Simplifying, stabilizing, and scaling continuous-time consistency models

OpenAI has published research on continuous-time consistency models, a class of generative models that can produce high-quality images in as few as two sampling steps. According to the OpenAI Blog, the team simplified and stabilized the training process for these models and successfully scaled them to achieve sample quality on par with leading diffusion models. Consistency models were introduced as an alternative to diffusion models, which typically require many iterative denoising steps to generate an image. The new work addresses earlier challenges in training consistency models in continuous time, making them more practical. For developers building AI workflows, this advancement means faster inference without sacrificing output quality. Two-step generation could enable real-time image creation in interactive applications, reduce cloud compute costs, and open up possibilities for on-device deployment. The research also suggests that further scaling could narrow the remaining gap with diffusion models.