T8 Storyboard Pack（分镜创作包）

Storyboarding is where a video project actually gets made - and most people do it in their head, because writing a shot list by hand is tedious and asking an LLM to do it costs you a call per table. T8 Storyboard Pack（分镜创作包） collapses the whole thing into a single LLM request that returns a complete creative pack: a global prompt, a per-shot list, keyframe image prompts, and a transition/sound table. That's four deliverables for one paid call.

The efficiency trick is in the request contract. The model is asked to return only one JSON object - global_prompt plus a shots array - where each shot carries all its fields in one place: timing, purpose, composition, subject action, camera, continuity, media bindings, dialogue/text, sound, keyframe prompt, transition in/out. The keyframe prompts and the transition/sound table are then derived locally from that per-shot contract. The pack deliberately doesn't make the model repeat itself into separate tables, because that's how you end up paying twice for the same content - the README's acceptance notes call this exactly: the model returns one shot contract, and the tables are assembled on your machine.

The inputs

• concept (required) - your creative direction, ideally the candidate you picked with the Selector.
• creative_brief - optional, from the Director, for locked identity/world/visual rules.
• reference_role_map - optional, from the Reference Role Mapper, so @素材N bindings only appear where a real asset exists.
• duration_seconds, shot_count (AUTO or fixed 1–20), model_target, rewrite_mode, output_language, seed, plus the usual api_key / provider_config.

What comes out

• global_prompt - the master prompt for the whole piece.
• shot_list_json - every shot with its full executable chain.
• keyframe_prompts_json - still-image prompts for keyframes, derived locally. The contract explicitly bans impossible temporal actions in keyframes ("a person running" is fine; "a person running then stopping" is a video prompt, not a keyframe).
• transition_sound_json - cut types and sound design per boundary, derived locally.

Feed shot_list_json into whatever render loop or segment pipeline you use, and keyframes into your image model. There's no built-in renderer - this is a planning node, and it says so.

Installing

In the comfyui-minimax-h3-prompt-enhancer-T8 pack. Manager: search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

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

Restart ComfyUI and Ctrl+F5 the browser. Cloud mode adds no Python dependencies. The single call uses the pack's shared provider chain - Seedance NZ by default, or a T8 LLM Provider Config for AI Workshop / OpenAI-compatible / local Qwen.

Honest notes

The pack doesn't generate, assemble, or preview footage - you get text, and the rendering is your job. And like the rest of the suite, it's a third-party layer by T8 (T8mars), not an official MiniMax product, and it never calls the H3 or Seedance generation APIs. The one thing that surprises people: because the model returns a single shot contract, if it returns slightly non-JSON text the node keeps the raw response rather than dropping it - so garbage output is visible in shot_list_json's unparsed_response field instead of silently vanishing. Check it once, then trust the local derivation.