😶‍🌫️Seedance2全能导演@炮老师的小课堂

ByteDance's Seedance 2.0 video model expects a "director's cut" level of prompt: camera, character mapping, continuity across shots. DapaoLocalSeedance20DirectorNode ("Seedance2全能导演") is the local LLM that plays director - it compiles the prompt, maps materials to characters, carries serial story state between segments, and even reviews finished segments and repairs failed generations. Like the rest of the local tool family, it's a dapaoAPI node repointed at your GGUF: same inputs, same structured validation, no API key.

The "director" framing isn't marketing. Look at the optional inputs and you'll see the job: 📦 上一个项目状态JSON and 🎬 上一段成片观察 exist so the node can continue a serial story with the state of the previous segment - real multi-segment continuity, not a one-shot prompt. And Review-成片复盘 / Repair-失败修复 are actual task modes in the 🎛️ Seedance任务 dropdown (auto-detect, 独立提示词, T2V, I2V, V2V, R2V, FLF2V, Edit, Extend, Sequence, Review, Repair).

The inputs that matter

• 🎛️ Seedance任务 - the 12 task modes. This decides what the LLM writes for.
• 🎨 创作类型 - 30 presets from 通用导演 through 极简产品广告, 真人电影叙事, 动作大片, 悬疑惊悚·真相反转, 误解喜剧·结尾反转, 一本正经·荒诞反差, 整蛊打脸·连环反转, 萌宠治愈日常, 邵氏复古武侠…
• 📝 原始视频需求 - your idea. The default is a solid brief: "电影感镜头，主体动作自然，画面稳定，声音与动作同步."
• ⏱️ 目标时长(秒) (4–15), 📐 视频比例, 🧩 目标分辨率 (自动/480P/720P/1080P), 🔊 原生音频.
• 🎬 首帧图 / 🏁 尾帧图 - for I2V / FLF2V. 🎞️ 每个视频采样帧数 (2–8, default 5) for reference videos.
• 🎧 参考音频原声直传LLM - send the audio spectrogram plus raw WAV when the model supports input_audio.
• 🖼️ 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 material ports.

The 🌐 输出中文提示词 toggle keeps official formatting while switching prose to simplified Chinese.

The outputs

🎬 Seedance最终提示词 is the one you wire into your Seedance node. The rest are the director's paperwork: 🎛️ 识别任务, 📑 导演与素材分析, 📦 项目状态JSON (feed this back into 上一个项目状态JSON for the next segment), 🧾 当前片段合同, 📄 LLM完整响应, and ℹ️ 处理信息.

Where people get burned

Serial work is where the node shines and where people trip: if you skip the 📦 项目状态JSON round-trip between segments, Sequence/Extend tasks lose their continuity and you get five disconnected clips instead of one story. For Review/Repair you need to actually describe the finished segment in 🎬 上一段成片观察 - the node reviews what you tell it, not what it can see. And like every vision node in this pack, reference images and videos demand a matched mmproj on the model.

Installing it

Standard pack install:

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

then requirements + restart. Pack-wide gotchas: llama-cpp-python pinned to the JamePeng fork at 0.3.47+ (stock PyPI 0.3.35 fails on Qwen3.8 GGUFs), NVIDIA users overlay a GPU wheel from the JamePeng releases, GGUF + mmproj in ComfyUI/models/LLM/. For a director that reads reference frames you want a real vision model - a text-only GGUF leaves the 参考图 inputs unusable.