saev - Sparse Auto-Encoders for Vision

Training sparse autoencoders (SAEs) on vision transformers (ViTs), implemented in PyTorch.

Docs

• Docs
• Colab Notebook for SAE Inference
• User guide
• API reference

Research

This framework was developed for a series of projects leveraging SAEs with vision models.

• [Feb 2025] Interpretable and Testable Vision Features via Sparse Autoencoders
• [Nov 2025] Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoders

Trained SAE checkpoints are available at:

• Huggingface Models

Citation

If you want to cite the software, please use the "Cite this repository" link on the GitHub page, which will provide you with the appropriate citation format.