Elucidating the design space of deep stochastic processes for image enhancement

ICLR 2026 Conference Submission20710 Authors

Keywords: image enhancement, diffusion models, Ornstein-Uhlenbeck processes, diffusion bridges, stochastic differential equations

TL;DR: This paper unifies diverse image enhancement methods under a stochastic process framework, implements and benchmarks 11 approaches across four tasks, and identifies key factors behind their performance differences.

Abstract:

In this work, we investigate deep stochastic processes for image enhancement. We show that existing approaches can be interpreted as instances of Ornstein–Uhlenbeck processes, diffusion bridges, or diffusion processes, each represented by a stochastic differential equation. As a result, we consolidate 11 methods into a unified mathematical framework and present them in a systematically structured table. This perspective separates the definition of processes from the schedulers and samplers that were originally used. Furthermore, we provide a modular library that implements the proposed methods and facilitates the integration of additional approaches with minimal coding effort.
In order to perform comprehensive empirical evaluation among considered approaches, we evaluate them on four image enhancement tasks: super-resolution, colorization, low-light enhancement, and deraining with identical backbones and training protocol ensuring fair and meaningful comparison. The experiments highlight that, while most methods achieve similar results, there are exceptions that make some refinement strategies more effective than others, which we further analyze and explain.

Primary Area: generative models

Submission Number: 20710