Accelerating Convergence of Score-Based Diffusion Models, Provably

arXiv:2403.03852v1 Announce Type: new
Abstract: Score-based diffusion models, while achieving remarkable empirical performance, often suffer from low sampling speed, due to extensive function evaluations needed during the sampling phase. Despite a flurry of recent activities towards speeding up diffusion generative modeling in practice, theoretical underpinnings for acceleration techniques remain severely limited. In this paper, we design novel training-free algorithms to accelerate popular deterministic (i.e., DDIM) and stochastic (i.e., DDPM) samplers. Our accelerated deterministic sampler converges at a rate $O(1/{T}^2)$ with …

abstract arxiv convergence cs.ai cs.it cs.lg diffusion diffusion models function generative generative modeling low math.it math.oc modeling performance practice sampling speed stat.ml type