🎞️08漫剧视频提示词@炮老师的小课堂

The 🐠 漫剧 text pipeline reaches its video moment at DapaoDramaVideo (advanced node 08). It takes the storyboard's shots and writes one complete MOTION-* video prompt per shot: what moves, the camera, the performance, and a sound/dialogue timeline. But read the description once more: "仅输出文字" - text only. This node has never generated a frame and never will. Its job is to hand the video model (or you, paste-by-paste) prompts it can actually execute.

The dialect dropdown you set back at the project node

Here's the detail that tells you this was built by people who run real Chinese AI-video stacks: the prompt style is chosen per dialect, set once in the project config - 通用 (generic), Seedance 2.0, Seedance 2.5, or MiniMax H3. Each loads a grammar reference so the phrasing matches what that model family actually responds to (ByteDance's Seedance is the short-drama workhorse; MiniMax's H3 is the omni-modal open release with native audio). The node even writes a dedicated music-timeline section if you ask, without pretending it can call a music model.

The catch worth knowing: the dialect only changes how the LLM writes prompts. The LLM doing the writing is whatever you pick in the shared 🤖 LLM模型 dropdown, and that model's real, current capabilities are your responsibility. The docs are blunt that the dialect grammar files are a snapshot, not a guarantee about what a given model version can produce today.

The one-to-one contract

Video prompts must map exactly to storyboard shots: every SHOT-* gets exactly one MOTION-*, in order, with matching 时长 (duration) and 输入参考图 (reference-image binding). The node mechanically cross-checks every entry against the storyboard and flags drift - a reference binding that changed, a shot skipped, a "待补参考图" (reference still to be filled) that slipped through. If the project is in 挂图计划 (hang-image plan) mode, the prompt is written to consume specific reference frames you'll provide; only in explicit 明确文生视频 mode does it write a self-contained text-to-video prompt with no image needed.

Inputs and outputs

Required: 剧本.md, 视觉设定.md and 分镜.md (from node 07), via bundle or direct text. Outputs: 📦 漫剧资料包, 📄 视频提示词文档, and the 🧩 视频提示词列表 - one complete prompt per shot. Because the full list and the storyboard are one-to-one, the paired duration lists (available on the guided finish node) let you hand prompt-plus-duration to a video node that wants timing as data.

Installing and gotchas

Pack install: Manager → "dapaoAPI" or clone into custom_nodes; restart; Ctrl+F5. No weights - the only spend is LLM usage on the dapaoAI relay (fixed seed + same content = 48-hour reuse; changes re-bill).

Two traps. First: copy the whole quoted block for each MOTION entry - multi-paragraph prompts including sound sections are one prompt, and grabbing the first line silently loses your audio and camera direction. Second: don't expect this node to keep characters consistent by itself. The suite's text checks guarantee the words match across shots; the actual rendered consistency still depends on your reference images and the video model's i2v behavior. Its prompts are the map - the model still has to walk the terrain.