JuneYaooo/rea-video

What it does

rea-video is an agent-driven workflow that breaks a reference video into individual shots and explains how each one was likely made. It covers live-action talking-head and tabletop footage, AI-generated clips, digital humans, 3D/animation, code-driven motion graphics, screen recordings, and mixed edits. For every shot it reports visual style, on-screen text, editing rhythm, and audio treatment, then proposes a recreation plan (9).

The package is designed for command-capable AI agents such as Claude Code and Codex rather than as a standalone GUI (8). It combines offline media probes (frame-rate regularity, noise, file metadata, AIGC markers), optional vision-model inspection, and audio analysis for BPM, key, and speech versus singing, giving hard evidence priority over appearance (15).

Quick start

Prerequisites: Python 3.10 or newer and ffmpeg installed on the machine. In Claude Code, Codex, or a similar agent, send an instruction such as: install the repository, then analyze /path/to/video.mp4 (11). The agent installs the command-line tool, reads the decomposition procedure, and runs the pipeline end to end.

The output includes a one-line conclusion, a recreation plan split by live-action, AI generation, and code, and a report.html with per-shot screenshots, evidence, prompts, shooting notes, and audio breakdown. To improve accuracy on compressed platform downloads, provide a vision-capable model such as Gemini or Doubao by giving the agent its endpoint and key; the tool still runs offline without one (20).

Constraints

• The workflow depends on an agent that can run shell commands; it is not described as a one-click desktop application.
• ffmpeg and Python 3.10+ are required before analysis can run.
• Offline-only analysis is reported as accurate on original files but noticeably weaker on transcoded social-platform downloads; README figures put the production-method accuracy at 88.6% on transcoded samples versus 99%+ with a vision model or agent review (13).
• Automatic recreation is partial: AI shots can be generated and placed back on the timeline with the companion capsule-cinema tool, while live-action shots still need to be filmed and code shots need a Remotion project written (21).
• The README explicitly says not to reuse other people's footage, music, characters, or likeness, and to follow platform rules and local AI-content labeling requirements (24).
• No license is declared in the repository metadata.

HotRepo's take

On the HotRepo new-project board for Python on 2026-11-03, the repository recorded 111 stars and 25 forks, adding 6 stars in the roughly one-hour window between 19:45 and 20:45 UTC. That is a single snapshot with no daily all-language rank recorded, so it supports only the observation that the project drew brief early attention, not a trend claim. The more substantive signal is the evaluation table covering 41 test videos with separate tuned and held-out sets and explicit accuracy differences between offline and vision-assisted runs (14); that kind of disclosed methodology is what makes the reverse-engineering claims checkable.