MiniMax H3 Face Cast Merge / 2-3人角色表 (Advanced)

If Face Character Profile builds the individual character cards, this is the node that stacks them into a cast. It chains two or three unique character profiles into one in-memory H3_T8_MULTIFACE_CAST, which is the object the track-assignment node will then use to label faces in your video. It's a small, boring node with two rules that matter: it refuses duplicate IDs, and it never writes biometric embeddings to disk.

The wiring is a chain, and it reads like a linked list. The profile input is required and takes a character profile; the previous_cast input is optional and takes an existing cast. Connect the first character into profile, leave previous_cast empty, and you get a cast of one. Now add the next node: feed the previous cast into previous_cast and the new character into profile - you've got a cast of two. Chain a third the same way for three. The node refuses duplicate character IDs, so you can't accidentally add the same character twice and get a confused cast with two "Alice" slots. If you need more than three characters, this particular merge node tops out at three per chain by design - you'd run separate casts.

Outputs: face_cast (the merged H3_T8_MULTIFACE_CAST), reference_contact_sheet (an IMAGE grid of the cast's reference crops - genuinely handy for eyeballing the whole cast at a glance), and report_json.

Two practical notes. First, the contact sheet output is the thing you'll actually look at most: it's a single image showing all characters side by side, which makes it easy to spot a weak reference (a profile where the reference crop is too small, too dark, or has the wrong person). Second, remember what a cast is for in this pipeline: it's the roster that Face Track Assign consults when it decides which SAM track is which person. A sloppy cast means sloppy assignments, so the reference quality you put into the profiles upstream is what determines whether the two-character dialogue scene keeps both faces straight.

This is squarely in the "experimental multi-person" tier of the pack, and the honest framing is the same as the profile node: the embeddings are matching aids, not identity proof, and the whole chain fails closed rather than guessing. The cast object lives in memory for the duration of the workflow - nothing is persisted to disk.

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. This node just combines objects you already built - the heavy lifting is upstream in the profile and downstream in the tracking.