Machine learningOther

Classifier-Free Guidance is a Predictor-Corrector

The paper explores the theoretical aspects of classifier-free guidance (CFG), debunking misconceptions about its interaction with DDPM and DDIM, and explaining its function as a predictor-corrector method, offering a deeper understanding of CFG within the context of principled sampling methods.