We examine the theoretical foundations of classifier-free steerage (CFG). CFG is the dominant technique of conditional sampling for text-to-image diffusion fashions, but in contrast to different elements of diffusion, it stays on shaky theoretical footing. On this paper, we disprove widespread misconceptions, by exhibiting that CFG interacts in another way with DDPM (Ho et al., 2020) and DDIM (Music et al., 2021), and neither sampler with CFG generates the gamma-powered distribution p(x|c)^γp(x)^{1−γ}. Then, we make clear the conduct of CFG by exhibiting that it’s a form of predictor-corrector technique (Music et al., 2020) that alternates between denoising and sharpening, which we name predictor-corrector steerage (PCG). We show that within the SDE restrict, CFG is definitely equal to combining a DDIM predictor for the conditional distribution along with a Langevin dynamics corrector for a gamma-powered distribution (with a rigorously chosen gamma). Our work thus gives a lens to theoretically perceive CFG by embedding it in a broader design area of principled sampling strategies.
