Summary: We propose REPresentation Alignment (REPA), a method that aligns noisy input states in diffusion models with representations from pretrained visual encoders. This significantly improves ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results