We're open-sourcing an SAE trained on the image generation model Stable Diffusion XL-Turbo. The SAE is trained on block down.2.1 of SDXL-Turbo's UNet.
For an example of loading the model and exploring the top features for the model's activations, see inference_example.ipynb. This notebook downloads and loads the SAE model from Hugging Face and explores the top features in our sample activations.