Search papers, labs, and topics across Lattice.
This paper introduces Jeffrey guidance, a novel framework that enhances the control of diffusion models by explicitly targeting prescribed distributions during sampling. By leveraging Jeffrey's rule of conditioning, the method updates marginal distributions while preserving the conditional structure, leading to significant improvements in output quality. The authors demonstrate its effectiveness through applications that reduce FID scores on CIFAR-10 and FFHQ, and enforce fairness in attribute independence on CelebA-HQ.
Jeffrey guidance can significantly improve diffusion model outputs while ensuring fairness across attributes, challenging the limitations of standard sampling techniques.
A key strength of diffusion models lies in their flexibility, since their outputs can be controlled at sampling time through guidance. However, beyond simple cases such as conditional sampling, the target distribution is often left implicit, defined only through a sampling rule or a heuristic energy function. To address this, we propose Jeffrey guidance, a principled framework that extends diffusion-model control to applications beyond what standard guidance can express. It leverages Jeffrey's rule of conditioning to update marginal distributions towards a prescribed target, preserving the conditional structure and minimally perturbing the joint distribution. We first demonstrate Jeffrey guidance by targeting a prescribed embedding distribution. With Inception embeddings as the target, this leads to substantial reductions in FID on both CIFAR-10 and FFHQ. We further apply Jeffrey guidance to fairness on CelebA-HQ, updating an unconditional diffusion model to enforce independence between attributes.
