Exploring the Design Space of Diffusion Bridge Models

TL;DR

This paper broadens the design space of diffusion bridge models for image-to-image translation by introducing new techniques like preconditioning and optimized sampling, achieving state-of-the-art results and addressing diversity issues.

Contribution

It unifies and extends diffusion bridge models with novel methods, improving image quality, sampling efficiency, and diversity in I2I tasks.

Findings

Achieved state-of-the-art image quality in I2I translation

Enhanced sampling efficiency through optimized algorithms

Improved output diversity by modifying base distributions

Abstract

Diffusion bridge models and stochastic interpolants enable high-quality image-to-image (I2I) translation by creating paths between distributions in pixel space. However, the proliferation of techniques based on incompatible mathematical assumptions have impeded progress. In this work, we unify and expand the space of bridge models by extending Stochastic Interpolants (SIs) with preconditioning, endpoint conditioning, and an optimized sampling algorithm. These enhancements expand the design space of diffusion bridge models, leading to state-of-the-art performance in both image quality and sampling efficiency across diverse I2I tasks. Furthermore, we identify and address a previously overlooked issue of low sample diversity under fixed conditions. We introduce a quantitative analysis for output diversity and demonstrate how we can modify the base distribution for further improvements.