Image models usually VAE decode first, then super-res later, which is slow and mostly reconstructs rather than creates detail. This NVIDIA research makes the decoder itself a pixel diffusion model, so a 512^2

→ View original post on X — @askalphaxiv