<ul data-eligibleForWebStory="true"><li>Diffusion models have been instrumental in generative modeling advancements.</li><li>Even though current models generate high-quality average samples, individual samples may be of low quality.</li><li>Detecting low-quality samples without human intervention is still a complex task.</li><li>A Bayesian framework is proposed to estimate generative uncertainty of synthetic samples.</li><li>The framework aims to make Bayesian inference practical for large generative models.</li><li>A new semantic likelihood is introduced to handle challenges in high-dimensional sample spaces.</li><li>Experiments demonstrate that the proposed generative uncertainty method effectively identifies poor samples.</li><li>The Bayesian framework can be applied post-hoc to pretrained diffusion models using the Laplace approximation.</li><li>Simple yet effective techniques are suggested to reduce computational overhead during sampling.</li></ul>

Generative Uncertainty in Diffusion Models

Discover more