MiniMax H3 Motion Director

v1.12.0 what's in itFixed<ul><li>The "gets laggier over time" UI freeze. The per-node audio-role panel rebuilt its DOM from <code>onDrawBackground</code> (runs every canvas frame) and re-scheduled itself on every rebuild → measured 4,613 mutation batches / 13,950 DOM mutations per second, one browser core pinned at 96 %, whole ComfyUI UI queued behind it. Now signature-gated rebuilds, 250 ms draw-hook floor, timer ladder coalesced → 4 batches / 99 mutations per second.</li><li>Two document-wide observers stopped scanning the whole page per mutation batch (report cards + SAM note): filtered to their own DOM, one coalesced sweep per frame with a 400 ms floor and a 2 s safety sweep, per-card attribute watching, no work in a hidden tab. Node polls skip when the tab is hidden or the node is gone.</li></ul>Added<ul><li>VRAM fit check in Validate — <code>ok</code> / <code>tight</code> / <code>over</code> per segment with numbers and the concrete fix (reference size → frame cap → canvas); warning only, never a blocker. New <code>vram_budget.py</code>.</li><li>Post-OOM retry buttons — "Retry S4 at 243 frames" / "Set reference size to 1024 px and retry". New <code>minimax_oom_retry.mjs</code>.</li><li>Red <code>missing asset</code> chips are clickable — relink that one mention from the assets the prompt can actually name.</li><li><code>drift_report.py</code> — chain drift report (seams, colour, detail, noise, structure fingerprint), CPU-only, safe during a render, seam judged against the incoming segment's own frame step.</li></ul>Changed<ul><li>All four CUDA OOM paths report identically via <code>vram_budget.oom_message</code> (segment, sampled frames, tokens, attention workspace, free memory).</li></ul>v1.11.0 what's in itA release about the things you set once. The whole pack now has a Settings panel (machine profile, backend awareness, baselines, run defaults, cache manager, diagnostics), exported videos stop carrying your project unless you ask them to, and a reference added from the Material Library finally reconnects to the prompt mention that names it.v1.10.0 what's in itNew capabilities<ul><li>Generated rows in a Character Replace chain — a row can be a generate row ("kind": "generate"): no source window, renders from its own prompt/refs, mid-chain or past the end of the footage; joins the previous segment as context (r2v) or as a literal first frame (i2v). Panel labels it G vs W, keeps mask controls off it, shows an n generated badge, and Long-form re-cut keeps those rows.</li><li>Story to segments (was already in the code, now usable): one brief → a shared world paragraph + one beat per segment, each beat then enhanced and reviewed before anything is written.</li><li>Story to segments is reachable in t2v / i2v / r2v and Long-form — new button in the batch and fl2v panels (the prompt rows that opened it were hidden in exactly those modes); a prompt batch may create the cards it needs, at 17k+5 frames / 24 fps.</li><li>Regenerate one prompt — or all — in the review list — re-runs from that row's own original text, task and block, so fixing one segment doesn't cost a 10-pass batch.</li><li>Your own prompt recipes in a file a pack update can't overwrite (recipes.json) + recipe_templates starters, and a Browse… to load them.</li></ul>Behaviour changes<ul><li>Vision frames follow the segment, not the file — sampled inside the window (was: the same three moments of the whole video for every window).</li><li>Vision frames (1-5) is a setting, also capping RefMod decode frames.</li><li>Movement pairs from 4 frames on — two frames a few frames apart, because stills can't show a movement.</li><li>A replace window keeps your own words about the motion (the caption can't see movement that leaves no frame difference).</li><li>Pre-flight warning when the project frame rate isn't 24 (frame_rate_not_24, with the recommendation to use a 24 fps source for replaced footage).</li></ul>Fixes<ul><li>Output stays at the model's 24 fps in an off-rate project — segment audio length, the merged export's fps stamp, Audio Drive validation/overlay and seam diagnostics were all using the project rate (h3_rate.py now owns the rule).</li><li>&lt;Picture N&gt; replace windows hold the source pose (the pose line existed only in the RefMod variant — a live run reached over the back instead of under it).</li><li>Enhancing the same window twice no longer nests the motion note.</li><li>The action caption describes the action first (a live run returned a room and no action, cut off mid-word).</li><li>A caption that's really the model's own notes is never pasted into the block.</li><li>The motion note no longer duplicates a whole structured prompt.</li><li>A stale cached .mjs can no longer disable the enhancer in silence (three layers).</li><li>Reference pictures in folders are readable; one unreadable picture no longer discards the rest.</li></ul>Added Model Prompt Enhancer (local or API)An enhancer that runs inside ComfyUIThe prompt enhancer is no longer a service beside ComfyUI. It loads a GGUF model from your own <code>models/LLM</code> folder with llama.cpp and runs inside ComfyUI's process — no Ollama, no API key, no second thing to start — and it can now build the prompt instead of rewriting it.Caption mode is a port of the external Character Remake workflow ("Masked Motion + QwenVL"), which handled character replacement reliably where prompt rewriting did not. The reason was never a better prompt: it is that the language model is never asked to write the structure. Two narrow vision calls answer two narrow questions ("what does this character look like", "what happens in this source window") and this pack's own code writes the section block around them, so every header, role line, retention rule and the discard sentence is code-owned and cannot be reordered, forgotten or invented. Identity sees only the reference images; action sees only the source frames.Nothing here changes generated frames of an existing project. The enhancer is a new path in the panel, started by a button. The SAM3 and mask-pre-flight fixes change what a newly configured replace window does, not what an already rendered segment produced. The RefMod digest only adds a cache key, and only when a RefMod is connected — a project without one keeps its caches. All cached work stays valid unless a RefMod is actually connected to it.Two behaviour changes are deliberate: the enhancer's SAM3 prompt default is now <code>the woman</code> (a window with its own prompt is untouched), and a <code>frames</code> mask with no folder is now reported as a pre-flight error instead of failing silently at render time.<ul><li>Loads any GGUF under ComfyUI's <code>LLM</code> model folders with <code>llama-cpp-python</code> and runs chat completions in-process. Nothing heavy is imported until you use it, so the panel and the catalog work on a box without llama.cpp installed.</li><li>The model list is what is actually on disk, plus the curated quants that can be fetched — the recommended Qwen3.8 27B abliterated is one click away, with its size shown before anything is downloaded and real progress while it downloads (measured from huggingface_hub's own staging directory, so it survives a panel reload).</li><li>A tight card degrades instead of failing: loading walks a VRAM ladder (full GPU offload → fewer GPU layers → smaller context) and prints why it dropped a rung, with free VRAM and model size.</li><li><code>Unload model</code> frees the VRAM for a render, and <code>Unload the model afterwards</code> does it at the end of an enhancement rather than after every retry pass.</li><li>The panel reports the backend that actually loaded (CUDA / Vulkan / Metal / CPU) and warns once when a GPU build is present but its backend never mapped into the process — the wheel built against a different CUDA major that silently runs on the CPU.</li></ul>Keep the learned-latent upscaler resident — opt-in, off by default<code>latent_upscale_cache_model</code> keeps the built upscaler on the device instead of freeing it, so a multi-segment run stops paying a checkpoint load per call.The upscale stage runs after the H3 DiT is unloaded, so an uncached call is a full <code>load → free</code> cycle stacked on the DiT's own unload and reload. A resident upscaler trades that churn for VRAM that stays held for the rest of the session, and it then shares the card with the reloaded DiT.That trade-off is why it ships off, with the note written beside the checkbox (English + 简体中文). It pays off only when you have headroom.<ul><li>Cache is a single entry keyed on <code>(checkpoint path, dtype, device)</code> — a precision or device change rebuilds.</li><li>A stale model is released before its replacement is built, so the two are never resident together.</li><li>Clearing the flag releases it. <code>clear_resident_model()</code> and <code>has_resident_model()</code> are exposed for explicit reclaim.</li><li>It changes speed, not frames: the first-pass cache is unaffected.</li></ul>Where: Global Refine → Upscale → H3 Learned Latent block, next to Device.FixedThe upscaler re-read and re-converted its checkpoint on every callEach call re-read the file from disk and ran the <code>float8 → fp16</code> conversion again — pure churn repeated on every segment. The last two decoded state dicts are now kept in CPU RAM, keyed on the path plus its size and mtime, so replacing the file on disk still invalidates the entry.The decoded dict is read-only for every caller (<code>build_model</code> reads shapes, <code>load_state_dict</code> copies), so it is handed out without copying. <code>clear_checkpoint_cache()</code> is available for file swaps.The upscaler's VRAM was never actually returnedIts <code>finally</code> block dropped the model and the input, then called <code>empty_cache()</code> while <code>mean</code> and <code>std</code> were still live device tensors. <code>empty_cache()</code> only returns blocks that are already unused, and unreachable-but-uncollected tensors keep holding their memory — the same trap the segment VRAM cleanup hit.Every device reference is now dropped and collected before emptying. This is the half that matches the original report (fp16 latent upscaling, degrading over a session, cleared by a ComfyUI restart): the pressure accumulated because the memory was never handed back.

Files with SHA256: 3f97a668bd89807af98b14c9b0e3ea403356a7af6739c23144e659a63b98abe6