MiniMax H3 Face Refine MANUAL512 REL Baseline / 人工验收512相对模式基线 (Advanced)

If the Quality Gate is the automatic regression tripwire, this node is the formal one: the manual-acceptance baseline for the author's MANUAL512 REL (relative-mode) face-refine recipe. The name is long because the contract is long. It's a fail-closed pass-through - it doesn't improve your candidate, it verifies that the chain you built matches a specific, documented reference recipe, and it refuses to bless the run if the reports disagree.

Here's the contract, verbatim from the description, because it's the whole point of the node. Your chain must have: a 512×512 crop canvas, crop_factor 2.5, 21/51 smoothing, BGR YOLO detection, crop-derived relative video denoise mapping small faces to 0.8 and large to 0.35, locked zero-mask audio, and the author's centred-mask/source-space 24/24 stitch. And it must agree with the reports - latent_report_json, denoise_report_json and stitch_report_json are all inputs, not outputs. You feed this node the three reports your chain produced, plus the face_plan and the candidate frames, and it cross-checks everything: did the canvas match, did the smoothing match, did the denoise curve match, did the reports agree, and is any explicit one-frame H3 alignment tail accounted for?

That's the unusual design here: this isn't a node that computes a baseline. It's a node that attests a baseline. The profile input is locked to manual512_relative_author_parity_v2 - there's one recipe, and the node will not bless a run that deviates from it. minimum_crop_face_height_px (200) is the only real dial: the minimum face height in pixels the crops must satisfy to be taken seriously.

Inputs: candidate_frames, face_plan (parity type), and the three report JSONs. Outputs: candidate_frames (passed through) and baseline_report_json - the attestation you can keep next to the renders as evidence of what was verified.

When would a normal user reach for this? Honestly, when you're being asked to formally verify that a face-refine render followed the audited recipe - a quality pass on a production job, a comparison study, or a "did I build this right" sanity check. The description ends with the standard pack disclaimer: it does not use proxy scores to replace the candidate and is not a universal quality guarantee. The gate checks the recipe, and the recipe being right is not the same as the face being good. You still look at the video.

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. All the actual refinement happened upstream - this node only reads reports and validates.