🧙‍♂️H3专用提示词框@炮老师的小课堂

MiniMax H3 is the video model whose prompt format is built around explicit material references - you don't say "use that clip," you literally write <Picture 1> or <Video 2> inside the prompt, and the model grabs the matching asset. DapaoH3PromptBoxNode is the pack's helper for exactly that: a two-input, two-output box that makes sure your prompt and your material list are in the shape H3 expects, and hands the reference map to downstream nodes as a typed object.

What it does

Two required inputs:

• 📝 H3提示词 - the prompt you're writing, where you embed the <Picture/Video/Audio N> markers yourself or with the help of the prompt-generating nodes.
• 🧩 H3素材清单 - the material manifest: which asset is which number.

Two outputs:

• 📝 H3提示词 - the (possibly normalized) prompt string, ready to feed a node that calls H3.
• 🧩 H3素材标记 - a DAPAO_H3_REFERENCES typed object carrying the parsed material markers, for the sibling DapaoH3VideoPromptNode and any H3-generation node in the pack to consume.

That's the whole node. It's plumbing - deliberately so. If you've ever hand-typed H3 tags and had the model silently ignore a <Audio 3> because the tag didn't match the manifest, you understand why a box that keeps the two in lockstep is worth having.

Where it sits

It's part of the H3 workflow in this pack: DapaoH3VideoPromptNode generates a full H3 prompt from a plain-language request (using the Context-IR spec, T2VA/I2VA modes, and up to 9 images + 3 videos + 3 audio references), and this box is the manual/adjacent counterpart - when you want to write the H3 prompt by hand but still keep the material tags structured and consistent. The pack's Skill list even includes h3-prompt-writing for the chat workbench, so there are three roads to the same destination.

The real-world catch

H3's official prompt format rewards discipline: a missing or misnumbered marker means the model just doesn't use that asset, and nothing tells you. The box catches manifest-vs-prompt drift by keeping both in one place, but it won't fix a <Picture 2> that points at an asset you meant to be <Picture 3>. Read the markers as literal contracts - that's the lesson H3 teaches everyone once.

Install

It ships in the ComfyUI-dapaoAPI pack, so the install is the pack story: ComfyUI Manager → search "dapaoAPI" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git
cd ComfyUI-dapaoAPI
pip install -r requirements.txt

No API key lives in this node - it does no network I/O, which makes it the friendliest thing in this pack. The key shows up downstream where the actual H3 call happens. Chinese-only labels, as with the whole pack, but with two inputs and two outputs you can't really get lost.

If it's this thin, why does it exist?

Because the pack's H3 prompt generator needs a typed reference object to hand off, and a raw string would let you smuggle in tags that don't line up with the material list. It's the difference between a self-describing workflow and a stringly-typed one. Annoying? A little. But the first time a video generation silently drops your reference audio, you'll appreciate the structure.