One still image → a 12.25 s video with sound (voice, one spoken line, room tone) where the face does not drift. Runs on a 16 GB GPU in 8 steps.
The trick: the same still goes into first_frame AND last_frame of MiniMaxH3ImageToVideo. With a first frame only, the face turned into another person around 6-8 s; anchoring both ends pins the identity while the motion in between stays free.
- 2 ComfyUI workflows (API format), core H3 nodes only: POV with one spoken line, and an oral variant
- Canvas sized to the still's aspect ratio (no squashed portraits)
- README with the prompt rules that fixed real defects: locked camera (no push-in), no wide-open mouth, anatomy wording, male fingernails, how to get exactly one spoken line
- Optional post-processing: GPEN face restoration with temporal smoothing, RIFE 48 fps with scene-cut detection
Multi-shot version (4-5 cuts, dialogue between characters): MiniMax H3 Multi-Shot Storyboard Workflow. Adults only: use generated, clearly mature characters.
Description
v1.1 — fix + clearer README. The workflows themselves are unchanged. Fix in the optional face_restore.py script: it crashed as soon as two faces were tracked at once (a second person, a face on a TV or a poster), and the whole video was then left unrestored.
README: new FAQ. You only need the .json (core ComfyUI nodes, nothing to install). “API format” is just ComfyUI's “Export (API)” format: drag it onto the canvas, it is not App mode. The two .py files are optional scripts I run outside ComfyUI after the render; they are not custom nodes and do not go in any ComfyUI folder.