Diffusion Model

• Iterative alpha-(de)Blending

• Inversion by Direct Iteration

  In image restoration, direct one-step restoration with model trained on L_p norm cause blurrly results (i.e. the regression to the mean effect) because the space of all valid restoration is large.

  Instead multiple small-step restoration largely avoid the regression-to-the-mean effect because the set of 'slightly-less-bad' images is relatively small for each small step. (link)