Code2Video

Why we included this project

Code2Video takes a different route from the usual text-to-video pipeline. Rather than having a model paint frames directly, a planner, a coder, and a critic work together to write executable Manim code that is rendered into the finished lesson. That fits educational content well: topics like algorithms, math proofs, and puzzles need exact, deterministic visuals, and since every frame comes from code, mistakes are cheap to fix and clips are easy to re-edit. The repo also ships a benchmark of 117 curated learning topics in the style of 3Blue1Brown, giving people building agentic video pipelines real data and evaluation scripts to test against. One caveat: this is a research prototype, so you'll need to supply your own LLM, VLM, and asset API keys.