Code2Video
showlab/Code2VideoReviewed by AI
Checked by RepoPicks Radar on
What is Code2Video?
Code2Video is research code from Show Lab (NUS) that makes 3Blue1Brown-style teaching videos. You type a topic such as "Linear transformations and matrices" and three AI agents work together: one plans the storyboard, one writes Manim animation code and debugs it, and one looks at the result and fixes the layout. It comes with MMMC, a benchmark of 117 learning topics, and scripts to score the videos. You run it from a terminal with Python and Manim installed, and you pay for the AI services it calls.
How easy is it to install Code2Video?
D · terminal, several tools or a specific graphics card · Windows
- Windows D
- macOS D
- Linux D
The same on all three systems, Linux being the usual one for this kind of code: there is no download and no window. You clone the repository, install Python packages, set up Manim Community 0.19 with the FFmpeg and LaTeX pieces its own installation guide asks for, write your keys into api_config.json, then edit and run a shell script.
Coming soon: install in 1 click with the RepoPicks app.
Code2Video on GitHub
- Stars
- 2,355on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +114stars in 6 days
Stars over time
Facts
- Runs on
- Windows, macOS, Linux
- You need to install
- Python, FFmpeg, Other tools
- Disk space
- not determined
- Accounts or keys
- LLM API (Claude or another model, for the Planner and Coder), Gemini API (vision model, for the Critic that fixes layout), IconFinder
- Cost
- Free, API costs extra · The code is MIT. Each video is made by calling paid AI services; the README recommends Claude-4-Opus for the code and gemini-2.5-pro-preview for the layout checks, and names token usage as one of its efficiency metrics.
- License
- MIT (Permissive)
- Can replace
- Veo 3
- Last change on GitHub
- Aug 24, 2026
Code2Video risk signals (5)
- Attentioncost per video unknown
README, Configure LLM API Keys: the pipeline needs a paid text model ("Best Manim code quality achieved with Claude-4-Opus") plus a paid vision model for the Critic, and run_agent.sh has a PARALLEL_GROUP_NUM setting that runs several topics at once. The repository gives no price per video, only "Token usage" as a metric to measure.
AI review
- Infoyour topics go to outside ai services
README, Run Agents: every video is produced by calls to an LLM API and a VLM API configured in api_config.json, so the topic, the generated code and the rendered frames are sent to those companies' servers.
AI review
- Infoicons come from stock libraries
README, Acknowledgements: "High-quality visual assets (icons) are provided by IconFinder and Icons8", and an update note says the IconFinder account had problems so the icons were copied to a HuggingFace dataset "as a temporary alternative". Whatever those icons' licences say applies to any video you publish.
AI review
- Inforesearch code not a packaged program
The repository is the code for an ICML 2026 paper: no releases, no installer, no version numbers, and it is driven by editing variables inside run_agent.sh and run_agent_single.sh. The last commit was 78 days before this check.
AI review
- Infoquality compared only with its own benchmark
README, Evaluation: the quality claims rest on MMMC, a benchmark the same authors built, scored by their own scripts (eval_TQ.py, eval_AES.py), with 3Blue1Brown's lessons used as the reference for what a good video looks like.
AI review
Evidence is shown in its original language.
What Radar could not check about Code2Video
- how much the AI services charge for one finished video, and how often a run fails halfway and has to be paid for again
- whether the generated animation code runs without errors for topics outside the 117 in the benchmark
- how long one video takes on a normal computer, and how much disk the rendered frames and cached icons use
- whether Manim, LaTeX and FFmpeg install cleanly on Windows and macOS for this exact version
- what the licences of the downloaded icons allow if you publish or sell the videos