What is Rea Video?

The 30-second answer

Rea Video is a browser-based video reverse-engineering tool. You drop a clip into the workbench; it finds the cuts, extracts reference frames, measures each shot’s light, color, motion and rhythm, and writes a shot-by-shot prompt pack for AI video generation. The file is decoded locally with the browser’s File API, Canvas 2D and Web Audio; there is no upload endpoint.

rea-video (the hyphenated name) is a separate open-source project by JuneYaooo. It is a Python CLI that runs ffmpeg on your machine, reads the file’s metadata, does OCR, separates stems, and can call a vision model. Its one-line brief is “reverse-engineer any video, shot by shot: filmed, AI-generated or coded, and how to remake it.”

What the open-source rea-video CLI does

The project is built for AI agents: you point an agent at the repo and a video, and it installs the CLI, learns the workflow and runs the breakdown. It needs Python 3.10+ and ffmpeg. The first run downloads a roughly 67 MB voice-separation model; a 30-second clip takes about a minute without a vision model.

It sorts each shot into one of eleven production categories:

• Live footage, AI-generated video, avatar / lip-sync
• 3D render, 2D animation, stop motion, film / TV excerpts
• Code or template rendering, screen capture, still image with motion, static hold

Around the visual call, it measures cuts, repeated frames, optical-flow motion, texture and sensor noise, frame cadence, container metadata, overlays and audio. It can export a breakdown.json, a human-readable report.html, a storyboard.jpg, first and last frames for every shot, and a packaging/ folder with SRT/ASS captions and a timeline.

The project publishes accuracy numbers from its own evaluation sets: 98.3% on its tuning set, 99.6% / 98.9% on its holdouts, and 88.6% on re-encoded WeChat videos with the offline tools alone. Those are the project’s self-reported numbers on its own samples, not this site’s. The methodology and the reproduction steps live in its eval/RESULTS.md.

What this browser workbench does

The browser version is deliberately smaller. It samples the video at 5 fps (3 fps for longer clips), detects cuts with a histogram distance plus a persistence gate, then measures each shot: luma key, contrast, color temperature, a five-color palette, edge density, flat-block ratio, estimated global motion, letterboxing and a beat grid from the audio. It extracts a first and mid frame per shot and writes prompts.md, prompts.json, shots.csv, storyboard.jpg, breakdown.json and a self-contained report.html.

Which one should you use?

Use this browser workbench when you want a fast, private first pass: you have a short clip, you need the shot list and reference frames, and you want a prompt scaffold without installing anything. It is also useful when the video is sensitive and should not leave your machine.

Use the rea-video CLI when the question is forensic: is this shot AI-generated or filmed? What does the container metadata say? Are there AIGC or C2PA tags? What do the captions say? What are the separated voice, music and effects? It is the right tool for a long-form edit or a verdict you need to defend.

Install the CLI

The project documents the manual path. You need Python 3.10+ and ffmpeg:

git clone https://github.com/JuneYaooo/rea-video.git
cd rea-video
pip install -e ".[ocr]"
rea-video --help
bash scripts/install_skill.sh   # optional agent skill links

Run the deterministic breakdown first, then add a vision model if you need it:

rea-video decompose ref.mp4 -o out/ref
# optional:
export REA_VLM_BASE_URL=https://.../v1
export REA_VLM_API_KEY=...
export REA_VLM_MODEL=gemini-2.5-pro
rea-video decompose ref.mp4 -o out/ref --vlm
open out/ref/report.html

Attribution: the commands above are from the project’s own CLI documentation, read on 8 Nov 2026. The project is Apache-2.0 licensed. This site does not bundle or redistribute its code.

The name, disambiguated

“Rea video” is not a unique phrase. There is the JuneYaooo rea-video project, the older REA (Reverse Engineer Anything) idea by morluto that the project credits as its inspiration, and a large amount of music content around the musician Chris Rea. This site is none of those things: it is an independent browser workbench that borrows the workflow name and credits the open-source project.

If you searched for Chris Rea’s music videos: wrong reel. If you searched for the open-source CLI: the GitHub repository is the source of truth. If you want to try the workflow without installing anything, the workbench is on the home page.