Comment by dangond

I believe diffusion image models learn to model a reverse-noising function, rather than reverse-blurring.