rea-video: Open Source AI Video Reverse Engineering Tool, Analyzing Each Shot to Reconstruct the Video's Production Recipe

1. What is rea-video

rea-video is an open-source video reverse engineering tool specifically designed for AI Agents. It can analyze any finished video shot by shot, determining whether the visuals are real-life footage, AI-generated, 3D rendered, code-based animations, or screen recordings. Based on this analysis, it breaks down the production details such as camera setup and lighting, subtitle styles, editing rhythm, and audio composition. The tool integrates with AI Agents like Claude Code and Codex via a command-line interface and Skill, generating HTML analysis reports for each shot and executable replication plans. It supports reverse-engineering the prompts used for each AI-generated shot, and when combined with tools like cheap-claude-video or Hypit, can replace the original characters in the video to produce new content. To run locally, you need to install ffmpeg and Python 3.10+. Accuracy improves further when integrated with visual models such as Gemini or Douyin.

Technical Positioning and Domain: rea-video is a reverse engineering tool in the field of AI video analysis, focusing on the automatic identification and deconstruction of video production methods. Its technical direction spans computer vision, audio-video signal processing, and multimodal content understanding. In the video understanding space, it targets the niche of "reconstructing production methods," differentiating itself from conventional video content understanding and video summarization tools.

Development Background: The project was open-sourced on GitHub (https://github.com/JuneYaooo/rea-video) and is maintained by independent developer JuneYaooo. The motivation behind its development stems from the information asymmetry in the AI video creation domain—creators often face a large volume of high-quality videos but struggle to determine their production methods, making it difficult to quickly replicate similar effects. The project adopts an Agent-native design philosophy, encapsulating the complex video analysis process into a CLI tool and Agent Skill, thereby reducing the technical barriers to entry.

Core Value: The core problem that this tool solves is "how was this video made?" It uses evidence-driven multimodal analysis to break down the video into quantifiable production parameters—from camera placement and lighting, to shot movement, subtitle styles, and audio composition—and outputs directly executable replication plans. According to official self-testing data, the accuracy of determining the original video type is approximately 99%, which is a significant advantage compared to similar tools.

Technical Features: rea-video employs a decision-making mechanism that prioritizes hard evidence, using quantifiable indicators such as frame rhythm, sensor noise, and AIGC metadata as the basis for its judgments. When hard evidence conflicts with model intuition, it automatically flags the questionable items for review by the Agent, keeping the misjudgment rate low. Additionally, its built-in capability table for video, image, and audio models allows it to plan replication strategies based on the actual capabilities of the target model, deciding whether to generate a segment using a video model or have the LLM write Blender/Three.js/Remotion code to rebuild it.

2. Key Features

• Video Type Identification: Determines the type of each shot—whether it is live-action, AI-generated, 3D-rendered, code animation, or screen recording. For mixed-edited videos, the system automatically splits the video by shot boundaries and independently analyzes each segment, preserving the original timeline structure when reassembling.

• Production Method Breakdown: Outputs camera setup and lighting plans, camera movement trajectories, and color correction parameters for each shot. For AI-generated shots, it can also reverse-engineer the prompts and types of reference materials used. This feature enables creators to systematically understand the methodology behind video production.

• Subtitle and Decorative Element Extraction: Accurately identifies the start and end times, color, outline style, and entrance animation for each subtitle line, supporting export as standard SRT/ASS subtitle files. It also locates the position and duration of stickers, emojis, and picture-in-picture elements within the video frame.

• Editing and Audio Analysis: Marks all cut points, analyzes the correlation between editing rhythm and music beat, distinguishes between human voice and AI-generated voiceovers, identifies AI-generated or library music, extracts BPM values, and annotates the timing of sound effects.

• Replication Plan Generation: Organizes each shot into a generation plan based on the capabilities of the target video model, clearly distinguishing between parts requiring live-action, generation, or coding. The final output is an HTML report containing shot-by-shot screenshots and audio breakdowns.

• Multimodal Evidence Collection: Quantitatively measures frame rate, sensor noise, and whether the entire image moves as a single plane. It also checks for AIGC tags in metadata; visually, the model examines hand structure, text rendering, and physical plausibility; and audibly, it analyzes BPM and tonality to distinguish between singing and speaking.

• Agent Native Integration: Available as a CLI with Skill integration for agents like Claude Code and Codex. A single instruction can automatically complete environment installation, dependency configuration, and full analysis workflow. It can operate in basic mode without connecting to external visual models.

• Model Capability Planning: Includes a built-in table of capabilities for mainstream video, image, and audio models. It batches shots into a generation plan that minimizes the number of required generations based on the target model's generation characteristics, avoiding unnecessary generation attempts.

3. Application Scenarios

• E-commerce Live Streaming Replication: Analyze the camera setup, scene breakdown, and subtitle pacing of trending live streaming videos. After generating a complete list of production parameters, replace the products with your own to mass-produce live streaming videos in the same style.

• MCN Competitive Analysis: Analyze each shot of competitor videos to determine their composition ratios—real footage percentage, AI-generated segments, and code animations—to accurately estimate production costs and workflow complexity, providing data support for content strategy decisions.

• AI Short Film Creation Efficiency: Reverse-engineer reference AI short films to identify the prompts and generation parameters used for each shot. This eliminates the trial-and-error cost of figuring out prompts, allowing for direct secondary creation based on existing results.

• Hybrid Content Dissection: Identify the production method for each segment of a hybrid video that combines real people, AI characters, and screen recordings. Reconstruct the video according to the original timeline to create your own version, solving the challenge of uniformly processing hybrid content.

• Educational/Animation Replication: Reconstruct 3D educational or code animation videos using the logic of tools like Blender or Remotion. Replicate similar effects at a lower cost through script-based methods, ideal for quickly producing tutorial-style content.

• Video Content Audit: Conduct a technical audit of an organization's or individual's video assets to clearly identify the production methods, source materials, and tools used for each video segment, providing a basis for copyright management and content compliance.

4. How to Use

1. Environment Setup: Install ffmpeg on your local computer (used for video decoding, shot splitting, and audio track separation) and Python 3.10 or higher version (used to run the CLI tool). The operating system is not limited—Windows, macOS, and Linux are all supported.

2. Select Agent Environment: Use it in Claude Code, Codex, or any AI Agent capable of executing commands. The design goal of rea-video is to be natively called by an Agent, so the Agent must have the ability to execute terminal commands and read files.

3. Install the Project: Send the instruction “Install https://github.com/JuneYaooo/rea-video” directly to the Agent. The Agent will automatically clone the repository, install Python dependencies, and configure environment variables. Once the installation is complete, you can call rea-video-related commands.

4. Start Video Analysis: Send the instruction “Analyze this video /path/to/video.mp4” to the Agent. The Agent will run the full analysis process—shot splitting, audio track separation, multimodal evidence collection, type determination, and report generation.

5. Integrate with Vision Model (Optional): Provide the interface address and API Key for Gemini or Doudian in the Agent's configuration file. After integration, the system can automatically reverse-engineer prompts for each shot, significantly improving accuracy and reducing questionable items. If no vision model is integrated, the Agent will directly review screenshots to verify questionable shots.

6. View Analysis Results: After the analysis is complete, the user will receive three parts of output—a one-sentence summary, a replication plan categorized into “real shot / generated / coded,” and a report.html file containing individual shot screenshots and detailed audio track analysis. The report can be opened directly in a browser.

7. Proceed with Replication: Pass the generated HTML report to video replication tools such as cheap-claude-video or Hypit. Replace elements according to the recipe to generate a new version. At this point, a complete workflow loop from analysis to replication is formed.

5. Pros and Cons Analysis

6. Comparative Analysis with Similar Tools

From their positioning differences, rea-video and Hypit naturally form an upstream-downstream collaboration relationship — the former is responsible for analyzing "how the video was made," while the latter executes "making it according to the recipe." Together, they can achieve a complete workflow from deconstruction to replication. Choosing rea-video is suitable for user scenarios requiring in-depth research into video production methods; choosing Hypit is more appropriate for commercial scenarios where a clear recipe exists and multiple video variations need to be generated quickly. For users without a technical background who focus more on creative generation, Runway ML and Descript offer more user-friendly graphical interfaces, but at the cost of lacking the detailed breakdown capabilities provided by rea-video.

7. Editor's Summary

rea-video occupies a unique ecological position in the AI video toolchain — video reverse engineering. While many tools are focused on solving "how to generate better videos," it instead answers the question "how was this video generated." The innovation of this technical direction is reflected in its evidence-driven detection architecture: it uses hard metrics such as frame rhythm quantification, sensor noise analysis, and AIGC metadata inspection as the basis for judgment, rather than relying solely on model intuition. This multimodal forensic approach is a first in video analysis tools. The official self-testing accuracy rate of 99% for original videos demonstrates the effectiveness of its technical approach, especially in segment-by-segment judgment of mixed editing content, showcasing its ability to handle complex inputs.

In terms of practical value, rea-video advances video analysis from "understanding content" to "understanding production." The output — a lighting setup list, reverse-engineered prompts, and editing rhythm data — forms a directly executable replication plan. This product concept of "analysis as a recipe" significantly shortens the path from video inspiration to final output, offering direct application value for scenarios such as e-commerce live streaming, MCN competitor research, and AI short film creation. Its agent-native tool form also aligns with the current trend in AI toolchains, eliminating the friction of traditional software installation and configuration.

In terms of target users, the tool is aimed at three typical user groups: MCN organizations and content researchers who need systematic breakdowns of competitor video production methods, independent creators who want to accelerate AI short film creation by reverse-engineering prompts, and e-commerce operations teams that require batch generation of video variants. For general video editing users, its command-line interface and agent dependency present a certain usage barrier, requiring basic command-line operation skills.

As an open-source project maintained by an individual developer, rea-video currently has a small community, but its unique project positioning provides it with clear growth potential. As the quality of AI-generated video continues to improve and creators increasingly demand transparency in production methods, video reverse analysis will gradually become a standard part of the content creation workflow. It is recommended to follow the project's future version updates, especially optimizations in visual model integration and expansion of the community ecosystem.

8. FAQ

Q: Does rea-video require a GPU to run?
A: No. rea-video's core analysis logic is primarily based on ffmpeg video processing and Python computation, and does not involve local large model inference. It can run smoothly in a CPU environment. The visual judgment component is completed by calling cloud-based models such as Gemini or Doubao via API, so no high-performance GPU is needed locally.

Q: What is the difference between connecting to a visual model and not connecting?
A: When connected to visual models like Gemini or Doubao, the system can automatically generate reverse-prompting instructions for each AI-generated shot, significantly improving the accuracy of shot-by-shot judgment and greatly reducing questionable items. Without connection, the analysis process can still run, but verification of questionable shots requires the Agent to manually review screenshots, resulting in lower accuracy and automation levels.

Q: What is the relationship between rea-video and Hypit?
A: The two are in an upstream-downstream collaborative relationship. rea-video is responsible for video reverse analysis, outputting an HTML report that includes camera setup, reverse-prompting instructions, and replication plans. Hypit is responsible for executing the video replication, reading the recipe generated by rea-video, and replacing elements such as products or people to generate new video versions. Using both together enables a complete workflow from analysis to replication.

Q: What video lengths can rea-video handle? How long does the analysis take?
A: rea-video has no strict limit on video length, but processing time is directly related to the video duration, number of shots, and whether a visual model is connected. The more shots there are and the longer the segments requiring visual verification, the longer the total processing time. In practice, analysis of short videos lasting a few minutes typically completes within a few to十几 minutes.

Q: Can the reverse-prompting instructions generated by rea-video be used directly with other video generation models?
A: Yes, but they need to be adapted based on the target model's capabilities. rea-video includes a capability table for mainstream video, image, and audio models, and will batch process shots into a plan that minimizes the number of generations required based on the target model's actual capabilities. Users can also use the reverse-prompting instructions as a reference, manually adjust them, and then use them with any video generation model.

Q: Can the analysis results be used for commercial purposes?
A: Yes. rea-video is released under an open-source license, allowing free use and modification of the code. However, users must assess the copyright compliance of the analyzed videos themselves. Analyzing others' videos for learning and research is typically acceptable, but when directly replicating creative elements from commercial videos for use in your own commercial content, you should comprehensively consider related legal risks such as trademarks, copyrights, and commercial competition.

9. Project Links

• GitHub Repository: https://github.com/JuneYaooo/rea-video (Official open-source code repository, containing complete CLI tool source code, usage documentation, and Agent Skill configuration guide)