MiniMax H3提示词扩写指令·高级汇总（Tlant）

The six "MiniMax H3高级配置" nodes in this pack are just parts; MiniMax H3提示词扩写指令·高级汇总（Tlant） is where they bolt together. This is the hub of the advanced mode: it takes whichever category configs you've connected - basic generation, identity, action, camera, scene, audio - and turns the whole bundle into a single stateless English instruction for a remote LLM to write your MiniMax H3 prompt.

Same honest framing as the rest of the family: it doesn't run H3 and it doesn't call MiniMax. It writes the instruction. Connect its LLM扩写指令 STRING output to a local LLM node (Tlant's Llama Server Chat works perfectly) and that LLM returns the clean, final H3 I2VA prompt - the first-line anchor plus integrated_multimodal_description, overall_soundscape, and non_diegetic_music - for your video workflow.

The genius of the design is what happens with the slots you don't fill. Every unconnected category input gets filled with reproducible random values drawn from the 随机种子. Each field samples independently from seed + field name, so you can precisely control, say, identity and camera while letting action and scene roll the dice - and a fixed seed reproduces the exact combination every time. Want full control? Connect all six. Want to nudge one thing? Connect just that category and leave the rest random. It's the most flexible layout in the pack and it's the reason the "advanced" workflow stays usable.

Inputs in full:

• 随机种子 (required) - the only required input. Set ComfyUI's seed control to fixed for reproducibility.
• 图像 (IMAGE) and 图像描述 (STRING) - you need one or the other. With an image connected, the instruction declares the image authoritative and passes it through untouched.
• The six config inputs - 基础生成配置, 人物一致性配置, 人物动作配置, 镜头剪辑配置, 场景效果配置, 音频配置. Each is a custom typed port, so only the matching config node will plug in.

Outputs: LLM扩写指令 (STRING) and 图像 (IMAGE passthrough). Wire the image output alongside the instruction into a vision LLM so the model can actually see your first frame.

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.

Troubleshooting

• "Please connect an image or a non-empty image description." The classic first-run error. If you've got an image in the graph, wire it into 图像; otherwise give 图像描述 a non-empty string from an upstream captioner.
• A category doesn't seem to do anything. Check you plugged it into the matching port - the custom types exist specifically to stop you connecting an audio config to the scene slot, and they make mismatches fail loudly rather than silently.
• Output not showing as random when expected. Remember only genuinely unconnected categories get randomized; a connected category with everything set to 随机 still rolls from the seed. And ComfyUI's seed control must be fixed for results to repeat.
• No video appears. Expected - this node is upstream of generation. It makes the prompt-writing instruction; the LLM and the H3 workflow do the actual work.