Classifier-Free Diffusion Guidance

Classifier guidance is a post-training method for conditional diffusion models that trades off mode coverage against sample fidelity, in the same spirit as low-temperature sampling or truncation in other generative models. It works by combining the diffusion model's score estimate with the gradient of a separately trained image classifier, which raises the question of whether guidance can be achieved without such a classifier at all.

The paper answers that question by introducing classifier-free guidance: a single pure generative model jointly trains a conditional and an unconditional diffusion model, and the resulting conditional and unconditional score estimates are combined at sampling time. This attains a trade-off between sample quality and diversity similar to classifier guidance, but without training or running any separate image classifier, simplifying guided diffusion generation.