MiniMax H3 Unified Cast / 统一角色表 (Advanced)

Character continuity across multiple shots is the hardest problem in AI video, and H3's answer is basically "remind it, in every shot, who everyone is." This node is the reminder. MiniMaxH3UnifiedCastT8Advanced builds one validated character-continuity contract - a structured list of your characters and their visual identity - that the Studio Timeline compiles into every shot prompt. Define the cast once, and every shot in the timeline carries the same identity text instead of you hand-typing "consistent facial identity and hairstyle" into fourteen prompts.

Crucially, it stores text metadata only. The description is explicit: it does not load a face, voice, or vision model. This isn't an identity-LoRA trainer or a reference-image packer - it's a structured prompt object. Which is exactly why it's cheap and safe: no VRAM, no model downloads, no way for it to hallucinate a face. It's the "write the character sheet" step, done properly.

How it works

One main input: cast_json, a JSON array of character entries. The default shows the shape - {"id":"hero","name":"Hero","visual_identity":"consistent facial identity and hairstyle"} - and you extend that with one entry per character. Each id is the handle the timeline uses to bind a shot to a character; name and visual_identity are what get compiled into prompts. The strict_identity toggle (default on) is the validation gate: in strict mode, the node rejects ambiguous, missing or duplicate identity bindings rather than silently guessing. Leave it on - the whole point of a cast sheet is that a shot referencing "hero" is unambiguous.

Outputs:

cast (the H3_T8_UNIFIED_CAST object you feed into the Studio Timeline's optional cast input) and report_json summarizing what was validated. Wire cast into the Timeline and every shot you tag with a character id inherits the identity text.

Installing:

part of MiniMax H3 Audio T8. Manager → search "MiniMax H3 Audio T8", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart after. No mandatory pip deps.

The honest caveat

A text contract makes identity consistent across prompts, but H3 still renders that identity fresh each shot - this doesn't lock a specific face the way a reference-image or identity-model pipeline can. If you need exact facial continuity, you'll want to pair the cast sheet with reference images or a face-refine pass. And as always with this pack, the license reality: H3's local weights are restricted in the US/EU/UK/Korea, which gates every workflow you'd build here.