Rea Video

8K stars
Apache-2.0
Active this week
Last commit today

Last verified Nov 14, 2026

Analyze any video shot by shot with an open source CLI tool that classifies footage, extracts prompts, and generates a production plan.

Visit website GitHub8K

Read the full breakdownWho it is for, how it works, trade-offs, the quick start and the FAQ

who it's for

Who Rea Video is for#

Plan a production from a reference video

Feed a reference video to rea-video before your production begins. Get a shot list showing what to film, what to generate, and what to build in code, with AI prompts for generated shots and lighting notes for filmed ones. Good for creators without a dedicated motion designer who need to understand a reference before budgeting or scripting.

Skip if:

You need a finished video immediately. rea-video produces a plan, not output footage. Use a generation tool directly if the goal is to create new video without first studying how a reference was made.

Connect to an AI video generation pipeline

Install rea-video as an agent skill inside Claude Code or Codex and use it to analyze reference clips before an automated generation run. The VIDEO.md output tells the agent which shots need generation, what parameters to use, and which video model fits the content type. Good for developers building automated video content workflows.

Skip if:

You need a browser-based or no-code workflow. rea-video is a local CLI and requires shell access from an AI coding assistant or a terminal.

Audit which shots in a video are AI-generated

Run rea-video on a finished video to verify which shots were filmed versus AI-generated, with measurement evidence per shot. The same detection pipeline that underpins production planning works as a shot-level audit log. Good for teams with AI content disclosure requirements or for reviewing content before publication.

Skip if:

You need a legally certified AI detection result. rea-video is an inference tool that flags its own uncertainty; it is not a certified or legally recognized detector.

the problem

The problem it solves#

When a creator studies a reference video, the central question is how it was made. A product promo might be 83% code-rendered motion graphics, 12% screen recording, and a handful of filmed moments. A TikTok ad that looks live might be entirely AI-generated. Without knowing this, a team might budget for a full live shoot when most of the work could be done in Remotion, or waste generation credits on footage that only needed a phone camera.

The difficulty is that standard video tools do not answer this question. Editing software handles footage you already have; AI generation tools produce new footage from prompts. Neither one works backward from a finished video to identify how each shot was made and what it would take to recreate it. rea-video fills this gap by measuring every shot: frame cadence, sensor noise, AIGC metadata tags, visual signals, and audio, then labeling each one and writing a material list and a route for the remake.

how Rea Video solves it

How it solves it#

Shot-type verdict with evidence

Every shot receives a classification: live footage, AI-generated, avatar, 3D or animation, code-rendered, or screen recording. Hard signals like AIGC metadata tags and frame cadence override softer visual signals. Conflicts are flagged for a visual check rather than resolved by guessing. The final verdict comes with the evidence that produced it.

VIDEO.md output for AI agents

Every breakdown is written in the open VIDEO.md format, structured in levels so an AI coding assistant reads only what a question needs. A full breakdown can run to hundreds of thousands of tokens; the layered format lets an agent read a card summary, an overview, a one-line-per-shot timeline, or a full shot file, depending on the query.

Audio and caption extraction

Captions and styled text are logged with start and end times, exported to SRT and ASS format, and annotated with color, outline, and entrance style. BPM, music key, AI-generated versus library music, and sound effect placement are all measured. Stickers, picture-in-picture elements, and text animations are mapped with their durations on the timeline.

Per-shot remake route

After classifying each shot, rea-video writes the production path: film it from the included shooting plan with lighting and framing notes, generate it with a suggested video model and a reverse-engineered prompt, or build it as a Remotion project with layout, text, and motion timing. AI shots are batched according to what the target model can actually handle.

Offline by default, vision model optional

Tools-only mode runs locally without any model endpoint and reaches above 98% accuracy by shot duration on original files. For videos downloaded from platforms, recompression strips noise evidence; adding a Gemini or Doubao vision model, or using an image-capable agent, brings accuracy back above 99%. There are no mandatory API keys or cloud dependencies.

strengths · trade-offs

Strengths and trade-offs#

Strengths

Evidence-first classificationEvery verdict traces to measurable signals. The classification order is hard evidence first: AIGC metadata and frame cadence; soft evidence only when hard signals are absent. The output states what evidence was used, so you can review any disputed shot without having to trust a black-box result.
Built for AI coding assistantsrea-video is designed to run inside Claude Code or Codex, not as a standalone app. An agent installs it, runs the breakdown, and reads the VIDEO.md output without any additional configuration. The whole pipeline runs from a single instruction to your AI coding assistant; you receive the results.
Handles all production types in one passLive footage, AI video, avatars, 3D and animation, code-rendered motion graphics, screen recordings, and mixed edits are all handled within a single breakdown. Each shot type has its own detection rules and remake path. A mixed-media clip with five different production methods requires one run, not five separate tools.

Trade-offs

-Requires a local video file and Python setuprea-video processes a file on disk, not a URL. You need to download the reference video before running a breakdown. Python 3.10 or later and ffmpeg must also be installed. The AI coding assistant path handles the Python setup automatically, but developers who prefer manual setup need to meet both prerequisites.
-Recompressed downloads need a vision modelPlatform downloads from TikTok, Douyin, and similar services are recompressed during upload, which removes the noise evidence that the offline tools rely on for certain calls. Tools-only accuracy drops to around 88% on these files. Adding a vision model endpoint or using an image-capable agent restores accuracy above 99%.
-Analysis only, no video outputrea-video produces a plan and prompts; it does not render footage. Live shots require a camera and filming equipment. AI shots require a generation tool like Runway or MiniMax. Code shots require a Remotion environment. The companion capsule-cinema tool automates AI shot generation from the breakdown, but live shots and code shots still need human execution.

versus alternatives

Rea Video vs alternatives#

rea-video vs Runway

Runway is a paid AI media platform that generates and edits video: image-to-video, text-to-video, AI effects, and motion brush controls. It does not analyze how an existing video was made. If you bring Runway a reference clip, you can use it as a visual guide for generation, but it will not identify which shots were filmed versus AI-generated, tell you the camera setup behind a particular angle, or produce a code plan for a motion graphic. rea-video and Runway address different steps: rea-video works backward from finished footage to explain how it was produced, while Runway works forward from prompts and assets to generate new footage.

rea-video vs Descript

Descript is a paid AI video and podcast editor that works from transcripts. Its AI features include voice cloning, automatic filler-word removal, screen recording, and timeline editing. These tools are oriented toward editing video you already own, not analyzing how a reference was produced. Descript does not classify shots by production method, detect AIGC signals, or generate a remake plan. If the goal is to edit a finished recording, Descript is a capable tool. If the goal is to understand and replicate a reference at the shot level, the two tools do not overlap.

What rea-video does not do

rea-video produces a production plan, not a finished video. It does not render footage, provide a timeline editor, or offer a no-code export path. After a breakdown, you still need generation tools for AI shots, a camera for live shots, and a code environment for code-rendered content. The companion capsule-cinema project automates the AI shot generation step from the VIDEO.md output, but live shots and code shots require human execution. rea-video belongs at the analysis and planning stage; production tools follow it.

install · quick start

Quick start#

bash

Requires Python 3.10+ and ffmpeg. Clone and install from GitHub:

```bash
git clone https://github.com/JuneYaooo/rea-video
cd rea-video
pip install -e .
```

tech stack · detected from GitHub

What it's built on#

Languages: Python
Tooling: FFmpeg

frequently asked

FAQ#

Does rea-video need a paid model to run?

No. Tools-only mode is fully offline and free. On original, uncompressed video files, the offline tools reach above 98% accuracy on the production method classification weighted by shot duration. If you want to improve accuracy on recompressed platform downloads, connecting a vision model (Gemini or Doubao both work) helps, but that step is optional. If your AI coding assistant can see images, it can also check uncertain shots itself without a separate model endpoint.

What do I need to get started?

You need Python 3.10 or later, ffmpeg, and a local video file. The quickest path is to open Claude Code or any AI coding assistant that can run shell commands and ask it to install rea-video from GitHub and run a breakdown. The agent handles the install and runs the full pipeline. For a manual setup, clone the repository and install it with pip.

What is VIDEO.md?

VIDEO.md is an open structured format for video breakdowns. Rather than dumping all analysis data at once, it lays out the breakdown in levels: a card summary, an overview, a one-line-per-shot timeline, and one file per shot. Each level states its size, and observations from visual inspection are kept separate from tool measurements. rea-video writes one into every output folder so AI agents read it token-efficiently. Other tools can read it too using the videomd Python package.

How accurate is the shot-type classification?

On original uncompressed video files, the offline tools reach above 98% accuracy weighted by shot duration, and above 99% with a vision model or agent visual review. On recompressed platform downloads, tools-only accuracy is around 88%; adding a vision model brings it back above 99%. Shot boundary detection reaches 95% recall and 94% precision on hand-labelled clips. Full numbers and test cases are in eval/RESULTS.md in the repository.

Can rea-video produce the finished video?

No. rea-video analyzes a reference and writes a production plan. AI shots get reverse-engineered prompts and a generation batch plan. Code-rendered shots get layout, text, and motion timing for a Remotion project. Live shots get a shooting plan with camera framing and lighting notes. For the AI shot generation step, the companion capsule-cinema tool automates output from the breakdown, but live shots and code shots still require human execution.

also worth a look