BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference
Core Concepts
Proposing BayesDiff for estimating pixel-wise uncertainty in diffusion models using Bayesian inference.
Abstract
BayesDiff introduces a novel approach to estimate pixel-wise uncertainty in diffusion models. It leverages Bayesian inference to characterize uncertainty dynamics and improve image quality. The method shows promise for diverse applications, including text-to-image tasks.
BayesDiff
Stats
Diffusion models have impressive image generation capability.
Lack of proper sample-wise metric for identifying low-quality generations.
Integration of Bayesian uncertainty and diffusion models is challenging.
BayesDiff proposes a framework for estimating pixel-wise Bayesian uncertainty.
Extensive experiments demonstrate the efficacy of BayesDiff.
Quotes
"Bayesian posterior delivers low uncertainty for training data and high uncertainty for others."
"BayesDiff enables simultaneous delivery of image samples and pixel-wise uncertainty estimates."