Abstract

Recently, Rissanen et al., (2022) have presented a new type of diffusionprocess for generative modeling based on heat dissipation, or blurring, as analternative to isotropic Gaussian diffusion. Here, we show that blurring canequivalently be defined through a Gaussian diffusion process with non-isotropicnoise. In making this connection, we bridge the gap between inverse heatdissipation and denoising diffusion, and we shed light on the inductive biasthat results from this modeling choice. Finally, we propose a generalized classof diffusion models that offers the best of both standard Gaussian denoisingdiffusion and inverse heat dissipation, which we call Blurring DiffusionModels.

 

Quick Read (beta)