marcelo-earth/generative-manim

Generative Manim – AI‑driven animation creation

What it is

• An open‑source suite that lets you describe a scene in plain English and have a Large Language Model (LLM) write the corresponding Manim Python code, then render it into a video.
• Built around the idea that the Manim renderer itself is a deterministic verifier: the generated code either produces a correct animation or crashes, which can be used as a reward signal for training.

Core components

Component	Purpose
LLM → Manim code engine	A collection of model wrappers (GPT‑4o, Claude, Gemini, open‑weight models via Featherless, LiteLLM passthrough, etc.) that turn natural‑language prompts into Manim scripts.
API	A REST‑style Animation Processing Interface that accepts a prompt, selects a model, runs the generation, and returns the rendered video (or the raw Python code).
Web demo	Hosted at generative-manim.vercel.app; shows the end‑to‑end flow in a browser.
Animo desktop app	A packaged, local‑rendering version of the same pipeline, giving users privacy and zero extra cloud cost.
Training pipeline	Open‑source scripts that fine‑tune open‑weight models (Qwen 2.5‑Coder 7B, DeepSeek‑Coder V2 Lite, CodeLlama 7B) using a three‑stage process: Supervised Fine‑Tuning → Direct Preference Optimization → Group Relative Policy Optimization, with the Manim renderer as the reward.
Benchmark suite	Reproducible evaluation (pass@k, render‑based scoring, structural checks) for any model that produces Manim code.

Key features

• Prompt‑to‑animation: Write a sentence like "A ball falls, bounces, and turns into a star" and get a ready‑to‑play video.
• Multiple model back‑ends: Official support for OpenAI, Anthropic, Google Gemini, Moonshot, plus open‑weight models accessed through Featherless or LiteLLM.
• Local rendering with Animo: No need to upload code or video; everything runs on the user’s machine.
• Exportable code: The generated Manim script is provided so you can edit, version‑control, or reuse it.
• Open‑source training: Scripts to train your own code‑generation model, with a benchmark that measures how often the model produces runnable Manim code.
• Community resources: Discord for discussion, a curated Manim Awesome Skills repo, and detailed deployment docs (Docker/Render).

Typical workflow

1. Write a description – plain English, no Manim knowledge required.
2. Select a model – choose a high‑capability model (e.g., GPT‑4o) or a cheaper open‑weight model.
3. Generate code – the API calls the LLM, receives a Python script, and optionally streams it back.
4. Render – the script is run through Manim; the resulting video is returned (or saved locally by Animo).
5. Iterate – edit the prompt or the generated code and repeat.

Getting started

• Try the demo: Visit https://generative-manim.vercel.app and type a prompt.
• Run the API locally:

  git clone https://github.com/360macky/generative-manim.git
  cd generative-manim/api
  # set your OpenAI/Anthropic/Google API keys
  export OPENAI_API_KEY=…
  uvicorn main:app --reload   # or use Docker as described in docs/cloud-deployment.md
  

• Use Animo: Download the desktop app from https://animo.video, install, and start creating without any server.
• Train your own model: Follow the training/ README – it shows how to fine‑tune Qwen 2.5‑Coder 7B with the provided 5k+ prompt‑code pairs and run the benchmark.

Who might benefit

• Educators & content creators who need quick explanatory animations without learning Manim.
• Developers looking for a code‑generation benchmark that has a deterministic, visual reward.
• Researchers interested in LLM‑driven programming, especially for domains where the output can be automatically verified (rendered video).
• Students who want to explore physics or math concepts via animated visualisations generated from natural language.

Community & support

• Discord: https://discord.gg/SNdbPU2AMM
• Issues & pull requests are welcomed on GitHub.
• Sponsorship: currently backed by The Astronomical Software Company.

---

All information above is taken directly from the repository’s README; no additional features have been inferred.