OpenAI’s Consistency Models Support Fast One-Step Generation for Diffusion Models

Diffusion-based AI models continue to wow the world with remarkable image, audio, and video generation capabilities. This performance however comes at a price, as diffusion models’ iterative sampling process that progressively removes noise to generate high-quality outputs typically requires 10–2000 times more compute than conventional single-step generative models such as generative adversarial networks (GANs). As is the case in other AI domains, generative AI researchers are increasingly focused not only on improving model performance, but also on reducing their compute burdens.

In the new paper Consistency Models, an OpenAI research team introduces a family of compute-efficient generative models that achieve state-of-the-art performance on single-step sample generation without adversarial training. The proposed consistency models can be trained either to distill pretrained diffusion models or as standalone generative models.

The team’s goal was to design a new family of generative models that facilitate efficient, single-step generation yet preserve the advantages of diffusion models’ iterative refinement process — such as the ability to trade-off computation complexity and sample quality when necessary or perform zero-shot data editing tasks.

Consistency models are built upon the probability flow (PF) ordinary differential equation (ODE) found in continuous-time diffusion models. Given a PF ODE that smoothly converts data to noise, a consistency model learns to map any point at any time step to the trajectory’s starting point for generative modelling. The outputs are thus trained to be “consistent” with initial points on the same trajectory — a key factor that informs the name given to the model family.

The researchers introduce two consistency model training approaches, which support either distillation or isolation modes. The first training approach uses numerical ODE solvers and a pretrained diffusion model to obtain pairs of adjacent points on a PF ODE trajectory and efficiently distill a diffusion model into a consistency model. This significantly improves sample quality while enabling zero-shot image editing. The second approach trains without dependence on pretrained diffusion models, effectively establishing consistency models as an independent generative model family.

In their empirical study, the team applied consistency models to real image datasets CIFAR-10, ImageNet 64×64, LSUN Bedroom 256×256, and LSUN Cat 256×256. In the experiments, distillation via consistency models achieved new state-of-the-art FID scores of 3.55 on CIFAR-10 and 6.20 on ImageNet 64×64 for one-step generation, while standalone consistency models also outperformed existing single-step, non-adversarial generative models.

This paper demonstrates how the proposed consistency models can achieve much more efficient sampling while performing single-step generation. The team believes such models also can offer exciting prospects for the cross-pollination of ideas and methods among different AI research fields.

The paper Consistency Models is on arXiv.

---

Author: Hecate He | Editor: Michael Sarazen

---

We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.