Apple researchers: one layer is enough to adapt visual encoders for image generation

Apple ML Research proposes a method that adapts pretrained visual encoders for image generation using only one additional trainable layer. The approach challenges the need for complex latent space compression in diffusion models.