• present Diffusion Classifier: strong classification results with pretrained conditional diffusion models, with no additional training! (View Highlight)
  • Diffusion Classifier approximates the learned p(img | class) with the ELBO. This consists of adding noise to the image and using the diffusion model conditioned on each class to predict the noise. We find p(class | img) with Bayes’ rule to determine the most likely class. (View Highlight)