🎭05漫剧视觉资产@炮老师的小课堂

The most frequently cited unsolved problem in AI image generation is keeping one character looking like the same person across shots - and the 🐠 漫剧 suite's answer is to do the hard part in text, before any image generator gets involved. DapaoDramaAssets (advanced node 05) is the stage that takes your finished script and breaks it into the visual assets an episode needs: 人物 (characters), 造型 (looks/variants), 地点 (locations) and 道具 (props), each with identity anchors, and - the key concept - 连续性锁 (continuity locks).

What it does (and deliberately doesn't)

Its description draws the line sharply: it decomposes people, looks, places, props, variants, sound direction and continuity locks, and it does not write image prompts. That's node 06's job. Think of it as the production bible for the episode - the "who is this character, what defines their identity, what can change between shots and what can't" document (视觉设定.md). When a character only has a voice and never appears on screen, it says so instead of inventing a face - the suite is allergic to fabricating unconfirmed details.

The continuity-lock mechanism is the clever bit for multi-shot drama. Locks are written as text phrases like "blue school uniform with the red scarf" that the image, storyboard and video stages are required to preserve verbatim in their prompts. Later stages get mechanically checked for those exact phrases. It's the text-layer equivalent of an IP-Adapter - and unlike an adapter, it costs nothing at inference and transfers to any image model you happen to use that week.

How it works

Standard LLM stage plumbing: 🔑 API密钥 (or DAPAO_API_KEY), 🤖 LLM模型, 📝 本阶段要求, seed, token budget, timeout. It requires the script (剧本.md) - either through the upstream bundle or by pasting it into the 📄 剧本.md input - because you can't break a story into assets you never read. Output is the updated 📦 漫剧资料包, the 📄 阶段文档 (视觉设定.md), a processing report, and an empty 🧩 提示词列表 socket (don't expect prompts here).

Where it sits in the flow

Assets feeds two children at once: DapaoDramaImage (node 06, which turns assets into IMG- identity-board prompts) and DapaoDramaStoryboard (node 07, which needs both script and assets). That's why the branch-and-merge shape exists: you can run asset-image prompts down one branch while the storyboard consumes the same visual bible down another, then re-join with DapaoDramaMerge (node 11). In the guided three-node flow, the equivalent work happens automatically inside DapaoDramaVisual (compact 02).

Installing and gotchas

Pack install, once: Manager → "dapaoAPI" or git clone https://github.com/paolaoshi/ComfyUI-dapaoAPI into custom_nodes, restart, Ctrl+F5. No weights.

Real-world warnings: (1) the continuity-lock scheme only works if you keep the same IMG assets and their original prompts through the whole project - the suite's cross-episode rules forbid silently rewriting a locked identity; if a character's look genuinely changes, you record it as an explicit variant with before/after state, not a quiet edit. (2) These are text locks, so they constrain your prompts, not the diffusion model's output - the docs are upfront that literal phrase-presence can't guarantee a model renders a consistent face. You still need good reference images at generation time; this node just makes the prompt half of that fight winnable.