MiniMax H3提示词扩写指令·简单模式（Tlant）

MiniMax H3 is the open-weight omni-modal video model everyone's been talking about - native stereo audio, 4–15 second clips, unified text/image/video/audio context. The catch for anyone who actually wants to use it well: it's prompt-hungry, and the prompt format (the I2VA-style first-line anchor plus integrated_multimodal_description, overall_soundscape, and non_diegetic_music fields) is not something you want to hand-type for every clip. MiniMax H3提示词扩写指令·简单模式（Tlant） exists so you don't have to. It's the one-node entry point into the Tlant MiniMax H3 family.

Here's the honest mental model: this node does not run H3, and it doesn't call MiniMax's API. It's a prompt-engineering node. You give it a handful of Chinese options plus your starting image (or a text description of it), and it emits a long, stateless English instruction - a prompt for a remote LLM to write your H3 prompt. Wire the LLM扩写指令 output into a local LLM node (Tlant's own Llama Server Chat is the natural pairing), and that LLM returns the clean, final MiniMax H3 I2VA prompt for your video workflow.

It's called "simple mode" for a reason: twelve visible options, everything else handled for you. The ones you'll actually think about:

• 视频时长 - 4 to 15 seconds, MiniMax H3's official range.
• 画面比例 - defaults to 跟随原图 (follow the source image), which is usually the right call for I2V.
• 主题预设 / 自定义主题 - pick a genre like dance, MV, or cinematic fashion; a non-empty custom theme completely overrides the preset.
• 创意幅度 - how far the video may drift from the static image, conservative to wild.
• 人脸保护 - defaults to 严格 (strict). For realistic people this is the setting you most want left alone; face drift is the classic H3 failure.
• 随机种子 - the reproducibility key. Set ComfyUI's seed control to fixed and the same seed + same inputs give you the same option combination every time.

Whatever you don't set gets filled by seed-stable random sampling - each hidden field rolls independently from seed + field name, so changing one thing doesn't scramble the others. It's a genuinely nice design: you control what you care about and let the node handle the long tail.

Two outputs: LLM扩写指令 (STRING, the instruction for the remote LLM) and 图像 (IMAGE passthrough). Connect the image output alongside the instruction into a vision LLM node, because when an image is attached the instruction tells the LLM the image is authoritative over any text description.

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 - this pack is pure Python.

The gotchas

• The UI is Chinese. Every label, dropdown, and tooltip is in Chinese. The options are Chinese, the output is always English. If you can't read Chinese, the hover tooltips are still your best map.
• You must provide an image or a description. No image connected and an empty 图像描述? The node refuses with a clear Chinese error. Connect one or the other.
• It's upstream, not the generator. If you queue it and see no video, that's expected - you still need an LLM to turn the instruction into an H3 prompt, and an H3 workflow to run it. It's the blank-page remover, not the renderer.
• License reality check. H3's community license excludes the US, EU, UK, and South Korea from running the local weights; in those regions the hosted Hailuo API is the licensed path. This node doesn't change that - it just writes prompts.