showlab/Code2Video

What it solves

Code2Video addresses the difficulty of creating high-quality, coherent, and reproducible educational videos. Unlike traditional text-to-video models that generate pixels directly, this framework ensures clarity and precision by using executable code as the medium for video generation.

How it works

The system uses an agentic, code-centric paradigm that leverages the Manim library to generate videos. It employs a modular tri-agent design:

• Planner: Expands knowledge points into a detailed storyboard.
• Coder: Synthesizes debuggable Manim code to implement the storyboard.
• Critic: Refines the layout and aesthetics using anchors and visual feedback.

To enhance visual quality, the framework can integrate external icons via the IconFinder API. The process is evaluated using the MMMC benchmark, which tests knowledge transfer, aesthetic quality, and efficiency.

Who it’s for

This tool is designed for educators, content creators, and researchers who want to automate the production of precise, mathematically or technically accurate educational animations similar to those produced by 3Blue1Brown.

Highlights

• Code-Driven Generation: Uses executable Manim code instead of pixel-based generation for better coherence and reproducibility.
• Tri-Agent Architecture: Separates planning, coding, and critiquing to ensure structured and refined output.
• MMMC Benchmark: Includes a specialized benchmark of 117 learning topics to evaluate educational video generation.
• Multi-Dimensional Eval: Assesses videos based on knowledge transfer (TeachQuiz), aesthetics, and computational efficiency.