Repolode
Submit
showlab

showlab/Code2Video

Generates educational videos from knowledge points by having LLM agents write executable Manim code instead of producing pixels directly.

View on GitHubWebsite
Stars
2,141
Forks
294
Open issues
2
Last commit
2 months ago

What is it?

Code2Video is a framework for producing educational videos from a topic or knowledge point. Rather than using a pixel-based text-to-video model, it has agents write executable Manim code, so the resulting animations are structured, reproducible and easy to edit at the code level.

The pipeline uses three agents: a Planner that expands a topic into a storyboard, a Coder that synthesizes debuggable Manim code, and a Critic that refines layout using anchor points. Generation can run for a single query or across a full list of learning topics defined in a JSON file. Optional icon assets can be pulled from the IconFinder API.

The project also includes the MMMC benchmark, which covers 117 learning topics inspired by 3Blue1Brown, and evaluation scripts for knowledge transfer, aesthetic quality and efficiency metrics. Setup requires Python, Manim Community v0.19.0, and API keys for an LLM and a vision-language model.

Key features

  • 01Generates educational videos by producing executable Manim code.
  • 02Uses a three-agent pipeline of Planner, Coder and Critic.
  • 03Supports single-topic generation and batch runs over a topic list.
  • 04Can enrich videos with icons from the IconFinder API.
  • 05Includes the MMMC benchmark with 117 learning topics and evaluation scripts.

Use cases

  • Creating explainer videos for math or computer science concepts.
  • Researching code-driven video generation and agent pipelines.
  • Benchmarking educational video quality and knowledge transfer.

Topics

#video-generation#manim#multi-agent#llm#education#code-generation#coding

FAQ

What is Code2Video?
Generates educational videos from knowledge points by having LLM agents write executable Manim code instead of producing pixels directly. Code2Video is a framework for producing educational videos from a topic or knowledge point. Rather than using a pixel-based text-to-video model, it has agents write executable Manim code, so the resulting animations are structured, reproducible and easy to edit at the code level.
Is Code2Video free to use? What license does it have?
Code2Video is open source under the MIT license. Check the license terms before using it in commercial products.
Is Code2Video still maintained?
Updated within 6 months. The last commit to Code2Video was on Aug 24, 2026.
What can I use Code2Video for?
Creating explainer videos for math or computer science concepts. Researching code-driven video generation and agent pipelines. Benchmarking educational video quality and knowledge transfer.
Where can I get Code2Video?
The source code and installation instructions are on GitHub: https://github.com/showlab/Code2Video

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects