🦊H3本地视频提示词生成@炮老师的小课堂

MiniMax H3's video workflow has a specific shape: you feed it text, images, video frames, and audio, and it needs everything organized into its official Context-IR prompt format - with the right material markers, field order, and timing. DapaoH3LocalPromptNode ("H3本地视频提示词生成") does that organization with a local GGUF vision model instead of an API. The pack frames it bluntly: independent local inference, no online H3 node, no online LLM API, no API key.

What makes this worth your attention is the scope of what it juggles. It auto-detects the H3 task mode - T2VA (text-to-video), I2VA (first-frame), FL2VA (first+last frame), L2VA (last frame), and Ref2VA (all-round reference) - and it accepts up to 9 source images, 3 video frame batches, and 3 audio tracks (12 mixed materials max). Video arrives as time-sorted IMAGE frame batches, so there's no video-decode dependency to install; audio is analyzed locally for spectrum, energy, and silence by default, with a switch to pass raw audio through when your model supports it.

The inputs that matter

The model side is the pack's standard cluster: 🤖本地模型文件, 🔌本地对话处理器, 🖼️本地mmproj文件, context length, the 💾本地显存限制(GB) budget (same tooltip math: 8GB card → 6, 12GB → 10, leave ~2GB headroom), image min/max tokens, thinking mode, and the Qwen3.8 reasoning-strength dropdown.

Then the H3-specific controls:

• 🎛️ H3生成模式 - auto-detect or force one of T2VA/I2VA/FL2VA/L2VA/Ref2VA.
• 🎨 创作类型 - 31 presets from 通用H3 through 极简产品广告, 3D动画短片, 真人电影叙事, 美食料理广告·ASMR, 时尚美妆大片 and more. This is where the prompt's genre voice comes from.
• 🎬 首帧图 / 🏁 尾帧图 - the anchor images for I2VA/L2VA/FL2VA; the tooltips are explicit that these act as precise anchors in the prompt.
• 🎞️ 每个视频采样帧数 (2–8, default 5), ⏱️ 目标时长(秒) (4–15), 📐 视频比例 (21:9 … 9:16), and 🔊 原生音频.
• 📝 原始视频需求 - your actual idea. This is the creative input; everything else is scaffolding around it.
• 🧩 H3素材标记 - optional DAPAO_H3_REFERENCES from the DapaoLocalH3PromptBoxNode, which locks numbering to the official H3 node's actual material order.
• 🖼️ 参考图1 through 🖼️ 参考图9, 🎞️ 参考视频1–🎞️ 参考视频3, 🎵 参考音频1–🎵 参考音频3 - the Ref2VA material ports (each video/audio 2–15s, total ≤15s).

The outputs

🎬 H3最终提示词 (STRING) is the one that matters - wire it to the official MiniMax H3 node's text input. The rest are diagnostics and plumbing: 🎛️ 识别模式, 📑 素材与制作分析, 📄 LLM完整响应, and ℹ️ 处理信息. The 🌐 输出中文提示词 toggle flips the prose to simplified Chinese while keeping the official H3 field format and markers intact.

Where people get burned

Two things. First, the audio reality check: raw-audio passthrough only works if your chosen GGUF actually supports input_audio - the default local spectrum/energy analysis is the safe path, and the tooltip says so. Second, model pairing: the pack recommends multi-image-capable vision models like Qwen3.5 / Qwen3-VL with a matched mmproj; a weak or mismatched projection means muddled material understanding, and you'll feel it in wrong references.

Installing it

Standard pack:

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

then requirements + restart. The pack-wide gotcha: llama-cpp-python is pinned to the JamePeng fork at 0.3.47+ (stock PyPI 0.3.35 fails on Qwen3.8 GGUFs with missing tensor 'blk.64.ssm_conv1d.weight'); NVIDIA users install a matching GPU wheel from the JamePeng releases and restart. GGUF + mmproj go in ComfyUI/models/LLM/.