Consistency Models

Consistency models are generative models trained to map any point on a diffusion process's noise-to-data trajectory directly to that trajectory's endpoint, in a single evaluation. Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever introduced them in 2023 (arXiv:2303.01469). The model's core requirement, called consistency, is that mapping two different points on the same trajectory should produce the same output. This lets it collapse the many small denoising steps a diffusion model normally takes into one step, while still allowing extra steps to trade compute for quality.

A consistency model can be trained by distilling an already-trained diffusion model, using that model's own trajectories as a training signal. It can also be trained from scratch as a standalone generative model with no diffusion teacher. Trained by distillation, consistency models set a new state of the art for one-step sampling on ImageNet and CIFAR-10 at the time. Trained standalone, they outperformed prior one-step, non-adversarial generative models on the same benchmarks. They also carry over diffusion models' capacity for zero-shot editing tasks like inpainting and colorization without task-specific training.

Consistency models established few-step and one-step generation as a distinct research area alongside standard multi-step diffusion. Later work such as consistency trajectory models and latent consistency models applies the same idea inside latent diffusion pipelines for image and video generation. Recent methods avoid a distillation stage altogether, learning a one-step map directly from data, and position themselves explicitly against this original distillation-based formulation.