classifier guidance

A method for conditioning a diffusion modeldiffusion model, n.: An AI that learns to create images, sound or video by practicing noise removal: shown pictures buried under static, it learns to clean them up, then builds new ones starting from pure static. by adding, at each sampling step, the gradient of a separate classifier's log-probability for the desired class, computed on the noisy intermediate sample and multiplied by a guidance scale. By Bayes' rule this gradient converts the unconditional score into that of the class-conditional distribution; scales above one sharpen the distribution, improving fidelity and class adherence at the cost of diversity. The classifier must be trained on noised images at every noise level, an extra burden, and its gradients can push toward adversarial patterns rather than meaningful features. A variant, CLIP guidance, uses the gradient of image–text similarity from a CLIPCLIP, prop. n.: An OpenAI model that learned to match pictures with their captions from hundreds of millions of web examples, so it can judge how well any piece of text describes any image. model to steer generation toward a caption. Classifier guidance has been largely superseded by classifier-free guidanceclassifier-free guidance, n.: A setting that makes an image generator follow the prompt more strictly by comparing what it would draw with and without the prompt, then exaggerating the difference., which needs no extra model.