It Takes Two to Tango: Two Parallel Samplers Improve Quality in Diffusion Models for Limited Steps

TL;DR

This paper introduces a simple, plug-&-play method using two parallel diffusion samplers to enhance image quality when limited denoising steps are available, without extra tuning.

Contribution

The authors propose a novel parallel sampling approach that improves diffusion model quality under step constraints, without additional training or external models.

Findings

Two parallel samplers outperform single samplers in limited-step scenarios.

Naive integration of sampler outputs reduces quality.

Adding more than two samplers does not necessarily improve results.

Abstract

We consider the situation where we have a limited number of denoising steps, i.e., of evaluations of a diffusion model. We show that two parallel processors or samplers under such limitation can improve the quality of the sampled image. Particularly, the two samplers make denoising steps at successive times, and their information is appropriately integrated in the latent image. Remarkably, our method is simple both conceptually and to implement: it is plug-&-play, model agnostic, and does not require any additional fine-tuning or external models. We test our method with both automated and human evaluations for different diffusion models. We also show that a naive integration of the information from the two samplers lowers sample quality. Finally, we find that adding more parallel samplers does not necessarily improve sample quality.