MiniMax H3提示词扩写指令 V2·完整模式（Tlant）

Here's the whole V2 design in one sentence: stop letting the node guess. The V1 family (Simple Prompt, the six config nodes, the Assembler) fills anything you leave alone with seed-stable random values. MiniMax H3提示词扩写指令 V2·完整模式（Tlant） throws all of that out. It's a single node holding every category - identity, face, editing, camera, scene, audio - where every single option is explicit, nothing is random, and unset means unset, not "roll the dice."

The mechanism is a clean three-valued system applied to every optional element:

• 不指定 - the default. The element is simply not written into the instruction. The LLM is explicitly told an absent option is genuinely unspecified and must not invent a value to fill a template.
• 无 - explicit prohibition. The element is banned; for unavoidable visual properties (like a face existing), it means "don't deliberately change the source."
• 自行推断 - the LLM is asked to infer a safe, source-grounded value when relevant.

V2 has no local seed at all - no random roll happens inside the node, so reproducibility comes from repeating your settings, not from a lucky draw. What you don't select stays out of the prompt. It's the "explicitness over convenience" philosophy, and for a model as prompt-sensitive as H3 it's a defensible trade.

Common inputs for both V2 nodes: 视频时长 (4–15s), 主题预设 + 自定义主题 (non-empty custom theme overrides the preset completely), and 提示词长度最小值 - default 200, the minimum English words the final H3 prompt must contain. The node enforces it via the instruction, not by padding: the LLM is told to reach the minimum through useful detail and never by repetition.

The full list is big - fifty-odd dropdowns covering identity lock, facial/head motion, face direction, gaze, expression, cuts, shot count and scale, transitions, continuity, ending, every official camera move, amplitude/speed/energy, handheld, axis, depth of field, motion blur, background, scene expansion, lighting, atmosphere, hair, clothing, VFX, environment events, and the whole audio stack (mode, music, tempo, energy, beat sync, SFX, ambience, dialogue, language, voice style, lip sync).

Two audio policies are worth calling out because they're the design's best feature. Music is opt-in: the LLM may only generate background music if you explicitly choose music or a full mix, enable music, pick concrete music parameters, or set them to infer - otherwise the instruction forces non_diegetic_music: N/A. Dialogue is off by default: no invented speech, singing, narration, or on-screen text unless you ask. Both stop the "why did it add a soundtrack/voice I never wanted" complaints that plague generative video.

Outputs: LLM扩写指令 (STRING) and 图像 (IMAGE passthrough) - feed both into a vision LLM (Tlant's Llama Server Chat is the natural bridge) to get the clean final H3 prompt.

Install

ComfyUI Manager → search ComfyUI-Tlant-Toolkit → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Tlant/ComfyUI-Tlant-Toolkit

Restart ComfyUI. No dependencies, no model downloads.

Gotchas

An image or a non-empty 图像描述 is required, same as V1 - no image, no description, and the node refuses. And don't fight the philosophy: if you leave 身份锁定 on 不指定, the prompt won't mention identity strength at all, and H3's baseline continuity still applies but you've given up your say. That's not a bug, it's the point. Also remember V2 coexists with V1 in the same pack - they're separate nodes, and V2 won't change how your old V1 workflows behave.