🎬03漫剧分镜与交付@炮老师的小课堂

The guided three-node 🐠 漫剧 flow ends at DapaoDramaFinish (compact 03): feed it the bundle from DapaoDramaVisual (02) and it generates the storyboard (分镜.md, shot-by-shot with frozen keyframes), writes the per-shot video prompts (skipping video for static 漫剧), optionally runs the text review, and finishes with the same delivery output the advanced Deliver node provides - prompt lists, JSON manifest, combined documents, and a delivery-check report. It's the widest node in the pack, but don't panic: 90% of the inputs are optional and the defaults are sane.

交付模式 decides the whole run

• 自动（按制作形态） - the default. Storyboard always; video prompts added only for dynamic forms (2D animated, live-action, etc.); static form skips video cleanly.
• 动态漫剧 / 静态漫剧 - force one or the other, overriding the project form.
• 仅分镜关键帧 / 仅视频提示词（复用已有分镜） - run just one stage when the other already exists.
• 仅文本审查 - review only.
• 仅整理已有提示词（不调用LLM） - the free mode: just collect and validate what's already in the bundle. Useful when you want to re-export without paying.

🔍 文本审查 toggles one extra LLM call at the end (default off), with 🔎 审查范围 to aim it.

Timing: the part people actually fight with

⏱️ 时长规划 controls how shots are timed. 自动分镜 lets the model decide. The two manual modes let you dictate exact per-segment seconds - 6,6,8,10 style via 逐段时长, or the friendlier 按段数逐段设置 which exposes 32 per-segment duration fields (🔢 分段数量 sets how many are used). Here's the design that saves your wallet: ⏳ 总时长规则 defaults to "must equal the 01 target seconds" - if your segments sum to 28s but the project says 30s, it refuses before any LLM call and tells you to fix the numbers or explicitly opt into "以本节点分段合计为准". No silent re-timing of your script on your dime. Each segment also has to respect the single-shot min/max you set in Prepare, and a model that ignores your manual plan stops the chain rather than papering over it.

The output row - and the pairing trap

Eleven outputs. The essentials: 📦 漫剧资料包, the three prompt lists (资产图片 / 关键帧 / 视频提示词), 📋 提示词清单JSON, 📚 五文档合订文本, ✅ 交付检查, and 💾 完整资料包JSON（续集用）. Plus the batch-focused ones: 🧩 所选批量提示词 (filtered by 🔢 提示词选择), and ⏱️ 全部视频时长列表 / ⏱️ 所选批量时长列表 - FLOAT lists of seconds.

That last part is a real footgun, so the docs spend a paragraph on it: if a video node needs durations, you must feed it the paired prompt and duration lists - filtered prompts need the filtered durations, and vice versa. And since these are FLOAT lists, a video node that only accepts integer seconds needs a numeric-conversion step per value; if it exposes fixed duration dropdowns, you execute per supported bucket instead. Get the pairing wrong and your 8-second motion prompt lands on a 4-second clip.

Money and installing

LLM stages bill per internal stage on the dapaoAI relay - dynamic mode's default chain is roughly 2 responses (storyboard + video) plus 1 more if review is on; static is 1. The "仅整理" mode calls nothing. Fixed seed + same content = 48-hour reuse; failures don't auto-retry. Install once via Manager → "dapaoAPI" or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI into custom_nodes, restart, Ctrl+F5. No model downloads - and remember this suite writes text only; the prompts it hands out are your input to a real image/video model, which is where the actual rendering spend happens.