2026 年第二季度,全国新上线的微短剧里 AI 生成作品占比已超过六成。一部过去需要编剧组两周的案头工作, 现在一个工作周就能交付。这篇指南把整条链路拆成八步,每步给出“做什么 + 用什么”。
短剧生死在前 3 秒。先定题材(战神归来 / 复仇打脸 / 先婚后爱),再写“每集一个钩子、三集一个大反转”的爆点分布表。 让大模型做竞品分析和受众画像,人只负责立意。
传统剧本是给演员和导演看的;AI 短剧剧本必须面向机器生产。每一场戏都要有:场景 ID、出场角色、对白、镜头类型、预估时长。 可用的 LLM:DeepSeek(逻辑严谨)、豆包(中文语境)、Kimi(长上下文,约 12.8 万 token 可撑 20-30 集)。
把剧本转成结构化格式:项目总纲 + 人物圣经(可视化标签)+ 场景圣经 + 分镜表。 这是决定成片一致性的关键一步——可以直接用我们的 剧本结构化转换器,粘贴文本即可生成含 videoPrompt 字段的分镜 JSON。原理见剧本工程化详解。
用图像模型(即梦、Midjourney、SD+ControlNet)生成角色定妆照和场景概念图。 这些参考图之后要喂给视频模型锁定一致性——脸一旦变了,观众三秒弃剧。
主流引擎:Seedance 2.5(字节系,30 秒连续生成 + 多模态参考)、可灵 3.0(动作与物理真实感)、 海螺 H3(性价比量产)、Sora / Veo 3.1(海外项目)。写提示词建议用 Prompt 构建器按平台分别生成。
关键省钱技巧:分镜分级。远景过场用低分辨率抽卡,特写情绪镜头才上高配,渲染成本能降约 40%。
抽卡可用率通常只有 30%,量产必须上自动审片(画面可用性、角色一致性、逻辑纠错), 像巨日禄这类 Agent 能把可用率拉到 90%。个人创作者也可以人工只审关键镜头。
中文首选讯飞智作(拟真度最高、支持方言),英文用 ElevenLabs。 Veo 3.1 和部分新模型支持原生音频输出,可以连台词一起生成。
剪映完成精剪、字幕、多平台导出。出海是增长最猛的口子:海外 AI 剧/漫剧市场规模同比暴涨 550%, 多语种译制(趣丸、腾讯云一站式方案)已能批量完成,人只打磨文化禁忌与本土化细节。
整个流程里,真正拉开差距的不是单镜画质,而是第 3 步的工程化程度。 从 结构化你的剧本 开始。
By Q2 2026, AI-generated titles made up over 60% of newly released microdramas in China. Case work that once took a two-week writers’ room now ships in under a week. Here is the full pipeline, broken into eight steps with tool picks for each.
Short drama lives or dies in the first 3 seconds. Pick a genre (billionaire revenge / sweet-marriage / underdog rise), then plan a “hook per episode, big reversal every three episodes” beat sheet. Use LLMs for competitor analysis and audience profiling; keep the theme human.
Traditional scripts are written for actors and directors. AI-drama scripts must be written for machines: every scene needs a scene ID, characters, dialogue, shot type and estimated duration. Good LLMs: Claude or GPT for creative range, Kimi or Gemini for long context (a 128k-token window covers roughly 20–30 episodes).
Convert the script into a structured format: project bible + character bible (visual tags) + scene bible + shot list. This step decides consistency more than anything else — run your text through our Script Formatter to get storyboard JSON with ready-to-paste videoPrompt fields. See Script Engineering explained.
Generate character reference sheets and concept art (Midjourney, Jimeng, SD + ControlNet). These references get fed into the video model to lock consistency — once faces drift, viewers bail in three seconds.
Main engines: Seedance 2.5 (30s continuous clips + multimodal reference), Kling 3.0 (physics-aware motion), Hailuo H3 (cost-effective volume), Sora / Veo 3.1 for western-facing projects. Generate per-platform prompts with our Prompt Builder.
Cost trick: shot tiering. Render wide transitional shots at low resolution and spend high settings only on close-ups and emotional beats — cuts render cost roughly 40%.
Raw acceptance rates hover around 30%. Volume production requires automated review (usability, character consistency, logic) — agent pipelines like Juirilu push acceptance above 90%. Solo creators can manually review only key shots.
ElevenLabs for English emotional range; Chinese productions favor Xunfei Zhizuo. Veo 3.1 and several newer models support native audio generation, including dialogue.
CapCut / Jianying for the final cut, captions and multi-platform export. Going overseas is the fastest-growing lane: the international AI-drama market is up ~550% YoY, and one-stop translation pipelines now handle bulk localization — humans focus on cultural taboos and local nuance.
The real differentiator across this pipeline is not single-shot beauty — it’s how much of Step 3 you engineer. Start by structuring your script.