小珠光 MiniMax H3 提示词

MiniMax H3 is one of the closed video models where the prompt is half the craft - its official prompt-writing guide expects structured, mode-aware prose, and slapping a random sentence in gives you a random video back. This node (小珠光 MiniMax H3 提示词) is a local prompt formatter for H3: you type a rough idea, it hands the text to a local Qwen LLM, and returns a structured prompt tailored to the generation mode and style preset you picked. The whole point is that the rewrite happens on your own GPU - no API, no cloud, no per-call cost.

Mechanically it's the classic local-LLM-in-the-graph pattern the KB's LLM essay covers. A single LLM call (the author deliberately avoided the double-call translate-then-structure architecture of the original BSAI node) asks Qwen to output a structured H3 prompt in Chinese and English. The interesting dependency wrinkle: it doesn't load the model itself. The qwen_model input takes a BSAI_QWEN_MODEL - the same model type the original BSAI MiniMAX H3 Prompt plugin's loader produces, and which this pack's own XiaozhuguangQwenModelLoader also emits. So you pair this node with either the BSAI loader or the pack's GGUF loader, feed it the loaded model, and it does the prompting.

The inputs that actually matter:

• qwen_model - the loaded model handle from the pack's Qwen Model Loader or BSAI's H3 loader.
• user_prompt - your rough idea. This is what gets rewritten; write it like you're telling a friend what the video should be.
• generation_mode - text-to-video (T2VA), image-to-video (I2VA), first+last frame (FL2VA), last frame (L2VA), or full reference (Ref2VA). Pick the mode you're actually generating in - the structure this emits depends on it.
• style_preset - H3's official style presets (minimalist product ad, 3D animated short, papercraft explainer, music-video aesthetics, and so on). Pass-through to H3's own styling vocabulary.
• seed - reproducibility of the rewrite.
• unload_after - unloads the LLM after the call to free VRAM for the video model. Leave it on unless you're chaining many prompts; if you're using BSAI's loader instead, the README notes to turn it off so the loader's own management can do its thing.
• image_1…image_9 - optional reference images, up to nine. These get passed to the VLM for the image/reference modes, where the model describes what it sees into the prompt.

One output: prompt_output (STRING), which feeds your H3 text encoder or API call.

The honest caveats from the LLM-in-graph essay apply here. A local 8B Qwen is not writing better prose than a frontier API, and it can drift from what you asked - it's a convenience that kills the blank-page problem, not a guarantee. And it needs a real LLM to work: no model loaded into the qwen_model port, no output, just an error. Budget VRAM for two models at once if you leave the Qwen resident while H3 runs - that's what unload_after is for.

Install the pack (ComfyUI Manager → ComfyUI-xiaozhuguang, or git clone into custom_nodes, restart), then get a GGUF Qwen into ComfyUI/models/LLM/ - see the pack's Qwen Model Loader node for that. This node specifically drags in llama-cpp-python and the transformers stack, which are in the pack's declared dependencies; the registry install pulls them automatically, and a git install means running the pack's requirements. It's a niche node for a niche job, but if you run H3 workflows, structured prompts are worth exactly this much trouble.