Classifier-free guidance in diffusion models

WHAT THIS TESTS: Whether you understand the mechanism, implementation, and trade-offs of the dominant technique for strengthening conditional diffusion generation.

A GOOD ANSWER COVERS: Classifier-free guidance removes the external classifier that earlier classifier guidance required. During training, the conditioning input, such as the text embedding or class label, is randomly replaced with a null or empty token some fraction of the time, often around ten to twenty percent. The same network thus learns to denoise both with the condition and without it, giving an unconditional model for free. At inference you run the network twice per step, once conditioned on the prompt and once unconditioned, then form a guided prediction by taking the unconditional estimate and extrapolating toward the conditional one, scaled by a guidance weight. A weight of one recovers ordinary conditional sampling; larger weights push harder toward the prompt, increasing fidelity and prompt adherence at the cost of diversity and sometimes saturation or artifacts. It outperforms classifier guidance because it avoids training a separate classifier on noisy images, sidesteps that classifier's noisy and sometimes adversarial gradients, and keeps everything in one consistent model.

COMMON WRONG ANSWERS: Confusing it with classifier guidance that uses gradients of an external classifier; saying it requires two separately trained models; claiming it improves diversity rather than trading diversity for fidelity; forgetting the conditioning dropout during training; thinking the guidance scale has no downside.

LIKELY FOLLOW-UPS: What does cranking the guidance scale too high do to image quality? How does it interact with negative prompts? What is the extra inference cost? Why are noisy-image classifiers hard to train, motivating the classifier-free approach?

ONE CONCRETE EXAMPLE: A text-to-image model is trained with the prompt embedding dropped to an empty string ten percent of the time. At generation, for the prompt an astronaut riding a horse, each step computes the conditioned noise prediction and the unconditioned one, then the sampler moves further in the conditioned direction with a guidance scale of about seven. The astronaut and horse appear crisply and on-prompt; lowering the scale yields more varied but looser interpretations, while pushing it very high oversaturates colors.