Scalable Fast Sampling Methods For Diffusion Models

In recent years, diffusion models have demonstrated impressive capabilities in multimodal generation tasks, including image, video and speech synthesis. However, their sampling speed remains a significant bottleneck for real-time generation due to the nature of iterative denoising. In this talk, I will discuss some recent advanced distillation methods for diffusion models, including variational score distillation and continuous-time consistency models, with their theoretical connections with traditional MCMC methods and the practical techniques for stably scaling up the methods on the high-resolutional dataset ImageNet 512x512.