In a significant leap forward for generative artificial intelligence, researchers have introduced consistency models to enhance the speed and quality of image, audio, and video generation. These models aim to address the inefficiencies of diffusion models, which rely on iterative sampling processes that slow down generation. Consistency models can generate high-quality samples by directly mapping noise to data, facilitating rapid, one-step generation while retaining the option for multistep sampling to balance computation and sample quality.
One notable feature of consistency models is their ability to perform zero-shot data editing, such as image inpainting, colorization, and super-resolution, without requiring specialized training for these tasks. This versatility sets them apart in the field of generative AI. Consistency models can be trained either by distilling pre-trained diffusion models or as independent generative models.
The research team conducted extensive experiments demonstrating that consistency models surpass existing distillation techniques for diffusion models in both one- and few-step sampling. These models achieved a groundbreaking Fréchet Inception Distance (FID) of 3.55 on CIFAR-10 and 6.20 on ImageNet 64×64 for one-step generation, establishing new standards in the field.
When trained independently, consistency models mark the emergence of a new category of generative models. They outperform existing one-step, non-adversarial generative models on benchmarks such as CIFAR-10, ImageNet 64×64, and LSUN 256×256. This advancement underscores their potential to revolutionize generative AI.
The research was led by Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever, offering a promising direction for fast and high-quality data generation across various applications.
You can read the original article here: https://openai.com/index/consistency-models/
