# Parallel Sampling of Diffusion Models

**Authors:** Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, Nima Anari  
**arXiv:** [2305.16317](https://arxiv.org/abs/2305.16317)  
**Date:** 2023-05-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16317v1/

## Key points

- Authors address the challenge of slow sampling in diffusion models
- Previous efforts focused on reducing denoising steps at the expense of sample quality
- Paper introduces ParaDiGMS method using Picard iterations for parallel denoising steps
- ParaDiGMS accelerates sampling process without compromising sample quality
- Compatible with existing fast sampling methods such as DDIM and DPMSolver
- Results show 2-4x improvement in sampling speed compared to traditional methods
- Achieves state-of-the-art speeds on different diffusion model scenarios
- Speed gains achieved without degradation in task reward, FID score, or CLIP score metrics

## Summary

In their paper "Parallel Sampling of Diffusion Models," authors Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, and Nima Anari address the challenge of slow sampling in diffusion models. These generative models typically require 1000 sequential denoising steps for one sample, leading to significant computational overhead. Previous efforts have focused on reducing the number of denoising steps at the expense of sample quality; however, this paper takes a different approach by exploring parallelization of denoising steps. The authors propose a novel method called ParaDiGMS that uses Picard iterations to run denoising steps in parallel. By predicting future denoising solutions and iteratively refining them until convergence, ParaDiGMS accelerates the sampling process without compromising sample quality. This approach represents a significant advancement in diffusion sampling techniques as it allows for trading compute resources for speed. Notably, ParaDiGMS is compatible with existing fast sampling methods such as DDIM and DPMSolver, further enhancing its applicability across various robotics and image generation models. The results demonstrate a 2-4x improvement in sampling speed compared to traditional methods and achieve state-of-the-art speeds of 0.2s on 100-step DiffusionPolicy and 16s on 1000-step StableDiffusion-v2. Importantly, these speed gains are achieved without any measurable degradation in task reward, FID score or CLIP score metrics. Overall,"Parallel Sampling of Diffusion Models" introduces an innovative approach to accelerating diffusion model sampling through parallel denoising steps and offers a promising solution to the computational challenges associated with these powerful generative models.