Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models

While there are many score-based models with various diffusing strategies as well as many numerical schemes of the denoising process, only a few works have explored the score part of the generative SDE. This paper introduces a new generative SDE with score adjustment using an auxiliary discriminator. The goal is to improve the original generative process of a pre-trained diffusion model by estimating the gap between the pre-trained score estimation and the true data score. This is done by training a discriminator that classifies diffused real data and diffused sample data. The gap estimation is then used to adjust the pre-trained score network. In experiments, the method enables new SOTA FIDs of 1.77/1.64 on unconditional/conditional CIFAR-10, and new SOTA FID/sFID of 3.18/4.53 on ImageNet 256x256.