marcelo-earth/generative-manim

What it solves

Generative Manim allows users to create mathematical and technical animations without needing to know how to program in Manim or use complex video editing software. It bridges the gap between natural language descriptions and the programmatic animation engine Manim by using Large Language Models (LLMs) to translate text prompts into executable animation code.

How it works

The project provides a suite of tools, including a web app and an API, that send user text descriptions to an LLM (such as GPT-4o, Claude, or Gemini). The LLM generates the corresponding Manim Python code, which is then rendered into a video file. The project also includes a training pipeline to create specialized open-source models (like Qwen 2.5 Coder) using a three-stage process: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO), using the Manim renderer itself as a deterministic reward signal to verify if the code renders successfully.

Who it’s for

• Non-programmers who want to create high-quality animations for educational or technical content.
• Developers who seeking to build applications on top of the Animation Processing Interface (API).
• AI Researchers interested in fine-tuning LLMs for specialized code generation tasks with deterministic verification.

Highlights

• Multi-Model Support: Supports a wide array of frontier models from OpenAI, Anthropic, Google, and Moonshot AI, as well as open-weight models via Featherless.
• Deterministic Verification: Uses the Manim renderer as a reward signal for RL training, ensuring generated code actually works.
• Integrated Benchmark: Includes an executable benchmark MVP for evaluating the expert-level Manim code generation capabilities of different models.
• Desktop Evolution: Evolved into Animo, a desktop application for local rendering and increased privacy.