Code2Video

Code2Video is a research project from Show Lab at the National University of Singapore that generates educational videos with a code-centric approach. Instead of asking a text-to-video model for pixels, agents plan the lesson and write Manim animation code, and the README compares results with Veo3 and Wan2.2 on topics such as the Hanoi problem, large language models and Fourier series. It links a paper, a dataset (MMMC) and a project website, and its news section notes it passed 1,000 stars and that installation time was cut by 80 to 90% after a requirements cleanup.

Running it takes real setup. You install `src/requirements.txt` and Manim Community v0.19.0, then fill in `api_config.json` with a model for the Planner and Coder (the README says Claude-4-Opus gave the best Manim code), a vision-language model for the Planner Critic (Gemini 2.5 Pro for layout and aesthetics) and an IconFinder key for icons. `sh run_agent_single.sh --knowledge_point "Linear transformations and matrices"` makes one video, and `sh run_agent.sh` runs the benchmark topic list. No license is declared in the repository.