SDXL Activation Modulation Comparison

This app demonstrates activation modulation in Stable Diffusion XL using sparse autoencoders. It compares standard SDXL-Turbo outputs with modulated outputs that can steer the generation based on a separate concept.

Examples from the notebook:

• Main prompt: "A photo of a tree" with steering prompt: "tree with autumn leaves"
• Main prompt: "A dog" with steering prompt: "full shot"
• Main prompt: "A car" with steering prompt: "A blue car"

How to Use

1. Enter your main prompt (what you want to generate)
2. Enter a steering prompt (concept to influence the generation)
3. Adjust the modulation strength slider (λ) - higher values mean stronger influence
4. Click "Generate Comparison" to see the results side by side
5. Use advanced settings if needed to adjust seed, guidance scale, or steps

About

This app demonstrates activation modulation using a sparse autoencoder trained on SDXL text encoder layers. The modulation allows steering the generation toward specific concepts without changing the main prompt.