Venue
International Conference on Machine Learning
Diffusion models have significantly advanced the fields of image, audio, and video generation, but they depend on an iterative sampling process that causes slow generation. To overcome this limitation, we propose consistency models, a new family of models that generate high quality samples by directly mapping noise to data. They support fast one-step generation by design, while still allowing multistep sampling to trade compute for sample quality. They also support zero-shot data editing, such as image inpainting, colorization, and super-resolution, without requiring explicit training on these tasks. Consistency models can be trained either by distilling pre-trained diffusion models, or as standalone generative models altogether. Through extensive experiments, we demonstrate that they outperform existing distillation techniques for diffusion models in one-and few-step sampling, achieving the new state-ofthe-art FID of 3.55 on CIFAR-10 and 6.20 on ImageNet 64ˆ64 for one-step generation. When trained in isolation, consistency models become a new family of generative models that can outperform existing one-step, non-adversarial generative models on standard benchmarks such as CIFAR-10, ImageNet 64ˆ64 and LSUN 256ˆ256. transformers can make one strong gan, and that can scale up.
This paper introduces consistency models, a new class of models designed to improve the efficiency and quality of generative tasks in image, audio, and video generation, specifically addressing the slow iterative sampling process of diffusion models. Consistency models enable fast one-step generation while retaining the option for multistep sampling to enhance quality. They also allow for zero-shot data editing tasks such as image inpainting, colorization, and super-resolution, requiring no specific training for these tasks. The models can be trained by either distilling from pre-trained diffusion models or as standalone generative models. The paper demonstrates their superiority over existing distillation techniques, achieving state-of-the-art results on metrics such as Fréchet Inception Distance (FID) on datasets like CIFAR-10 and ImageNet 64x64. The authors show that these models can also outperform various other single-step generative models, leading to improved sample quality across multiple datasets while efficiently performing zero-shot editing tasks.
This paper employs the following methods:
- Consistency Models
- Distillation
- Multistep Sampling
The following datasets were used in this research:
- CIFAR-10
- ImageNet 64ˆ64
- LSUN Bedroom 256ˆ256
- LSUN Cat 256ˆ256
- FID
- Inception Score (IS)
- Precision (Prec.)
- Recall (Rec.)
- New state-of-the-art FID of 3.55 on CIFAR-10 for one-step generation
- New state-of-the-art FID of 6.20 on ImageNet 64x64 for one-step generation
- Superiority to existing distillation techniques in image generation
- Number of GPUs: None specified
- GPU Type: None specified