MiniMax H3 Motion Director

Fixing shot 3 of a five-shot video should not mean re-rendering shots 1, 2, 4 and 5. That's the whole reason MiniMax H3 Motion Director exists - and it's the thing most other H3 workflows in ComfyUI get wrong.

This is the headline node of the j955229/ComfyUI-MiniMax-H3-Motion-Director pack. Despite the film-studio name, it's a single node that runs the open MiniMax H3 weights locally: no API, no key, nothing to subscribe to. What you get instead of one H3 sampling chain per shot is a timeline inside the node - T2V, I2V, FL2V, R2V, V2V and RV2V segments mixed however you like, each with its own prompt, duration and media, plus Selective Run so you only regenerate the shots that actually need another pass. The author calls it "one Director, one shot to a complete multi-segment video project," and that's an honest description.

How it works

The clever bit is that the Director isn't a new model. It compiles every segment into an existing MiniMax H3 task and hands it to ComfyUI's official MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo conditioning nodes (tokenization happens internally). "Mixed" is a timeline concept, never something you send to the model.

The continuity machinery is where the real engineering went. With Motion Context enabled, each segment's final exported frames become real H3 motion history for the next one - context_length snaps to H3-valid runs of 1, 5, 22 or 39 frames. Audio Context carries the previous segment's generated audio forward, and Color Re-anchor tries to stop cumulative color drift. A Segment Result can reuse a decoded frame from an earlier segment as a later I2V start image or FL2V first/last frame. It's a lot of moving parts, but the defaults are sane and the per-boundary controls on the segment cards are optional.

Post-processing lives here too: a Global Refine pass (second sampling, optional upscale, and on supported NVIDIA hardware, RTX VSR or Deblur) and a Face Refine path that detects, crops and regenerates faces. If either stage fails, the Director keeps the first-pass result instead of discarding a finished generation - a genuinely thoughtful touch.

The inputs that matter

You won't touch most of the schema. The ones a beginner actually sets:

• model / video_vae / audio_vae / clip - feed these from ComfyUI's standard loaders: UNETLoader for the H3 UNET, the minimax_h3_video_vae, the minimax_h3_audio_vae (required for r2v / v2v / rv2v), and a CLIPLoader with type minimax (qwen3vl). Expect the model alone to be on the order of 42 GB, so this is not a light install.
• task_type - the seven-way enum: t2v, i2v, fl2v, r2v, v2v, rv2v, or mixed.
• global_prompt - the prompt, sent straight to the H3 conditioning node.
• frame_rate (24 - what H3 was trained at), width/height (864×480 default), total_frames (snaps to MiniMax's 17k+5 grid; 124 ≈ 5 seconds).
• steps (25, the official template), sampler_name (res_multistep), scheduler (simple), shift_video (12) / shift_audio (3).

The optional sampler / sigmas sockets let you swap in an external ComfyUI sampling chain instead of the internal one, and clear_vram_between_segments releases models between segments - slower, but a stability win on tight VRAM.

Outputs are exactly three and keep it friendly for downstream nodes: images (IMAGE list), audio (AUDIO list), and fps (FLOAT). The node is an OUTPUT_NODE, so it can sit at the end of a graph and still expose those to whatever follows.

Install

ComfyUI Manager: search "MiniMax H3 Motion Director". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/j955229/ComfyUI-MiniMax-H3-Motion-Director.git
cd ComfyUI-MiniMax-H3-Motion-Director
python -m pip install -r requirements.txt

Restart ComfyUI completely; if an update changes frontend files, hard-refresh the browser. Python deps are light (opencv-python-headless, imageio-ffmpeg, scenedetect), but you need a recent ComfyUI with official MiniMax H3 support and the H3 model files above. The pack also registers its own folders for face-detection (ultralytics), SAM, and latent upscale models.

Where people get burned

• Do not load the standalone ComfyUI-H3-Motion-Context pack alongside this one. Motion Context is integrated here, and running both fights each other.
• Forgot the audio VAE on an r2v/v2v/rv2v job - it won't run. It's only optional for t2v/i2v/fl2v.
• Continuity is real but not magic: the README is upfront that H3 can still drift in lighting, motion or identity across boundaries.
• The licensing asterisk: H3's community license geofences the weights out of the US, EU, UK and South Korea - outputs included. It doesn't call an API, but "runs locally" and "legal to run in your country" are different questions here.

If you're doing single-shot work, a plain H3 workflow is simpler. The Director pays for itself the moment you're stitching shots together.