🎬07漫剧分镜与关键帧@炮老师的小课堂

Between "here's the script" and "here's the video" sits the part most AI-video beginners skip and then regret: the shot list. DapaoDramaStoryboard (advanced node 07) is where the 🐠 漫剧 pipeline breaks an episode's scenes into individual SHOT-* entries - each with a source scene, a duration, a defined start → single action → end, a visual reference, a frozen keyframe prompt, and a plan for which reference images it needs. It needs both the script and the visual setting (视觉设定.md), because a shot that references a character who was never asset-listed is a shot that can't be drawn.

Why the per-shot contract matters

Every SHOT-* entry carries a fixed set of fields: 来源 (source scene, e.g. EP007-SC003), 时长 (duration in seconds), 起点 (start state), 唯一动作 (the one action), 终点 (end state), 视觉依据 (which 人物/地点/道具 it leans on), 图片提示词项 (which IMG assets apply), and 输入参考图 (which reference images to hang - the PLAN- mechanism). Under each shot sits a ### 冻结关键帧提示词 block - a frozen keyframe prompt that captures the shot's starting frame as an image you'd generate first, then animate into.

The suite mechanically verifies this contract and it's strict: every script scene must be covered by a shot or explicitly listed as 未拍 (unshot); visual references must actually exist in the assets doc; IMG references must exist; durations must respect the min/max seconds you set back on the project node; and the whole shot count must equal the manual timing plan if you gave one. Miss any of it and the node returns your draft for revision but refuses to pass it downstream - no auto-"fix", because every paid retry is a new bill.

How it works

Standard LLM-stage plumbing (key/model/requirements/seed/tokens/timeout), driven by three upstream docs. Optional 📄 图片提示词.md input matters if your IMG identity boards already exist - the storyboard can reference them as keyframe sources, which is how the串接 (serial) wiring 05 → 06 → 07 works. If you skipped image prompts, shots fall back to freezing their own keyframe.

Inputs and outputs that matter

Required upstream: 剧本.md and 视觉设定.md (via bundle or text inputs). Optional: 图片提示词.md. Outputs: 📦 漫剧资料包, 📄 分镜文档, and - the one you came for - 🧩 关键帧图片提示词列表, one complete keyframe prompt per shot, in the same order as the shots.

Installing and gotchas

Pack install: Manager → "dapaoAPI", or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI into custom_nodes, restart + Ctrl+F5. No model downloads; LLM spend on the dapaoAI relay.

Biggest conceptual trap: the storyboard is not a video. Its frozen keyframes and PLAN- references describe images that don't exist yet - the docs are adamant that PLAN- is a plan, never proof of generated pixels, and the code rejects any REF- claim that it verified real images. You generate the keyframes yourself, hang them where the plan says, and only then feed them to your image-to-video model. Think of node 07 as the blueprint and your own image runs as the construction. If that split annoys you, you're right that it's more manual work - but it's also the only honest way to keep a text-only suite from lying about what it can see.