Abstract
Classifier-free guidance has become a staple for conditional generation withdenoising diffusion models. However, a comprehensive understanding ofclassifier-free guidance is still missing. In this work, we carry out anempirical study to provide a fresh perspective on classifier-free guidance.Concretely, instead of solely focusing on classifier-free guidance, we traceback to the root, i.e., classifier guidance, pinpoint the key assumption forthe derivation, and conduct a systematic study to understand the role of theclassifier. We find that both classifier guidance and classifier-free guidanceachieve conditional generation by pushing the denoising diffusion trajectoriesaway from decision boundaries, i.e., areas where conditional information isusually entangled and is hard to learn. Based on this classifier-centricunderstanding, we propose a generic postprocessing step built uponflow-matching to shrink the gap between the learned distribution for apre-trained denoising diffusion model and the real data distribution, majorlyaround the decision boundaries. Experiments on various datasets verify theeffectiveness of the proposed approach.