H3 Face Stitch Back

H3 Face Stitch Back is the final node in the ComfyUI-H3-FaceRefine pipeline: it takes the crops that MiniMax H3 re-generated, warps each one back onto the exact spot the face came from, colour-matches it, feathers the seam, and composites it into the original frame. If the tracker is the spine of this pack, the stitch is the part the eye actually judges - a perfect refine with a bad paste reads as a pasted-on face, and it's the stitch that decides.

The pack exists because H3 renders distant heads as smears. The fix is the classic detailer loop Impact Pack made famous: detect, crop so the face fills the canvas, let the model re-render, paste back. This node is the paste step, and it inherits two important ideas from FaceDetailer along the way.

How it works

The clever part is that it's one batched grid_sample - every refined crop is warped back onto its float-precision box in a single GPU tensor op. That matters because the tracker smoothed its trajectory to sub-pixel accuracy, and a naive paste would re-quantize it, re-introducing exactly the jitter the whole pipeline exists to remove. Then a per-channel mean/std match brings the crop's brightness and tint in line with the region it replaces, the paste mask is blurred by feather, and the composite happens.

The design principle to internalise: only the face region composites. The wide crop exists to give the sampler context, not to be pasted. Pasting the whole crop covers ~88% of the canvas versus ~16% for the face box, and anything H3 changed in the hair or background comes back as a rectangle.

Inputs that matter

• base_images - the original frames, the same clip you fed to the tracker.
• refined_crops - the VAEDecode output of your refined crops.
• transform - from H3 Face Track + Crop; it records where every crop came from.
• paste_region - face_only (default) or face_ellipse paste just the detected face; full_crop risks a visible rectangle.
• feather (6) - blur radius on the paste mask in source pixels. This is the trap input: if it were canvas-relative, the blend would come out tightest exactly where the face is smallest - a hard edge appearing as a shot zooms out. Use ~24 with a rect mask, 4–8 with a SAM mask.
• colour_match (1.0) - without it the refined face comes back subtly brighter and reads as pasted on. Dial down if it over-corrects.
• blend (1.0) - global opacity of the refined face; below 1.0 mixes back toward the original, handy for dialing back over-sharpening.
• undetected_frames (fade_out) - what happens where no face was found. All frames still run through H3 either way (that's what keeps it temporally consistent); this only controls pasting. fade_out ramps the composite to zero, skip hard-cuts, composite_anyway risks H3 hallucinating a face onto the back of a head.
• masks (optional) - per-frame masks from H3 Face Mask (SAM), which override paste_region.

Output is a single images tensor - the finished frames, straight to your video save node.

Install and gotchas

The pack is one clone into custom_nodes/ (or search "ComfyUI-H3-FaceRefine" in ComfyUI Manager), then restart. For this node you need the transform wiring intact and, to run the actual pipeline, a face detector (face_yolov8m.pt in models/ultralytics/bbox/) plus the usual H3 model stack. The stitch itself needs nothing extra.

Where people get burned: frame count mismatch. If base_images, refined_crops and the transform disagree on length, the node truncates to the shortest and prints a warning - your clip's frame count must sit on H3's 17k+5 grid (5, 22, 39 …), which clips generated by H3 already do. And if the face looks pasted on, that's colour_match working or not working - check it's at 1.0 before you blame the geometry.