Steering SDXL Turbo Image Generation with Musical Motifs

Sauers_ · x · 2026-07-24

Leveraging Transformer block SAEs from the paper "Unboxing SDXL Turbo", a developer experimented with cross-modal image steering. By applying the autoencoder's decoder to activation space during every diffusion step, aspects of the music—specifically motifs from Aphex Twin's tracks—were used to dynamically control and manipulate the generated images.

Original post →

More from Multimodal

Multimodal channel →