展示HN:AI漫画生成器 – 将剧本转化为动画视频
我构建了一个可自托管的管道,它可以将剧本(或一句话的创意)转化为完成的动画视频——无需编码,也无需SaaS订阅。
管道步骤:
1. 编写或通过人工智能生成剧本
2. 提取角色 → 生成角色转身参考图(以确保视觉一致性)
3. 将剧本拆分成镜头,并添加摄像机方向和动作描述
4. 生成关键帧(每个镜头的第一帧和最后一帧),并保持连续性
5. 将每个镜头插值为视频片段
6. 使用FFmpeg合并所有内容,添加字幕
每个阶段都可以单独触发或批量处理,因此你可以保持控制。
技术:
- Next.js 16 应用路由 + SQLite(自包含,无需基础设施)
- 自带密钥:OpenAI、Gemini或Seedance用于文本/图像/视频
- Docker一行命令:docker run -p 3000:3000 -v ./data:/app/data twwch/aicomicbuilder:latest
- 无需账户——浏览器指纹(用户代理/屏幕/时区的SHA-256)作为用户身份,因此多个用户可以共享一个实例,同时数据完全隔离
我为什么要构建它:现有的视频生成工具要么是单次生成(一个提示→一个片段),要么是完全不透明的管道。我想要一个能够展示每个步骤的工具,这样我可以在提示上进行迭代,在项目中途更换模型,并理解实际发生的事情。
最难的部分是连续性链——镜头N的最后一帧成为镜头N+1的第一帧,这样可以在剪辑之间保持角色的视觉一致性,而无需微调。
GitHub: https://github.com/twwch/AIComicBuilder
自述文件中有演示视频。希望能收到关于管道设计的反馈,以及指纹作为身份验证的权衡是否适用于这个用例。
查看原文
I built a self-hostable pipeline that takes a screenplay (or a one-line idea) and produces a finished animated video —
no coding required, no SaaS subscription.<p><pre><code> The pipeline:
1. Write or AI-generate a script
2. Extract characters → generate turnaround reference sheets (for visual consistency)
3. Split the script into shots with camera direction and motion descriptions
4. Generate keyframes (first + last frame per shot) with continuity chaining
5. Interpolate each shot into a video clip
6. Concat everything with FFmpeg, burn subtitles
Each stage can be triggered individually or in batch, so you stay in control.
Tech:
- Next.js 16 App Router + SQLite (self-contained, zero infra)
- Bring your own keys: OpenAI, Gemini, or Seedance for text/image/video
- Docker one-liner: docker run -p 3000:3000 -v ./data:/app/data twwch/aicomicbuilder:latest
- No accounts — browser fingerprint (SHA-256 of UA/screen/timezone) is the user identity, so multiple people can share
one instance with full data isolation
Why I built it: Existing video generation tools are either single-shot (one prompt → one clip) or fully opaque
pipelines. I wanted something that exposed every step so I could iterate on prompts, swap models mid-project, and
understand what's actually happening.
The hardest part was the continuity chain — the last frame of shot N becomes the first frame of shot N+1, which keeps
characters visually consistent across cuts without fine-tuning.
GitHub: https://github.com/twwch/AIComicBuilder
Demo video in the README. Would love feedback on the pipeline design and whether the fingerprint-as-auth tradeoff
makes sense for this use case.</code></pre>