classifier-free

Definition

Naming a technique that gets the effect of classifier guidance without training a separate classifier. In classifier-free guidance the same diffusion model is run twice, once conditioned on the prompt and once unconditionally, and the difference between the two predictions is extrapolated to push samples toward the condition. The name is historical: earlier methods needed gradients from an external classifier trained on noisy images, which was fragile and expensive.

Explainers using this term

• CFG and Samplers — What the "Strength" Knob in Generative AI Really DoesClassifier-Free Diffusion Guidance