Diffusion Models (2020)

Ho et al.'s DDPM (2020) generates data by learning to reverse a noise diffusion process. Achieved image quality surpassing GANs with more stable training.