Diffusion Classifiers Understand Compositionality, but Conditions Apply

Recent advancements in diffusion models are reshaping our understanding of visual scenes, a key aspect of human intelligence. While traditional discriminative models have made strides in computer vision, they often fall short in grasping compositionality. However, generative text-to-image diffusion models have shown remarkable capabilities in synthesizing complex scenes, indicating a potential for deeper compositional understanding. This development is significant as it opens new avenues for applying zero-shot diffusion classifiers, enhancing the versatility and effectiveness of these models in various applications.