Comment by bertdb

Did you do any inpainting experiments? I can imagine a pixel-space diffusion model to be better at it than one with a latent auto-encoder.