One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models

Stable Diffustion XL multistep version

Note

If you encounter GPU time limit errors, don't worry, the app still works and you can use it freely.

Demo Overview

This demo showcases the use of Sparse Autoencoders (SAEs) to understand the features learned by the Stable Diffusion XL (Turbo) model.

How to Use

Explore

  • Enter a prompt in the text box and click on the "Generate" button to generate an image.
  • You can observe the active features in different blocks plot on top of the generated image.

Top Images

  • For each feature, you can view the top images that activate the feature the most.

Paint!

  • Generate an image using the prompt.
  • Paint on the generated image to apply interventions.
  • Use the "Feature Icon" button to understand how the selected brush functions.

Remarks

  • Not all brushes mix well with all images. Experiment with different brushes and strengths.
  • Feature Icon works best with down.2.1 (composition) and up.0.1 (style) blocks.
  • This demo is provided for research purposes only. We do not take responsibility for the content generated by the demo.

Interesting features to try

To get started, try the following features:

  • down.2.1 (composition): 2301 (evil) 3747 (image frame) 4998 (cartoon)
  • up.0.1 (style): 4977 (tiger stripes) 90 (fur) 2615 (twilight blur)