Abstract
Diffusion models are powerful generative models but suffer from slowsampling, often taking 1000 sequential denoising steps for one sample. As aresult, considerable efforts have been directed toward reducing the number ofdenoising steps, but these methods hurt sample quality. Instead of reducing thenumber of denoising steps (trading quality for speed), in this paper we explorean orthogonal approach: can we run the denoising steps in parallel (tradingcompute for speed)? In spite of the sequential nature of the denoising steps,we show that surprisingly it is possible to parallelize sampling via Picarditerations, by guessing the solution of future denoising steps and iterativelyrefining until convergence. With this insight, we present ParaDiGMS, a novelmethod to accelerate the sampling of pretrained diffusion models by denoisingmultiple steps in parallel. ParaDiGMS is the first diffusion sampling methodthat enables trading compute for speed and is even compatible with existingfast sampling techniques such as DDIM and DPMSolver. Using ParaDiGMS, weimprove sampling speed by 2-4x across a range of robotics and image generationmodels, giving state-of-the-art sampling speeds of 0.2s on 100-stepDiffusionPolicy and 16s on 1000-step StableDiffusion-v2 with no measurabledegradation of task reward, FID score, or CLIP score.