🔍09漫剧审查@炮老师的小课堂

Every production pipeline needs a reviewer, and in the 🐠 漫剧 suite that reviewer is DapaoDramaReview (advanced node 09). Its whole design can be summarized as: it audits the existing documents against each other and against the evidence, and it does not rewrite anything. Output is a problem list - each with where the issue is, what it affects, who owns it (which upstream stage must fix it), and what the revision should say. Fixing is deliberately someone else's job.

What it checks

Give it an upstream bundle and it reads every document that exists - 原著分析, 系列开发, 剧本, 视觉设定, 图片提示词, 分镜, 视频提示词 - plus the project config. It looks for the things that make AI drama fall apart: story contradictions, characters whose stated looks drift between the visual setting and the prompts, cross-episode continuity breaks, prompt-level errors, and structural failures. It even has a mechanical layer that runs before the LLM does, fingerprinting every doc against its sources to catch stale or internally-inconsistent material. The review document it writes records a 结论 (verdict) and its findings; if you set a fixed seed, the same state re-reviews from cache rather than re-billing.

The important limit, stated plainly in the code: this is semantic review of text. It cannot guarantee plot quality, real character consistency in rendered images, or anything about media that doesn't exist yet. It catches that the storyboard references scene EP007-SC003 which the script never wrote; it can't tell you the rendered hero's face changed. The suite leans on you to do the final media QC.

The inputs that matter

Beyond the shared LLM-stage set (key/model/requirements/seed/tokens), everything comes in through the 📦 上游资料 bundle and the per-document text inputs - you can connect 原著分析.md, 系列开发.md, 剧本.md, 视觉设定.md, 图片提示词.md, 分镜.md and 视频提示词.md individually if you're not running from a bundle. The 📝 本阶段要求 box is where you focus the audit ("check only character continuity between script and image prompts"). Outputs are the standard four: bundle, 📄 阶段文档 (审查.md), an empty 🧩 提示词列表, and the processing report.

Where it sits

In the advanced flow it's optional, after the video stage and before delivery - you review, then route findings back to whichever stage owns the problem, revise there, and re-run the affected chain. In the guided flow, the finish node (DapaoDramaFinish) offers review as a toggle, and a "仅文本审查" mode runs just this stage. The delivery node will quote the review verdict in its report if one exists, and refuse to treat a stale review as current.

Installing and gotchas

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

The mental model to keep: review is a dead end, not a sink. If you let this node rewrite your script for you, you've broken the whole architecture - the suite's contract is that review only locates problems, and every other node refuses to auto-"fix" bad documents because a blind paid retry is how you burn money. Plan for the loop: review → fix the owning stage → re-run from there. Annoying? A little. Cheaper than letting an LLM silently "improve" your locked continuity locks? Yes, by a lot.