Fixing the tiny faces MiniMax H3 never wanted you to see

What it does

MiniMax H3 renders faces badly when the head occupies a small fraction of the frame — a property of head-size-in-frame, not output resolution, so it persists at 720p and above. This pack detects the face on every frame, crops it so it fills the canvas, lets H3 re-generate it, and composites the result back into the original video. It’s Impact Pack’s FaceDetailer, adapted from stills to video.

The interesting bit

The hard part isn’t the refinement, it’s the tracking: the crop box has to sit still on a moving subject or the refined face “boils.” The pack leans on hard-cut detection (via PySceneDetect, fed from the decode pass already running) so smoothing, interpolation, and composite fades run per shot instead of being dragged across a cut. And when the subject’s face is lost — head turned, face covered — the tracker now refuses to jump to the nearest bystander; it interpolates across the gap and fades the composite out, keeping the original pixels.

Key highlights

• Per-frame face tracking with crop → H3 re-generation → stitch-back, modelled on FaceDetailer
• H3 Load Video + Face Select finds faces and hard cuts in one pass, with an interactive “Pick faces” UI for choosing the subject per shot
• Nine face-selection modes (largest_face, closest_to_xy, detector_score, etc.), with old workflows migrating on load
• Identity matching via InsightFace, CLIP Vision, or CCIP — the latter two for anime, where face_yolov8m.pt may not find faces at all
• Cut detection is a soft dependency: if PySceneDetect is missing, the run continues treating the video as one shot

Caveats

• Version 1.1.0 breaks workflows saved against 1.0.0 — H3 Per-Frame Denoise now requires a MODEL input routed through it
• You must supply your own face detector; the example workflows additionally need Impact Pack, VideoHelperSuite, and a third-party audio-lock node just to queue
• The README warns that installing onnxruntime-gpu alongside onnxruntime silently forces identity matching onto CPU

Verdict

If you’re generating H3 video with distant subjects and mushy faces, this is a focused, thoughtfully-engineered fix. If you don’t use MiniMax H3 in ComfyUI, it’s a well-documented case study in video face refinement and nothing more.