In a recent development, OpenAI has introduced a new family of generative models called consistency models. These models aim to improve the speed and efficiency of generating high-quality samples in image, audio, and video domains, which have traditionally relied on diffusion models. The key advantage of consistency models is their ability to generate samples in a single step by directly mapping noise to data, while still supporting multi-step sampling to balance computational effort with sample quality.
Consistency models also offer impressive zero-shot data editing capabilities, enabling tasks such as image inpainting, colorization, and super-resolution without the need for explicit training on these specific tasks. This flexibility stands in contrast to traditional methods, which often require extensive task-specific training to achieve similar results.
There are two primary ways to train consistency models: by distilling existing pre-trained diffusion models or as standalone generative models. Empirical testing shows that these models outperform existing distillation techniques in both one- and few-step sampling scenarios, achieving state-of-the-art Fréchet Inception Distance (FID) scores. Specifically, consistency models reached an FID of 3.55 on the CIFAR-10 dataset and 6.20 on ImageNet 64×64 for one-step generation, setting new benchmarks in the field.
When trained independently, consistency models have demonstrated their superiority over existing single-step, non-adversarial generative models across standard benchmarks such as CIFAR-10, ImageNet 64×64, and LSUN 256×256. These findings indicate that consistency models not only provide speed advantages but also excel in generating high-quality data.
The research was conducted by Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Their groundbreaking work promises to set new standards in the field of generative modeling and holds the potential to significantly accelerate advancements in AI-driven data generation.
You can read the original article here: https://openai.com/index/consistency-models/
