Diffusion Model with Perceptual Loss

arXiv:2401.00110v4 Announce Type: replace-cross
Abstract: Diffusion models trained with mean squared error loss tend to generate unrealistic samples. Current state-of-the-art models rely on classifier-free guidance to improve sample quality, yet its surprising effectiveness is not fully understood. In this paper, we show that the effectiveness of classifier-free guidance partly originates from it being a form of implicit perceptual guidance. As a result, we can directly incorporate perceptual loss in diffusion training to improve sample quality. Since the score matching objective …

abstract art arxiv classifier cs.ai cs.cv cs.lg current diffusion diffusion model diffusion models error form free generate guidance loss mean paper quality sample samples show state state-of-the-art models type