MiniMax H3 Face Character Profile / 多人角色参考 (Advanced)

This is the "who is this person" node for multi-person face refine on H3. You give it one or more single-person reference images, it detects the face(s) with OpenCV's YuNet + SFace on CPU, and it builds an in-memory character profile - an embedding plus the reference crops - that the rest of the multi-person pipeline uses to recognize that person in the video and decide which track is them. Name your characters, build profiles, merge them into a cast, assign tracks.

The input list is short on purpose:

• character_id - your label, default Character_A. This is what the track-assignment reports will call this person, so make it meaningful (Alice, Lead, Suspect #2).
• reference_images - one or more single-person reference shots. More angles = a more robust matching embedding, but each image must be a clean single-person reference.
• reference_face_policy - dominant_face_auto by default (use the biggest/clearest face), require_single_face (refuse if more than one face is detected - the strict option), or largest_face_exp.

Outputs: character_profile (the H3_T8_MULTIFACE_CHARACTER_PROFILE object), reference_preview (an image you can eyeball), and report_json.

The thing to keep front-of-mind, because the tooltip hammers it: the embedding is a matching aid, never persistent identity proof. It's used to suggest "this track looks like Alice," not to certify identity. Nothing here writes biometric embeddings to disk, and the whole pipeline is built to fail closed - if the similarity is below threshold, it says "not confident" rather than guessing.

Where this fits in the wider workflow: H3 doesn't give you multi-person identity control natively, so this pack builds a light character-consistency layer on top. The chain is: build a profile per character (this node) → merge up to 2–3 into a cast (Face Cast Merge) → track faces in the video with SAM3.1 → bind each track to a character (Face Track Assign) → run per-character refine. If you're doing a two-person dialogue scene where both faces need to stay consistent, this is the stack you'd want.

Honest expectations: this is the "experimental multi-person" tier of the pack, and the author's own notes say the multi-person route reuses the existing local YuNet and native SAM3.1 tracks - it's CPU face detection plus embeddings, not a trained identity model. It's a matching layer, and its value depends on your reference quality. Give it good, well-lit, single-person references and it'll do solid work; feed it group photos with dominant_face_auto and you're gambling.

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 during execution - YuNet/SFace come from your local OpenCV install.