The article examines diffusion models through the lens of Dense Associative Memories. It describes how these models create distinct attractors for each training sample in the small data regime. As the training set size increases, the models transition to generalization. A critical intermediate phase is identified where spurious states emerge, marking the first signs of generative capabilities. This phenomenon is demonstrated across a wide range of architectures and datasets.
Read the full story at nature.com.
Photo: geralt / Pixabay
