rea-video: A shot-by-shot breakdown of how a video is made

A video analysis tool for agents like Claude Code and Codex: it splits videos into shots, separates vocals, music, and sound effects, and then determines whether each shot is live-action, AI-generated, a digital human, 3D, or code animation based on frame rhythm, sensor noise, AIGC markers in metadata, and visual details. It outputs subtitle styles and timelines, editing rhythm, background music BPM, sound effect positions, and a replication plan detailing "what to shoot, what to generate, and what code to write" along with a report.html; when integrated with vision models like Gemini and Doubao, it also reverse-engineers prompts for each shot.

Original title: JuneYaooo/rea-video

A video analysis tool for agents like Claude Code and Codex: it splits videos into shots, separates vocals, music, and sound effects, and then determines whether each shot is live-action, AI-generated, a digital human, 3D, or code animation based on frame rhythm, sensor noise, AIGC markers in metadata, and visual details. It outputs subtitle styles and timelines, editing rhythm, background music BPM, sound effect positions, and a replication plan detailing "what to shoot, what to generate, and what code to write" along with a report.html; when integrated with vision models like Gemini and Doubao, it also reverse-engineers prompts for each shot.

Why it matters When replicating ads or selecting topics for short videos, clarifying what needs to be shot, what can be generated, and what can be coded helps avoid unnecessary detours.

Stars 2091 (+1184 / 24h)

Reasoning

  • 1184 new stars in 24 hours, total stars 2091 (indicating attention only).
  • Rapid star growth for a new repository is marked as "to be observed".

Usage

  • Requires ffmpeg and Python 3.10+; simply tell the agent "install this repository, then analyze a specific video"
  • Optional: Provide the agent with a model interface that supports video input to improve judgment accuracy

Risks

  • The accuracy rates in the README come from the author's custom test set; videos downloaded from platforms undergo secondary compression, so relying solely on local tools for judgment may yield poorer results
  • Detailed documentation is currently primarily in Chinese; for a new repository, star growth only indicates attention

Open

GitHub

Explore the latest AI Radar updates