MiniMax H3 Multi-Shot Storyboard Workflow (Ref2VA, 16GB) - v1.3 (writing rules + adaptive face restore)
NSFW
One MiniMax H3 generation = a 12.25 s multi-shot video: 4-5 shots with hard cuts, up to 3 characters and lip-synced dialogue, rendered in ONE pass so faces, room and lighting stay consistent across cuts. Runs on a 16 GB GPU.
- ComfyUI workflow (API format), core H3 nodes only (Reference-to-Video, 3 references at full size)
- Structured multi-shot prompt template + a complete real example
- Post-processing scripts: GPEN face restoration with temporal smoothing, and RIFE 48 fps with scene-cut detection (no ghost frames on cuts)
- README with the editing rules that fixed lip sync, eyelines and teeth
Full method: Multi-shot AI videos in one 12-second generation. Adults only: every character is generated and age-checked.
Description
Workflow .json unchanged from v1.2. This update is about the writing rules and the optional post-processing script.
- Adaptive face restore (optional script): stronger blend on full-frame faces, following your own setting. Measured on one close-up shot: about x2.6 more apparent detail instead of x1.6, but more flicker (+18 % frame-to-frame change instead of +10 %); lower FACE_FORCE if it bothers you.
- One shared description for the reference still and
subject_definitions(her face changed between cuts when they disagreed; small sample so far). - Avoid on-screen text: H3 garbles writing, so inserts show the object and a voice says the content. Template updated.
- Strict timing rules written out, and how to fix an overflowing line.
Core ComfyUI nodes only, nothing to install; the .py files are optional and run outside ComfyUI (see the README FAQ).