MiniMax H3 Face Refine Parity Plan / 原版机制规划 (Advanced)

The "Parity" family of face-refine nodes exists for a specific, slightly obsessive purpose: reproducing the original face-refine mechanism - the reference implementation the author audited - so you can build an A/B comparison instead of trusting vibes. This node is the planning half of that chain, and its whole personality is "the exact recipe, nothing else."

Where the regular Face Refine Plan node optimizes for practical results, this one optimizes for matching upstream mechanics. Read the description: "audited BGR YOLO input, separate 21/51-frame Gaussian center/size smoothing, per-frame float crops and the author's centred FaceDetailer-style mask geometry." Those numbers matter. center_smooth_window (21) and size_smooth_window (51) are deliberately different - center drift is smoothed fast, box size slower - because that's what the reference did. crop_factor defaults to 3 (the manual-acceptance recipe later requires 2.5, so you'd set it there). The detector runs BGR YOLO, and the outputs are per-frame float crops, not integers.

It's shot-aware like the main plan: scene_cut_threshold (0.28), max_track_jump (0.18) and max_gap_frames (4) control when a face is a continuing track versus a new detection, and smoothing resets at each cut. canvas_mode defaults to auto_capped_768, require_h3_grid keeps crops on H3's grid so the VAE accepts them, and analysis_chunk_frames (8) bounds how many frames the detector processes at once on CPU.

The honest footnote in the description is the bit people should remember: "The best source crop is only a convenient reference, not identity proof." The plan picks the cleanest-looking crop of each face so you have something to eyeball, but that's not an embedding, not a biometric, and not a promise the refined face will match it.

Inputs are the same shape as the main plan - frames, fps, detector_mode/detector_model/detector_device, confidence, manual ROI fallback, and the tracking/smoothing fields above.

Outputs: face_plan (the H3_T8_FACE_REFINE_PARITY_PLAN type the rest of the parity chain consumes - note it's a different plan type from the regular refine chain), crops, reference_crop, preview, report_json, and canvas_width/canvas_height/frame_count/reference_frame_index for layout.

Use this node when you want your face-refine results to be comparable to the reference - for the parity chain that feeds Parity Latent → Per-Frame Denoise → Parity Stitch → Quality Gate → the MANUAL512 acceptance baseline. If you just want a face fixed and don't care about the archaeology, use the regular Plan node instead. This is the "measure twice" branch.

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 downloads - YuNet runs from your local OpenCV, and nothing here touches the model weights.