JuneYaooo/rea-video

逆向一切视频：实拍口播、桌拍、AI 视频、3D、代码动效、混剪，逐镜拆出画面、花字、剪辑和声音是怎么做的，并给出复刻方案。Reverse-engineer any video, shot by shot

Summary

rea-video is a Python, agent-oriented tool that breaks a video into shots and estimates whether each shot was filmed, AI-generated, produced as 3D or animation, or built with code. It turns that analysis into a remake plan and HTML report rather than rendering a replacement video itself.

What it is useful for

Engineers can use it to understand a reference video before rebuilding it: extracting cut timing, camera and lighting notes, caption styling, music and sound-effect placement, and the assets that must be recorded or supplied. The README says caption timing can be exported as SRT or ASS, while code-driven sections include text, layout, and motion timing suitable for rebuilding as a Remotion project.

It handles mixed media shot by shot. For example, the README describes a tutorial combining AI dialogue, a coded interface, and eight seconds of real screen recording. Optional Gemini or Doubao integration adds reverse-engineered prompts for individual shots and can improve analysis of recompressed videos. No video-capable model is required: tools-only mode runs offline, while an agent that can inspect images can review uncertain shots.

This is analysis rather than generation. rea-video does not render the final video; automatic AI generation and timeline assembly are partly handled by the separate capsule-cinema project. The README warns that platform recompression removes useful noise evidence, so bundled re-encoded samples may produce different results from the original files. It also notes that the documentation is currently in Chinese and advises remaking videos only with material you have the rights to use.

How engineers can use it

The documented quickest path is through Claude Code, Codex, or another agent that can run commands:

Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4

FFmpeg and Python 3.10+ are required. For manual setup, the README documents:

brew install ffmpeg            # or: sudo apt install ffmpeg
git clone https://github.com/JuneYaooo/rea-video && cd rea-video
bash scripts/install_skill.sh
rea-video --help

The installer uses pip or pipx, or a local .venv if pip is blocked, and links the skill into ~/.claude, ~/.codex, and ~/.agents. Run the analysis on a local MP4, review report.html, and use the remake plan to decide what to film, generate, or implement in Remotion. For per-shot prompts, give the agent the endpoint and key for a supported video-capable model. For an automated remake workflow, consult the repository’s docs/replicate.md.