You can do this more controlled way of painting in diffusion models, turning it much more into a latent space brush, where you paint with sparse autoencoders features: https://fxtwitter.com/GoodfireAI/status/1927415017504165978?t=fvMPOGfH4yqLbPiq1wfuUQ&s=19 https://www.youtube.com/watch?v=mlZYRwJ2oJg It's a bit similar to this from long time ago (4 years ago) They are using sparse autoencoder features for this: https://fxtwitter.com/GoodfireAI/status/1927415034973491395