Classifier guidance steers diffusion sampling toward a desired class or condition with a classifier trained separately from the diffusion model. During reverse denoising, the classifier evaluates partially denoised images and supplies feedback about their alignment with the target. section 6.1.1
A guided sampling process proceeds in stages:
- Start from pure noise and iteratively reverse the learned noising process. section 6.1.1
- Present each intermediate, partially denoised sample to the separately trained classifier. section 6.1.1
- Use predicted class probabilities as feedback to steer later sampling toward the requested class or condition. section 6.1.1
Selecting an item class
An unguided model may generate any learned clothing class, such as a t-shirt, coat, or sandal. Classifier guidance biases the intermediate samples toward the selected item. section 6.1.1
Classifier guidance differs from classifier-free guidance: CFG trains the diffusion model to work with and without conditioning, embedding guidance in the generative model instead of requiring an external classifier. section 6.1.1 section 6.5
