剧本工程化:把故事变成可执行的分镜 JSON

更新于 2026-11 · 阅读约 5 分钟

很多新手把小说直接粘进视频模型,抽了 20 次卡还是得不到想要的画面。原因很简单:视频模型需要的是"镜头指令",不是"故事"。剧本工程化就是把故事翻译成机器指令的过程。

一份工程化剧本包含什么

Schema 长什么样

这是 DramaKit 转换器输出的核心结构(节选):

{
  "characters": [
    { "id": "c1", "name": "林晚",
      "visualTags": "25岁女性,黑色长发,白衬衫,工牌挂脖",
      "personality": "冷静、执拗" }
  ],
  "scenes": [
    {
      "id": "s1", "episode": 1,
      "location": "写字楼电梯", "intExt": "INT", "time": "day",
      "characters": ["林晚"],
      "dialogue": [{ "character": "林晚", "line": "……" }],
      "shots": [
        { "shotId": "s1-1", "type": "medium",
          "durationSec": 4,
          "videoPrompt": "中景,写实电影感……保持角色面部与服装一致。" }
      ]
    }
  ]
}

有了这个结构:写提示词时直接引用 visualTags 保持一致;配音按 dialogue 逐条生成;剪辑按 shots 里的 durationSec 排时间线。整个下游都被这一个 JSON 驱动。

5 分钟实操

  1. 打开 剧本结构化转换器,点"载入示例"看格式,或直接粘贴你的剧本
  2. 识别规则:第X集 / Episode 分集、内景 / INT. 场景头、角色名:台词 对白——都是最常见的写作习惯,不需要改格式
  3. 填入剧名、题材、视觉风格(会成为每个镜头 videoPrompt 的一部分)
  4. 解析后复制 JSON 下载存档;把每场戏的 shots 逐个喂给视频模型
  5. 对镜头语言有要求时,用 Prompt 构建器重写单个镜头

常见坑

一句话总结:AI 短剧的生产力差距,就是剧本工程化的差距。现在就去 结构化你的第一份剧本。

Script Engineering: Story to Executable Storyboard JSON

Updated Nov 2026 · ~5 min read

Beginners paste a novel straight into a video model and burn 20 generations without getting the shot they imagined. The reason is simple: video models need shot instructions, not stories. Script engineering is the act of translating story into machine instructions.

What an engineered script contains

What the schema looks like

Core structure produced by the DramaKit formatter (excerpt):

{
  "characters": [
    { "id": "c1", "name": "Lin Wan",
      "visualTags": "25yo woman, long black hair, white shirt, ID badge",
      "personality": "calm, relentless" }
  ],
  "scenes": [
    {
      "id": "s1", "episode": 1,
      "location": "office elevator", "intExt": "INT", "time": "day",
      "characters": ["Lin Wan"],
      "dialogue": [{ "character": "Lin Wan", "line": "..." }],
      "shots": [
        { "shotId": "s1-1", "type": "medium",
          "durationSec": 4,
          "videoPrompt": "Medium shot, cinematic realism... keep character consistent." }
      ]
    }
  ]
}

Once you have this structure: prompts reference visualTags for consistency; dubbing is generated per dialogue line; editing follows durationSec. One JSON drives the entire downstream.

Five-minute walkthrough

  1. Open the Script Formatter, click "Load sample" to see the format, or paste your own script
  2. The parser reads common conventions — Episode / 第X集 splits, INT. / 内景 headers, CHARACTER: line dialogue — no reformatting needed
  3. Fill in title, genre, visual style (these become part of every videoPrompt)
  4. Parse, copy and archive the JSON; feed shots to your video model one by one
  5. For finer camera control, rewrite individual shots in the Prompt Builder

Common pitfalls

Bottom line: the productivity gap in AI drama is the script-engineering gap. Go structure your first script.