A forensic pathologist for short videos
rea-video takes any clip apart shot by shot and tells you whether each frame was filmed, AI-generated, or coded — then how to remake it.

What it does
rea-video is a Python toolkit built to run inside coding agents like Claude Code or Codex. You point it at a video and it splits it into shots, then classifies how each one was made: real footage, AI generation, digital human, 3D render, code-driven motion graphics, or screen recording. It also extracts subtitles with timing (exportable as SRT/ASS), sticker placements, cut rhythm, and audio details — voice vs. AI dub, BPM, where sound effects land. The output is a one-line verdict, a per-shot report.html, and a replication plan split by method.
The interesting bit
The judgment pipeline is evidence-based, not vibes: AIGC metadata flags and frame-rate regularity count as hard evidence and outrank “looks AI-ish.” When tools disagree, the disagreement is flagged and the agent re-examines frames — which is how it caught an AI MV that fooled the vision model, by spotting fake Chinese characters on a prop. In one case a tool-only pass misjudged 35% of a real unboxing video as AI; agent review corrected it to 97% real.
Key highlights
- Handles mixed videos: each shot judged separately, then reassembled on the original timeline
- Reverse-engineers prompts for AI shots and suggests how many generations they’d merge into
- Tool-only mode runs fully offline and free; adding a vision model (Gemini, Doubao) improves accuracy on platform-compressed videos
- Reported accuracy on held-out sets: ~99% by shot duration with vision + agent review, 88.6% tool-only on transcoded clips
- Pairs with the author’s cheap-claude-video project for “swap-the-elements” remakes that keep the original’s motion and timing
Caveats
- Phone-screen footage filmed in real life tends to get misclassified as 3D/AI by tools alone — the Douyin cases only reached 97.5% after vision-model review
- Platform-compressed videos lose noise evidence, so high accuracy there depends on a paid vision model
- It doesn’t output a finished video itself; real-shot segments still require you to actually film them
Verdict
Worth a look if you make short-form content and want to understand — or cheaply imitate — how a clip was produced. If you just want a one-click same-style video, the README itself suggests tools like Hypit are the faster route.
Frequently asked
- What is JuneYaooo/rea-video?
- rea-video takes any clip apart shot by shot and tells you whether each frame was filmed, AI-generated, or coded — then how to remake it.
- Is rea-video open source?
- Yes — JuneYaooo/rea-video is open source, released under the Apache-2.0 license.
- What language is rea-video written in?
- JuneYaooo/rea-video is primarily written in Python.
- How popular is rea-video?
- JuneYaooo/rea-video has 4.6k stars on GitHub.
- Where can I find rea-video?
- JuneYaooo/rea-video is on GitHub at https://github.com/JuneYaooo/rea-video.