🧙‍♂️H3专用提示词框@炮老师的小课堂

MiniMax H3 is a video model with a particular prompt language: its prompts reference materials by official markers - <Picture 1>, <Video 2>, <Audio 3> - and if you hand it a prompt full of loose prose instead of those markers, it can't map your words to the actual files you attached. DapaoLocalH3PromptBoxNode ("H3专用提示词框") is a plain text box that knows that language. Type your prompt, @-select the materials already connected to the downstream official H3 node, and it produces an H3 prompt with the correct Picture/Video/Audio N markers baked in, plus a machine-readable references object the other H3 nodes can lock onto.

Unlike the rest of the pack's H3 story, this node runs no LLM at all. It's pure prompt assembly - no model dropdown, no VRAM, no inference. That's what makes it useful in two places: as a manual box for people who want to write H3 prompts by hand but hate counting markers, and as the @素材 authoring surface that feeds DapaoH3LocalPromptNode (the "H3本地视频提示词生成" node). The pack's wiring shows the H3素材标记 output (type DAPAO_H3_REFERENCES) plugging into the local prompt node's 🧩 H3素材标记 input, so the generated prompt and the reference numbering stay consistent.

Inputs and outputs

• 📝 H3提示词 - your prompt. It supports dynamic prompts and the @ material picker.
• 🧩 H3素材清单 - auto-maintained by the node UI from what's actually connected to the downstream official MiniMax H3 node. You normally never edit this.
• Outputs: 📝 H3提示词 (STRING, wire into the H3 text-encode or the local prompt node) and 🧩 H3素材标记 (DAPAO_H3_REFERENCES, the marker/order contract for other nodes).

The marker contract matters

The tooltip on the local prompt node's H3素材标记 input explains the payoff: connecting it uses the official H3 node's actual material order to lock the numbering. So the manual box and the LLM generator agree on what <Picture 1> means, which is exactly where hand-written H3 workflows fall apart. If you've ever generated an H3 prompt and watched the video reference the wrong image, this is the fix: one consistent marker source.

Installing it

Ships in the standard pack:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-llama_Dapao

then install requirements and restart. Because it's a pure text-box node, none of the llama-cpp-python GPU-wheel pain applies to this node specifically - but if you use it with DapaoH3LocalPromptNode, that sibling needs the full model setup: GGUF + matching mmproj in ComfyUI/models/LLM/, llama-cpp-python from the JamePeng fork (0.3.47+ for Qwen3.8), and a restart. Worth saying once: a pack that mixes a zero-dependency text box with heavy inference nodes is exactly why you read the node's inputs before installing - here the box is the harmless part.