marcelo-earth/generative-manim

Leverages LLMs (GPT-4o, Claude, open-source models) with custom system prompts to convert natural language descriptions into executable Manim animation code, eliminating the need for manual scripting expertise. Employs a three-stage training pipeline (SFT + DPO + GRPO) using Manim's deterministic renderer as a reward signal to fine-tune open-source code models like Qwen and DeepSeek. Includes a render-based benchmark suite with pass@k evaluation and comparative analysis tools for measuring code generation quality across model variants.

805 stars. Actively maintained with 5 commits in the last 30 days.