Towards More General Control of Diffusion Models Using Jeffrey Guidance

The paper introduces Jeffrey guidance, a novel framework for enhancing control over diffusion models during sampling. This method utilizes Jeffrey's rule of conditioning to adjust marginal distributions toward specified targets while maintaining the conditional structure, resulting in improved performance, as evidenced by significant reductions in FID scores on CIFAR-10 and FFHQ datasets. This approach is particularly relevant for practitioners seeking to refine diffusion models for specific applications, such as embedding distributions and fairness in generated outputs.