RetiDiff: Stable Underwater Image Color Reconstruction Based on Retinex and Diffusion Distillation
Loading...
Date
2026
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Underwater image color reconstruction remains challenging due to wavelength-dependent light absorption and scattering that cause severe color casts and visibility degradation. We propose RetiDiff, a Retinex-guided diffusion distillation framework that couples a physics-aware diffusion prior with a lightweight Retinex-based UNet for stable, single-pass color restoration. A conditional Diffusion Transformer (DiT), pretrained on physics-guided underwater–terrestrial pairs, is frozen and distilled via Score Distillation Sampling (SDS) into a Retinex-UNet that predicts reflectance R and illumination L. This distillation transfers domain-agnostic color priors while mitigating cross-domain feature entanglement and avoiding iterative diffusion. To further suppress artifacts from imperfect Retinex separation, an Inter-Component Residual (ICR) regularization penalizes cross-component correlation and gradient co-occurrence, reducing halos, ghosting, and color drift while preserving structural fidelity. Extensive experiments on UIEB, LSUI, and TEST-U45 demonstrate state-of-the-art perceptual quality and LAB-space fidelity, with RetiDiff achieving comparable or superior performance to diffusion-based baselines while requiring far fewer parameters, lower FLOPs, and an order-of-magnitude faster inference.
Description
@inproceedings{10.2312:egs.20261008,
booktitle = {Eurographics 2026 - Short Papers},
editor = {Musialski, Przemyslaw and Lim, Isaak},
title = {{RetiDiff: Stable Underwater Image Color Reconstruction Based on Retinex and Diffusion Distillation}},
author = {Qiu, Wenyao and Zhou, Zhuang and Zhang, Xin and Chen, Jiayi and Zhou, Shiping and Tao, Ran},
year = {2026},
publisher = {The Eurographics Association},
ISSN = {2309-5059},
ISBN = {978-3-03868-299-8},
DOI = {10.2312/egs.20261008}
}