rea-video README

项目的 README,来自 github.com/JuneYaooo/rea-video。合并为一个 markdown 文件:README.md。

rea-video

Reverse-engineer any video.

Shot by shot: was it filmed, AI-generated or coded, and how to remake it.

Every breakdown is written in VIDEO.md, an open format that agents read layer by layer.

Website | 中文 | English | Русский | 日本語 | 한국어 | Español | Português | Français | Deutsch | Tiếng Việt | বাংলা | Türkçe

Stars Python Agent Output: VIDEO.md

JuneYaooo/rea-video | Trendshift

Visit our website Join our Discord WeChat group Follow on X

⭐ Help more people find rea-video. Star this repo!

Left: TikTok original shot on a phone; right: someone's AI remake with a different product

Left: a TikTok ad shot on a phone. Right: an AI remake with another product. rea-video's verdict: the left is all live footage, one set and two camera positions; the right is all AI, 8 shots across 3 generations.

Quick start

rea-video is built for AI agents. In Claude Code, Codex or any agent that can run commands, say:

Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4

The agent installs the CLI, learns the workflow and runs the whole thing. You need ffmpeg and Python 3.10+.

Breakdowns are fully upgraded. Each one gives you two outputs:

You can also hook up a video-capable model (Gemini or Doubao both work) by giving the agent its endpoint and key. That makes the live/AI/3D calls more accurate and adds a reverse-engineered prompt for every shot. It works fine without one; if your agent can see images, it checks the uncertain shots itself.

Reports are written in whichever language you ask in (Chinese or English). The example reports below are in English.

What it can break down

Video type Examples What you get
Live footage Talking-head, tabletop, handheld product demo, vlog, store visit Setups and camera positions, framing, camera moves, lighting (12 setups with gear lists), color grade
AI video Text/image-to-video, AI short drama, AI music video Evidence, a reverse-engineered prompt per shot, how to batch into generations and which model to use
Avatar A portrait driven by audio, lip-sync Which shots are avatar shots, what portrait and audio you need
3D / animation Blender / C4D explainers, product animation, 2D animation, stop-motion Whether to generate with a video model or render from 3D code
Code / template Kinetic type, UI demos, charts, slide-style cards Text, layout and motion timing, ready to rebuild as a Remotion project
Screen recording Software tutorials, web demos Which parts are real material you'll need to record yourself
Mixed / edits AI characters + code UI + real screen capture; live footage with AI edits Each shot judged separately, then reassembled on the original timeline

Besides how the visuals were made, every breakdown also covers:

Examples

The sample videos in docs/cases/videos/ are re-compressed to keep the repo small, which strips the TikTok / Douyin platform tags. So if you run them yourself, you'll get different results from the reports below, especially without a vision model. To reproduce them, use the original downloads from the platform.


Mixed-media clip · 19 s · 10 shots
AI characters talking + code-built UI + 8 s of real screen recording. The 5 dialogue shots are broken up by UI segments but share one set, so a single generation covers all of them.
Video · Report

3D explainer · 67 s · 21 shots
All 3D, AI narration. You can do it in 10 video generations, or have an LLM write Blender / Three.js code.
Video · Report

Product promo · 70 s · 12 shots
83% code-driven motion graphics, 12% screen recording. No video model needed; one Remotion project does it.
Video · Report

AI music video · 165 s · 86 shots
Correctly flags the song as AI-generated (it was made with Suno). The vision model called some shots "live-action drama"; the agent spotted fake Chinese characters on a scroll and ruled them AI.
Video · Report

Live-action unboxing · 72 s · 29 shots
All real footage: an overhead phone rig, one rug, one camera position, no voice-over, room sound plus library BGM. After platform transcoding, the offline tools called 35% of it AI; after agent review it came out 97% live-action.
Video · Report

Remakes

After rea-video breaks a video down, the remake recipe in capsule-cinema generates a new version that keeps the original's camera, motion, pacing and transitions and swaps only the elements you pick. These were made with a self-hosted MiniMax H3, using the original shot as the motion reference and keyframes with the swapped elements (gpt-image-2) for the look. Below are three remakes OiiOii showed in its WeChat articles. We made two versions of each: one with the same swap as OiiOii, and one that also swaps the subject, setting and props. Each GIF plays four versions in sync: the original, OiiOii's remake, our remake and the element-swap version.

Overhead outfit change · 12 s
Same swap: woman → man, three outfits → matching menswear. Element swap: a girl in three historical Hanfu outfits in a Chinese study; the remote becomes a round fan.

HD comparison video (with audio) · same-moment stills
Paper-craft / watercolor animation · 30 s
Same swap: new clay girl as the lead. Element swap: a fox; forest → coral reef, clouds → starry sky, sunset → aurora.

HD comparison video (with audio) · same-moment stills
Handheld product video · 21 s · 11 shots
Same swap: an original new product, English sticker → Chinese caption. Element swap: pink product, kitchen → camping table, new ingredients.

HD comparison video (with audio) · same-moment stills

The HD comparison videos use the original audio. The originals and OiiOii versions come from the articles and are shown for comparison only.

You can also do the remake in ChatCut, a hosted AI video editor: there is no model to set up, and the generated shots land on a timeline you can keep editing. Sponsor

Sponsors

rea-video is free and open source. Sponsors help keep it maintained.

ChatCut ChatCut is an AI video editor for Claude Code and Codex that generates the shots, cuts them on a timeline and exports the finished video.
Try ChatCut for free

Sponsored · Become a sponsor

Accuracy

Accuracy of the how-it-was-made call, weighted by shot duration:

Tuning set (15) Holdout (16) Re-encoded WeChat videos (10)
Tools only (offline, free) 98.3% 99.6% / 98.9% 88.6%
+ vision model (Gemini) — 99.6% / 99.8% 99.6%
+ agent visual review — 99.6% / 99.8% 100%

The Holdout column shows two numbers: holdout (9 videos) and holdout2 (7 videos).

On original files, the tools alone are usually enough. Videos downloaded from platforms have been recompressed, which wipes out the noise evidence, so you need a visual check to get back above 99%.

Real-asset detection is 97–99%; on nine hand-labelled clips, shot boundaries reach 95% recall and 94% precision. Numbers and how to reproduce them: eval/RESULTS.md.

How it works

rea-video pipeline

  1. Split: cut the video into shots and split the audio into voice, music and effects.
  2. Gather evidence: measure (frame cadence, sensor noise, whether the whole image moves as one flat plane, AIGC tags in the metadata), look (hands, text, physics) and listen (BPM, key, singing vs speech).
  3. Decide: hard evidence like AIGC tags and frame cadence beats "looks like". Conflicts get flagged for the agent to check visually.
  4. Plan the remake: batch shots according to what the target video model can actually do, such as max clip length, multi-shot support and how many reference images it takes.

Star History

Star History Chart

FAQ

How is this different from one-click remake tools like Hypit?

Hypit takes a reference video and makes a new one: the agent reads frame grids to understand the original, generates footage with the models it supports (Seedance, PixVerse, MiniMax, Wan and others), and renders captions and motion graphics in HTML.

It doesn't ask how the original was made. A live-action source goes through the same pipeline; it won't tell you how many cameras were used or how it was lit. What AI video it can match depends on the models it supports: no Kling or Veo, no dedicated avatar or 3D path, and no AI music (you bring your own). It's strongest on talking-head, podcast and interview videos that follow a script, and weaker on music-driven edits, product motion graphics and cinematic shorts.

rea-video doesn't render a video. It works out shot by shot whether something was filmed, AI-generated or built in code, then gives a shooting plan, prompts or a code approach for each. For a quick lookalike, Hypit is faster. To understand how the original was made and remake it cheaply, use rea-video. They work fine together.

Do I need a video-capable model? What does it cost?

No. Tools-only mode is offline and free, and already accurate on original files. For recompressed platform downloads, add a video-capable model (Gemini or Doubao both work) or use an agent that can see images; you pay whatever that service charges. Automatic video generation (capsule-cinema) does a free dry run first and only calls paid models after you confirm.

Can it remake the whole video automatically?

Partly. With capsule-cinema installed, AI shots are generated and placed back on the original timeline automatically. You film the live shots from the shooting plan, and the agent writes the code shots as a Remotion project. See docs/replicate.md.

Can it break down a multi-episode short drama?

rea-video works on one video at a time. For multi-episode short dramas with a running story, see rea-toon.

What is VIDEO.md, and why does rea-video write one?

Every skill has a SKILL.md; every video gets a VIDEO.md. It is an open format for video breakdowns, made so that models read only what a question needs. A full breakdown runs to hundreds of thousands of tokens; VIDEO.md lays it out in levels: a card of a few hundred tokens, an overview and map, a timeline with one line per shot, then one file per shot. Every file states its size, and what is seen on screen is kept apart from what the tools infer.

rea-video writes one into every output folder (out/ref/VIDEO.md), and its agent skill reads it instead of the raw data. Other agents and tools can read it without knowing rea-video: pip install videomd, then videomd check out/ref. The spec and guides are at videomd.github.io/videomd.

Can I just repost someone else's video?

No. rea-video is for learning a video's structure, pacing and techniques, then remaking it with material you have the rights to. Don't reuse other people's footage, music, characters or likeness. Follow platform rules and local AI-content labeling requirements when you publish.

Docs

The docs are in Chinese for now.

Community

Questions, breakdowns and remake ideas: GitHub Discussions. Chat with us on Discord. Chinese speakers can also find us on linux.do.

Want a video broken down? Open a request issue with the link and we'll publish the result as a case.

Credits

Inspired by REA (Reverse Engineer Anything): work back from what you see to how it was made, back every claim with evidence, and verify by measuring.

Sound separation uses UVR MDX-Net for voice and music, and BandIt v2 (Watcharasupat, Wu and Orife; weights CC BY-SA 4.0) for sound effects. Both models are downloaded on first use and are not part of this repo.

rea-video (JuneYaooo/rea-video)

正在搜索数据库…

10.3K ·GitHub
这是什么?它会把视频逐镜头拆解:画面内容、字幕及其样式、剪辑节奏和声音,并给出复刻方案。适用于实拍素材、AI 视频、3D、代码动画以及混合片段。
它会替我生成成品视频吗?不会。它会把每个镜头分类为实拍、AI 生成或代码制作,然后给出拍摄方案、提示词或代码思路,用于复刻该镜头。
我该怎么把它跑起来?不需要手动安装。你告诉 Claude Code 或 Codex 等编程代理来设置它,代理会安装命令行工具并运行完整的拆解流程。
开始之前我需要准备什么?电脑上需要安装 ffmpeg 和 Python 3.10 或更高版本。支持视频的视觉模型(Gemini、Doubao)是可选的,但能提升实拍与 AI 画面的识别准确度,并增加逐镜头提示词。
识别结果总是准确的吗?并非总是准确。从平台下载的视频经过压缩,会丢失噪点特征,因此需要视觉模型或代理自身的图像检查,才能把准确率提升到 99% 以上。真实的手机屏幕录像也可能被误判为 3D 或 AI。
不用视觉模型,也能得到有用的结果吗?可以。纯工具分析可离线运行且免费,README 中提到原始文件的准确率已经很高。只有压缩过的下载视频才需要视觉模型,并且会通过其服务商产生费用。
为什么不直接用 Hypit 重制视频?Hypit 基于参考视频生成新视频,对同风格片段来说更快。rea-video 则逐镜头解释原片是如何制作的,以及复刻需要哪些条件。README 称两者可以配合使用。
报告包含哪些内容?一个 report.html,包含每个镜头的截图、每项判定的依据、AI 镜头的提示词、拍摄与布光说明,以及字幕、切点时间和声音(如 BPM、音效位置)的细节分析。
我能用输出结果转发别人的视频吗?不能。README 称该工具用于学习视频的结构、节奏和创作手法,你应使用有权使用的素材重新制作,不要复制他人的素材、音乐、角色或形象。
它能自动完整复刻一个视频吗?只能部分实现。使用 cheap-claude-video 方案进行元素替换式复刻,会保留镜头运动、时间安排和转场,只替换你选定的元素。AI 镜头可以自动生成,但实拍镜头仍需要你自己拍摄。
作者
JuneYaooo
大小
153.6 MB
星标
10.3K
时长
20天
提交 · 90 天