Diffusion probabilistic models have recently demonstrated state-of-the-art gener- ative performance in a wide range of applications, including inpainting, super- resolution, and semantic editing. While these models have proven highly effective in generative tasks, there is a growing interest in leveraging their generative ap- proach for discriminative tasks, which has shown promising results. This paper focuses on the ability of large-scale text-to-image diffusion models to serve as a zero-shot classifier, which is a type of discriminative task that involves classifying images based on textual descriptions, without any prior training on the specific image classes. To evaluate the performance of these models, we compare them to the current state-of-the-art zero-shot classifier, CLIP. To the best of our knowl- edge, this is the first study to investigate the potential of large-scale text-to-image diffusion models for zero-shot classification.
Find more in my paper.
Each experiment contains python notebook, which one can easily launch in Google Collab. VISITING NOISE-DENOISE SECTION IS MANDATORY!!!
Go to the experiment of interest and follow the readme instructions there.