Code2Video is a framework for generating high-quality educational videos using executable Manim code, addressing the limitations of pixel-based text-to-video models in clarity and reproducibility. It targets researchers and educators seeking automated, structured, and verifiable video content creation, offering benefits like enhanced coherence and systematic evaluation.

How It Works

The project employs a code-centric paradigm where Manim code serves as the unified medium for temporal sequencing and spatial organization. Its core is a modular tri-agent design: a Planner expands storyboards, a Coder synthesizes and debugs Manim code, and a Critic refines layouts using visual anchors. This approach ensures clarity, coherence, and reproducibility, differentiating it from traditional video generation methods.

Quick Start & Requirements

Installation involves navigating to src/ and running pip install -r requirements.txt. Key prerequisites include Manim Community v0.19.0, LLM API keys (Claude-4-Opus for core generation, Gemini for layout, IconFinder for assets), and configuring credentials in gpt_config.json. Videos can be generated from a single query using run_agent_single.sh or in full benchmark mode via run_agent.sh. Official documentation and the MMMC benchmark are available via links.

Highlighted Details

• Code-Centric Paradigm: Leverages executable Manim code for unified control over video elements.
• Modular Tri-Agent Design: Planner, Coder, and Critic collaborate for structured and refined video synthesis.
• **MMMC Benchmark