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。
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。
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。
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.
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.
Episode / 第X集 splits, INT. / 内景 headers, CHARACTER: line dialogue — no reformatting neededBottom line: the productivity gap in AI drama is the script-engineering gap. Go structure your first script.