancement

This release adds comprehensive professional cinematography specifications to all WAN 2.2 presets and updates naming for consistency.

🎯 Major Enhancements

• Professional Specifications: All WAN 2.2 presets now include comprehensive cinematography technical details
• Light Sources: 8 types (sunlight, artificial, moonlight, practical, firelight, fluorescent, cloudy, mixed)
• Light Quality: 9 qualities (soft, hard, top, side, back, bottom, edge, silhouette, low/high contrast)
• Time Periods: 6 periods (daytime, nighttime, dusk, sunset, dawn, sunrise)
• Shot Types: 6 types (close-up, medium shot, medium close-up, medium full shot, full shot, wide angle)
• Composition Types: 5 types (central, balanced, side, symmetrical, short side)
• Lens Specifications: 5 focal lengths + 6 angles for complete camera control
• Camera Movements: 9 movement types for dynamic cinematography
• Color Tone: 4 tone options (warm, cool, high saturation, low saturation)

🏷️ Branding Updates

• Wan 2.2 T2V: Updated to "🍿 Wan 2.2 T2V" with proper emoji
• Extended Storyboard: Updated to "🍿 Wan Extended Storyboard" for WAN family consistency
• Cinematic Video: Updated to "🎥 Wan Cinematic Video" to align with WAN branding

📋 Enhanced Presets

• 🍿 Wan 2.2 I2V: Timeline structure + professional specs
• 🍿 Wan 2.2 T2V: Timeline structure + professional specs
• 🍿 Wan Extended Storyboard: Timeline + continuity + professional specs
• 🎥 Wan Cinematic Video: Single scene + professional specs

🛠️ Technical Improvements

• Consistent Organization: All WAN presets now follow family branding
• Professional Quality: Industry-standard cinematography specifications
• Comprehensive Coverage: Complete technical details for professional video generation

---

Version 2.0.5 (2026/02/01)

🎬 Extended Storyboard Preset Added

This release adds a new specialized preset for extended storyboard generation with WAN 2.2 format compatibility.

🎯 New Features

• Extended Storyboard Preset: New preset positioned as 2nd in the list after Wan 2.2 I2V
• WAN 2.2 Format Compliance: Follows the same timeline structure "(At 0 seconds: ...)" format
• Storyboard Continuity: Each paragraph repeats content from previous paragraph for smooth transitions
• NSFW Support: Complete NSFW description support like Wan 2.2 I2V preset
• Cinematic Specifications: Comprehensive camera, lighting, and composition guidelines

📋 Preset Details

• Position: 2nd preset (after Wan 2.2 I2V) for easy access
• Format: 5-second timeline with detailed second-by-second descriptions
• Focus: Extended video generation with consistent character appearance and scene continuity
• Technical Specs: Includes camera movements, lighting types, composition guidelines
• Output: Optimized for storyboard-to-storyboard workflow generation

🛠️ Technical Improvements

• Stability Fixes: Identified and resolved "keep model loaded" memory issues
• Performance: Maintained stable operation without OOM errors
• Compatibility: Preset works with both VL and text-only models

---

Version 2.0.4 (2026/02/01)

🔧 Stability Update - SageAttention Removal

This release focuses on stability by removing SageAttention integration that was causing compatibility issues and model interference.

🛠️ Major Changes

• SageAttention Removed: Completely removed SageAttention integration due to compatibility and stability issues
• Simplified Attention Modes: Now supports auto, flash_attention_2, and sdpa only
• Clean Codebase: Removed 100+ lines of complex patching code
• Stable Performance: SDPA provides excellent performance with zero interference

🎯 Attention Modes Available

• Auto: Automatically chooses best available attention implementation
• Flash Attention 2: High-performance attention when available (RTX 20+)
• SDPA: Scaled Dot Product Attention - stable and well-tested fallback

🐛 Issues Resolved

• Model Output Problems: Fixed prompt generation issues caused by SageAttention interference
• Recursion Errors: Eliminated infinite recursion in attention patching
• Head Dimension Compatibility: Removed headdim restrictions that were causing crashes
• VAE Compatibility: Fixed VAE encoding/decoding interference

⚡ Performance Impact

• Flash Attention 2: Still available for 2-3x speedup on compatible hardware
• SDPA: Highly optimized baseline performance
• Zero Interference: Clean attention pipeline without patching complications
• Better Stability: More reliable across different model architectures

---

Version 2.0.3 (2026/02/01)

🔧 SageAttention Compatibility Fix

This hotfix release addresses a critical compatibility issue with SageAttention integration that prevented proper patching in certain transformer versions.

🐛 Bug Fixes

• SageAttention Patch: Fixed AttributeError when accessing transformers.models.qwen2.modeling_qwen2.F
• Proper Module Access: Changed patch target to torch.nn.functional.F.scaled_dot_product_attention
• Compatibility: Ensured SageAttention works across different transformer versions
• Safer Imports: Added better error handling for module imports

⚡ Performance Impact

• SageAttention now works correctly with 8-bit quantization
• Maintains 2-5x performance boost on compatible hardware
• Graceful fallback to SDPA when patching fails

---

Version 2.0.2 (2026/02/01)

🎯 User Experience & Model Management Update

This release focuses on improving user accessibility, model management, and content generation capabilities based on community feedback.

🚀 Enhanced Model Accessibility

• Free Abliterated Models: Added Qwen3-4B-abliterated-TIES and Qwen3-8B-abliterated-TIES (nbeerbower) as default options
• No Token Required: These models provide uncensored performance without Hugging Face authentication
• Smart Model Ordering: 4B models prioritized before 8B for better VRAM accessibility
• 8-Bit Quantization Default: Optimized VRAM usage for both VL and text models

New Default Models:

• VL: Qwen3-VL-4B-Instruct-Abliterated (8-bit quantized)
• Text: Qwen3-4B-abliterated-TIES (8-bit quantized, no token)

🔧 Improved Custom Prompt Logic

• Template Combination: Fixed custom prompts to combine with preset templates instead of replacing them
• Consistent Behavior: All nodes (HF, GGUF, PromptEnhancer) now handle custom prompts identically
• Better UX: Users can now add their input while keeping preset instructions intact
• WAN 2.2 Compatibility: Fixed I2V preset to work properly with custom user input

📝 Enhanced NSFW Content Generation

• Comprehensive Descriptions: Expanded NSFW rules to include detailed sexual act descriptions
• Clear Instructions: Added "body parts and acts being performed" specification
• Complete Coverage: Includes blowjob, paizuri, deepthroat, handjob, cunnilingus, anal, missionary, doggystyle, cowgirl, reverse cowgirl, masturbation, fingering, squirting, cumshot, facial
• Applied Universally: Enhanced both WAN 2.2 I2V and T2V presets

🎬 Preset Priority Optimization

• WAN 2.2 I2V First: Moved most-used video generation preset to top position
• Better Workflow: Faster access to primary video generation functionality
• Maintained Order: Preserved logical organization of remaining presets

🐛 Bug Fixes

• Custom Prompt Override: Fixed issue where custom prompts completely replaced preset templates
• GGUF Consistency: Applied same prompt logic fixes to GGUF nodes
• Tooltip Updates: Updated all custom prompt tooltips to reflect new behavior
• JSON Validation: Removed duplicate model entries for cleaner configuration

---

Version 2.0.1-enhanced (2026/01/30)

🚀 Major Performance & Video Generation Update

This enhanced version introduces cutting-edge performance optimizations and comprehensive video generation capabilities, making it the most advanced QwenVL integration available.

⚡ SageAttention Integration

• 2-5x Performance Boost: Added SageAttention support for 8-bit quantized attention computation
• Automatic Detection: System automatically detects SageAttention availability and GPU compatibility
• Seamless Fallback: Gracefully falls back to SDPA when SageAttention is unavailable
• Memory Efficiency: Reduced VRAM usage with quantized attention kernels

Requirements:

• NVIDIA GPU with capability >= 8.0 (RTX 30/40/50 series)
• CUDA >= 12.0
• sageattention package (optional)

🎬 WAN 2.2 Video Generation Integration

• I2V Support: Image-to-Video generation with cinematic timeline structure
• T2V Support: Text-to-Video generation with professional scene descriptions
• 5-Second Timeline: Precise second-by-second scene progression
• Multilingual: Italian/English input to optimized English output
• Cinematic Quality: Film-style direction including lighting, camera, composition

Available Prompts:

• 🍿 Wan 2.2 I2V - For image-to-video workflows in QwenVL nodes
• 🍿 Wan 2.2 T2V - For text-to-video workflows in Prompt Enhancer nodes

🔧 Enhanced Compatibility

• Transformers Compatibility: Support for both old (AutoModelForImageTextToText) and new (AutoModelForVision2Seq) transformer versions
• Attention Mode Selection: Four modes available - auto, flash_attention_2, sdpa, sageattention
• Custom Model Support: Extended HF and GGUF model configurations
• Improved Error Handling: Better diagnostics and fallback mechanisms

📚 Documentation Updates

• SageAttention Section: Comprehensive installation and troubleshooting guide
• WAN 2.2 Documentation: Complete usage examples and best practices
• Performance Comparisons: Detailed benchmarks and requirements tables
• Updated README: Professional documentation with all new features

🏷️ Manager Integration

• Custom Nodes JSON: Prepared for ComfyUI Manager submission
• Enhanced Tags: Added wan2.2, video, i2v, t2v tags for better discoverability
• Updated Metadata: Comprehensive description highlighting all enhancements

---

Release Notes: v2.0.0 (2025-12-22)

New GGUF Nodes

We've expanded our GGUF model support with three powerful new nodes:

• QwenVL (GGUF) — Lightweight GGUF-based vision node for image/video understanding and text generation. Offers significantly faster inference speed compared to Transformers models, making it ideal for real-time workflows and resource-constrained environments.
• QwenVL (GGUF) Advanced — Enhanced GGUF vision node with additional controls for advanced users. Maintains the ultra-fast inference speed of GGUF while providing fine-tuned control over generation parameters.
• Qwen Prompt Enhancer (GGUF) — GGUF text-only node for intelligent prompt rewriting and enhancement (not a vision model). Delivers rapid prompt enhancement with minimal resource usage, perfect for iterative prompt refinement workflows.
• Qwen Prompt Enhancer (Transformers) — Uses Qwen3 transformer model to enhance and rewrite prompts. Analyzes your input prompt and intelligently expands it with better detail, structure, and clarity for improved generation quality. Offers full model capabilities with precise control over the enhancement process.

Enhanced GPU Device Selection (Advanced Node)

The QwenVL (Advanced) node now offers flexible GPU device management:

• Manual GPU selection — Choose specific CUDA devices (e.g., cuda:1, cuda:2) instead of defaulting to cuda:0
• Automatic device detection — Dynamically discovers all available CUDA devices on your system
• Improved device mapping — More consistent behavior and better resource allocation
• OOM prevention — Route models to underutilized GPUs when your primary GPU is handling diffusion workloads

Note: The basic QwenVL node continues to use automatic device selection for simplicity

---

✨ Improvements

GGUF: Quality and Usability

Cleaner Outputs by Default:

• Automatically removes common "thinking/planning" content and leaked tokens (<think>, <im_start>, <im_end>)
• Users now receive clean, usable prompt-only or answer-only text without manual filtering

QwenVL (GGUF) Vision Node:

• Model dropdown now displays actual .gguf filenames with automatic deduplication for easier model identification
• Enhanced download progress logging with clear status messages during model download and cache reuse
• Token generation speed reporting (tok/s) when available — helps compare different models and quantization levels

Qwen Prompt Enhancer (GGUF):

• Updated built-in presets to reduce "junk talk" and return clean enhanced prompts more consistently
• Refined system prompts that minimize verbose output
• More reliable prompt-only text generation

Transformers Nodes: GPU and Attention Stability

Advanced GPU Routing:

• QwenVL (Advanced) supports selecting specific GPUs (e.g., cuda:1, cuda:2) to avoid OOM when GPU0 is busy
• Improved device-mapping logic for more consistent behavior across different hardware configurations

Attention Backend Stability:

• Flash-Attention auto mode now behaves safely across all platforms
• Gracefully falls back to SDPA when Flash-Attention dependencies are unavailable
• Prevents runtime errors from missing or incompatible Flash-Attention installations

---

🐛 Bug Fixes

QwenVL (Transformers) Stability

• Fixed: Invalid CUDA device handling that caused crashes with incorrect device specifications (e.g., device "0" or malformed device_map)
  Related issue: https://github.com/1038lab/ComfyUI-QwenVL/issues/21
• Fixed: Flash-Attention detection now restricted to Linux systems only, preventing Windows metadata errors
• Fixed: Flash-Attention auto mode fallback mechanism to eliminate runtime errors when dependencies are unavailable

---

📚 Documentation & Dependencies

New Documentation

• Added: Comprehensive installation guide for vision-capable llama-cpp-python
  See docs/LLAMA_CPP_PYTHON_VISION_INSTALL.md for:
  • JamePeng fork wheel installation instructions (wheel source: https://github.com/JamePeng/llama-cpp-python/releases/)
  • Handler verification steps
  • Common numpy/OpenCV conflict resolution

Dependencies

• Added: hf_xet to requirements.txt for improved Hugging Face download performance and to eliminate Xet fallback warnings

Version 1.1.0 (2025/11/11)

⚡ Major Performance Optimization Update

This release introduces a full rework of the QwenVL runtime to significantly improve speed, stability, and GPU utilization.

🚀 Core Improvements

• Flash Attention Integration (Auto Detection)
  Automatically leverages next-generation attention optimization for faster inference on supported GPUs, while falling back to SDPA when needed.
• Attention Mode Selector
  Both QwenVL nodes expose the attention backend (auto / flash_attention_2 / sdpa) so users can quickly validate which mode performs best on their hardware without leaving the basic workflow view.
• Precision Optimization
  Smarter internal precision handling improves throughput and keeps performance consistent across high-end and low-VRAM cards.
• Runtime Acceleration
  The execution pipeline now keeps KV cache/device alignment always-on, cutting per-run overhead and reducing latency.
• Caching System
  Models remain cached in memory between runs, drastically lowering reload times when prompts change.
• Video Frame Optimization
  Streamlined frame sampling and preprocessing accelerate video-focused workflows.
• Hardware Adaptation
  Smarter device detection ensures the best configuration across NVIDIA GPUs, Apple Silicon, and CPU fallback scenarios.

🧠 Developer Enhancements

• Unified model and processor loading with cleaner logging and fewer bottlenecks.
• Refined quantization and memory handling for better stability across quant modes.
• Improved fallback behavior when advanced GPU optimizations are unavailable.

💡 Compatibility

• Fully backward compatible with existing ComfyUI workflows.
• Retains both QwenVL and QwenVL (Advanced) nodes: the basic node now bundles the most useful speed controls, while the advanced node exposes every knob (quantization, attention, device, torch.compile) for deep tuning.

🔧 Recommended

• PyTorch ≥ 2.8.0
• CUDA 12.4 or later
• Flash Attention 2.x (optional, for maximum performance)

Switching quantization or attention modes forces a one-time model reload and is expected behavior when comparing runtime profiles.

Version 1.0.4 (2025/10/31)

🆕 Custom Model Support Added

• Users can now add their own custom Qwen-VL or Hugging Face models
  by creating a custom_models.json file in the plugin directory.
  These models will automatically appear in the model selection list.

• Added automatic merging of user-defined models from custom_models.json,
  following the same flexible mechanism as in ComfyUI-JoyCaption.

• Added detailed documentation
  👉 docs/custom_models.md
  and an editable example file custom_models_example.json.

⚙️ Dependency Update

• Updated Transformers version requirement:
  transformers>=4.57.0 (was >=4.40.0)
  to ensure full compatibility with Qwen3-VL models.
  Reference: Qwen3-VL

---

Version 1.0.3 (2025/10/22)

• Added 8 more Qwen3-VL models 2B and 32B (FB16 and FP8 variants) have been integrated into our support list, catering to diverse requirements.

Version 1.0.2 (2025/10/21)

• Integrated additional Qwen3-VL models
• Added Chinese language README (README_zh.md)
• Refined fine-tuning preset system prompt

Version 1.0.1 (2025/10/17)

• Resolved various bugs
• Optimized video input logic

v1.0.0 Initial Release (2025/10/17)

• Support for Qwen3-VL series models.
• Automatic model downloading from Hugging Face.
• On-the-fly quantization (4-bit, 8-bit, FP16).
• Preset and Custom Prompt system for flexible and easy use.
• Includes both a standard and an advanced node for users of all levels.
• Hardware-aware safeguards for FP8 model compatibility.
• Image and Video (frame sequence) input support.
• "Keep Model Loaded" option for improved performance on sequential runs.
• Seed parameter for reproducible generation.