AI - Manim Video Generator

Automatic video generation using an agentic LLM flow in combination with the manim python library.

🌟 Inspiration

Following the release of Deepseek-R1, users found it was particularly adept at generating manim code for mathematical animations. A good example of this ability can be seen in this X post and in the repository related to that post: Math-To-Manim

Goal

The Goal of this project is to experiment with the automatic creation of manim based videos. Based on any input data / information you want to visualize.

As of now the project consist of an (agentic) flow with one Code Reviewer and one Code Writer that interact in a feedback loop.

Manim code generated in a one shot fashion often contains errors or has visual inconsistency. The approach in this project tries to reduce the error rate and improve the quality of the videos through better prompting, feedback loops and the use VLMs (Image Capable LLMs).

The use of LiteLLM in the project ensures that future models can also be easily benchmarked and compared on this task.

Benchmarking future model releases on this task will also be interesting, since it requires good spatial understanding which most LLMs struggle with.

❗ It's important to note that creating a coherent video that compiles and has no issues is still pushing the current SOTA LLMs to their limits

Current Flow (subject to change)

🚀 Installation

1. Clone the repository:

git clone https://github.com/makefinks/manim-generator.git
cd manim-generator

2. Install the requirements

With uv (Recommended)

uv sync

Note: uv offers faster installation times and better dependency resolution

Or using the requirements.txt directly

pip install -r requirements.txt

You also need to install ffmpeg and for some cases a LaTeX distribution

Windows (using Chocolatey):

choco install ffmpeg
choco install miktex

macOS (using Homebrew):

brew install ffmpeg
brew install --cask mactex

Linux (Debian/Ubuntu):

sudo apt-get update
sudo apt-get install texlive texlive-latex-extra texlive-fonts-extra texlive-science

3. Setup Environment Variables

Either by creating a .env file or by defining the API key of your preferred provider system wide.

Recommendation: use openrouter to access all llms

⚡ Usage

1. Input any data you want to have visualised as a video into video_data.txt

2. Execute the script

python manim_generator.py

3. Tweak the config

Manually configure what llm the coder and reviewer should use inside manim_generator.py.

Images will only be used if the reviewer model supports image inputs.

Check compatibility:

• https://openrouter.ai/models?modality=text+image-%3Etext
• https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json (can be tweaked manually if necessary)

🤝 Contributing

Contributions are welcome!

Areas for Contribution

• Improving the prompt engineering
• Enhancing the review feedback loop -> extending it with more agentic patterns
• Improving code quality
• Adding / Proposing new features or optimizations

Known Issues:

• Streaming: current implementation performs no syntax highlighting and leaves artifacts in the console history when scrolling
• Prompting / Environment Setup: There may be a mismatch of what version the llm chooses to use and what is currently installed on the system.

Observations:

LLMs

• Claude 3.7 Sonnet creates unique and stable scenes, allocating extra thinking budget increases the complexity of the visualisations but introduces visual errors or invalid use of the manim api
• o3-mini seems to be the most reliable coder but does not create the most impressive scenes - more testing is neeeded (also no image inputs yet)
• Claude 3.5 Sonnet V2 performs very well as the reviewer and coder - to no ones suprise ;)
• Deepseek R1 sometimes creates more impressive compositions but code often has syntax errors
• Gemini 2.0 Flash struggles creating runnable scripts without additional prompting

Flow

• In the original flow the llm was instructed to only output one class (Manim Scene) so that it was easier to render the video with one command. Now each scene is its own class. This was done so I could extract the --last-frame for each scene and pass it to the review model. There might be a better way to do this!

🎬 Examples

You can find example videos in the examples directory.