It uses a conditional diffusion process to predict the image residual, generating high-fidelity $SR$ results that avoid the over-smoothing common in $MSE$-based models.

It maps a simple distribution to a complex one using a sequence of invertible transformations $f$, enabling exact likelihood $p(x)$ estimation and efficient sampling.

It scales vision learning by masking a large portion of an image ($75\%$) and forcing an asymmetric $ViT$ to reconstruct the original pixels, learning robust representations.