Consistency Models

TL;DR

Consistency models are a new type of generative model that enable fast, high-quality data generation and editing by directly mapping noise to data, outperforming previous diffusion-based methods in speed and quality.

Contribution

This paper introduces consistency models, a novel family of models that support one-step generation and zero-shot editing, trained via distillation or as standalone models, surpassing existing methods in benchmarks.

Findings

Achieve state-of-the-art FID of 3.55 on CIFAR-10 with one-step generation.

Outperform existing diffusion distillation techniques in one- and few-step sampling.

Can serve as standalone generative models surpassing traditional non-adversarial models.

Abstract

Diffusion models have significantly advanced the fields of image, audio, and video generation, but they depend on an iterative sampling process that causes slow generation. To overcome this limitation, we propose consistency models, a new family of models that generate high quality samples by directly mapping noise to data. They support fast one-step generation by design, while still allowing multistep sampling to trade compute for sample quality. They also support zero-shot data editing, such as image inpainting, colorization, and super-resolution, without requiring explicit training on these tasks. Consistency models can be trained either by distilling pre-trained diffusion models, or as standalone generative models altogether. Through extensive experiments, we demonstrate that they outperform existing distillation techniques for diffusion models in one- and few-step sampling, achieving the new state-of-the-art FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64x64 for one-step generation. When trained in isolation, consistency models become a new family of generative models that can outperform existing one-step, non-adversarial generative models on standard benchmarks such as CIFAR-10, ImageNet 64x64 and LSUN 256x256.