MiniMax H3 Face Refine Parity Stitch / 原版机制回贴 (Advanced)
The reference-mechanism stitch, bit-exact outside the mask
- base_frames
- refined_crops
- face_plan
- candidate_frames
- changed_mask
- fallback_mask
- fallback_count
- report_json
The parity-chain counterpart to Face Refine Stitch Audit: where the regular node uses a feather ellipse and a 0.65 color match, this one reproduces the reference mechanics verbatim - centred face rectangle, a chunk-midpoint 24px source feather, and post-warp source-coordinate colour match. Same idea as the practical stitch (paste the refined crops back, audit everything you didn't touch), but built so the result is directly comparable to the audited upstream recipe.
Mechanism details that define this node: paste_region defaults to face_only (a centred face rectangle, not an ellipse), mask_dilation (24) and feather_source_px (24) come from the reference's 24/24 recipe, and colour_match defaults to a full 1.0 - the reference applies it at full strength, and the parity chain keeps it that way. blend (1) scales the overall paste. If a frame's face change exceeds max_face_mean_abs_delta (1.0 here - deliberately loose, because the reference mechanism trusts the geometry), the frame's undetected_frames behavior kicks in, defaulting to fade_out.
The two things it guarantees: pixels outside the mask are audited bit-exact (nothing outside the face region changed, provably), and audio is never touched. It's a non-destructive review candidate - it doesn't overwrite your source or auto-accept anything.
Inputs: base_frames, refined_crops, and a face_plan of the parity type. Outputs: candidate_frames, changed_mask, fallback_mask, fallback_count, report_json. Like the practical stitch it's an output node - it ends a branch and hands you something to look at.
The honest caveat applies here more than anywhere in the chain: this node reproduces the recipe, and reproducing a recipe is not the same as reproducing quality. The mask geometry and feathering are audited; whether the reference mechanism actually beats the practical path is exactly the question this parity chain exists to answer, on your footage, with your eyes.
Use it as the sink for the parity chain: Parity Plan → Parity Latent → Per-Frame Denoise → this → Quality Gate → the MANUAL512 baseline for formal acceptance. If you don't care about reference comparison, the regular Stitch Audit is the one you'd actually ship with.
Install: ComfyUI Manager → "MiniMax H3 Audio T8", or:
cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8
Restart after. No forced pip deps, no model downloads - pure CPU image work at the end of the chain.
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| base_frames | IMAGE | — | |
| refined_crops | IMAGE | — | |
| face_plan | H3_T8_FACE_REFINE_PARITY_PLAN | — | |
| paste_region | COMBO | face_only | 3 options: face_only, face_ellipse, full_crop_exp |
| mask_dilation | INT | 240–256 | — |
| feather_source_px | FLOAT | 24.00–256 | — |
| colour_match | FLOAT | 1.000–1 | — |
| blend | FLOAT | 1.000–1 | — |
| undetected_frames | COMBO | fade_out | 3 options: fade_out, skip, composite_anyway |
| max_face_mean_abs_delta | FLOAT | 1.000.01–1 | — |
| processing_device | COMBO | cpu_memory_safe | 2 options: cpu_memory_safe, cuda_if_available |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| candidate_frames | IMAGE | — |
| changed_mask | MASK | — |
| fallback_mask | MASK | — |
| fallback_count | INT | — |
| report_json | STRING | — |