Consistency Models

ABSTRACT

Diffusion models have achieved significant success in image, audio, and video generation, but they depend on an iterative generation process that causes slow sampling and precludes real-time applications. To overcome this, we propose consistency models, a new family of generative models that support fast one-step or few-step generation by design. They learn to map noise to data directly by enforcing a self-consistency property over the probability flow ordinary differential equation (ODE) trajectory. Consistency models can be trained either by distilling pre-trained diffusion models or as standalone generative models from scratch, achieving state-of-the-art results for one-step generation on CIFAR-10 and ImageNet 64x64.