T8 Performance Director Config（表演导演）

Somewhere in the middle of hand-writing H3 and Seedance 2.0 prompts you notice the same thing everyone notices: describing what happens is easy, describing how a person performs it is not. A character "reacts" - but to what, in what order, with which visible cues? T8 Performance Director Config is a single-combo node that bolts a mini acting-director onto your prompt-enhancer request, and the best part is the price: zero additional LLM calls. It's a config object, not a model. No API key, no request, no cost per run.

Honest framing first, because the author says it plainly in the tooltip and the source: this is community-research-inspired, not an official MiniMax Skill, and its effects depend on model, media, and task. It's soft guidance threaded into the enhancer's system rules, not a proven cross-model law. The inspiration is credited to the phileiny/h3-storyboard-skill project at a pinned commit, and the emitted config carries an evidence_level of "narrow observations pending independent paired validation." Take it as a good default, not a guarantee.

What the modes do

• AUTO（按人物 / 表演意图） - the default. The structure only applies when your text or connected media actually contains a person, character, emotion, dialogue, reaction, or acting intent. Pure landscape, product, text-motion, and music requests skip it entirely.
• 强化（明确表演结构） - force the acting structure explicitly: organize character performance as trigger → reception → one primary response → settled end state, with each beat capped at three observable cue channels per character.
• 关闭（保持原编译） - restore the original H3/Seedance compile path with none of this added.

Under both active modes it translates abstract emotion into visible evidence - gaze, eyelids, breathing, shoulders/neck, hands, posture - establishes camera side and a concrete gaze target before prescribing head direction, preserves your dialogue verbatim and never invents speech, and forbids scheduling incompatible mouth actions inside a speaking span. It won't hide a change the user demanded stay continuously visible, and your explicit locks and hard constraints outrank everything it says.

Input, output, wiring

One combo input, mode, and one output, performance_director_config - a custom socket type (purple, per the README), not a STRING. You connect it to the H3 or Seedance 2.0 enhancer's optional performance_director_config input, or to T8 Storyboard Pack if you're using the creative suite. Unplug it and the enhancer silently falls back to AUTO-conditioning.

H3 and Seedance get separate rule sets - Seedance never receives H3's field tags, speaker IDs, millisecond timecodes, or frame-grid notation, so wiring one director doesn't cross-pollinate the two models' protocols.

Install and gotchas

Same pack as everything else in this suite: 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 after install. What bites people: the output is a custom type, so you can't substitute a text node or feed it through a STRING path - add the T8 Performance Director Config node and wire its socket directly. And don't expect a miracle: this shapes how the LLM writes the acting, it doesn't fix a model that won't hold a performance. Pair it with the pack's T8 Character Performance Bible for character intent and you've got a proper two-layer setup - the Bible supplies what the character wants, the Director imposes how the beats read. The catch with AUTO is it's conditional by design, so if your results aren't changing, check whether the enhancer actually sees acting intent before you blame the node.