🦊H3视频提示词生成@炮老师的小课堂

MiniMax H3 is a closed video model that is picky about how you talk to it. Its official Context-IR spec wants prompts structured around material references - first frame, reference videos, audio - written in a specific form, and a plain English paragraph gets you a video that ignores half your references. DapaoH3VideoPromptNode is the dapaoAI pack's translator: it calls an LLM to rewrite your plain-language video idea into a proper H3 prompt, and it does the material analysis for you, handling up to 9 reference images, 3 videos, and 3 audio files.

What you feed it

• 🔑 API密钥 + 🤖 LLM模型 - dapaoAI key and the usual 11-model menu (default gemini-3.7-flash).
• 🎛️ H3生成模式 - the mode set, default auto-detect: T2VA (text-to-video-with-audio), I2VA (image-first-frame-to-video-with-audio), FL2VA (first+last frame), L2VA (last-frame continuation), and so on.
• 🎨 创作类型 - 31 creative types (general H3, character, product, cinematic…). Picking one shapes the prompt's tone and shot vocabulary.
• 🌐 输出中文提示词 - false by default, so English prompts come out unless you flip it. Note the defaults elsewhere: ⏱️ 目标时长(秒) 5, 📐 视频比例 16:9, 🔊 原生音频 true, 🌡️ 温度 0.4.
• 📝 原始视频需求 - your idea in plain words.

Optional inputs are where the reference handling lives: 🎬 首帧图 and 🏁 尾帧图 (for FL2VA), 🖼️ 参考图1–9, 🎞️ 参考视频1–3, 🎵 参考音频1–3, plus a 🧩 H3素材标记 typed input from DapaoH3PromptBoxNode, and an 🎧 参考音频原声直传LLM toggle for whether audio goes to the LLM as-is. The 🎞️ 每个视频采样帧数 (default 5) controls how many frames get pulled per reference video for analysis - the node doesn't upload whole videos to the LLM, it samples representative frames, which is the sane way to keep a vision model's context from exploding.

What comes out

The output to wire downstream is 🎬 H3最终提示词 - the structured prompt for your H3 video generation node. The rest are diagnostics: 🎛️ 识别模式 (which mode it picked), 📑 素材与制作分析 (its read of the materials), 📄 LLM完整响应, and ℹ️ 处理信息.

How it works

The node builds a Context-IR-shaped instruction set, sends your text plus the sampled reference frames to the dapaoAI LLM, and asks it to emit the H3 prompt complete with the <Picture/Video/Audio N> markers that H3's own node will honor. It's the "LLM as translator to a model's official prompt guide" pattern that llm-in-comfyui.md describes - the community's top advice for models like this is literally "read the official prompt guide, then have an LLM translate what you want into it."

Gotchas

• This node makes prompts, not video. Pair its output with an actual H3 generation node - the pack's RH and dapaoAI video nodes, or MiniMax's own.
• Marker discipline is on the LLM. A model that ignores a <Video 2> tag will silently drop the reference. Scan the 素材与制作分析 output before paying for a generation.
• Reference-heavy requests cost more tokens - 9 images + sampled video frames + audio can push a call well past a bare text prompt.
• No auto-retry of paid POSTs, per pack policy; a truncated response means re-running and paying again.
• Chinese-only UI.

Install: ComfyUI Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI.git → pip install -r requirements.txt → restart. dapaoAI key from api.dapaoai.com (register → wallet → redeem code → apply key in the default group).