Diffusion Models Exhibit Remarkable Consistency Across Initializations and Architectures

Diffusion models with different initializations or architectures can produce remarkably similar outputs when given the same noise inputs, a rare property in other generative models.