Robust low-light image enhancement in the wild via data synthesis and generative diffusion prior

Wang, Zhihua, Lin, Qinghua, Liu, Feiyang, Zhang, Weixia and Zhou, Wei 2026. Robust low-light image enhancement in the wild via data synthesis and generative diffusion prior. Pattern Recognition , 113336. 10.1016/j.patcog.2026.113336

Full text not available from this repository.

Abstract

Deep neural networks (DNNs) have become the mainstream approach for low-light image enhancement (LLIE), achieving impressive progress in recent years. However, their real-world applicability remains limited by two major challenges. First, supervised LLIE depends on large-scale paired datasets of low-light and reference images, which are costly to collect and sensitive to reference fidelity. Second, real-world degradations—such as underexposure, uneven illumination, noise, color imbalance, and motion blur—are diverse and non-linear, making them difficult to model and hindering generalization. To overcome these issues, we introduce a framework that combines advanced data synthesis with generative diffusion priors to enhance robustness and adaptability. For data synthesis, we develop an ISP-based pipeline enhanced with JPEG compression and partial degradations, producing images that contain both well-exposed and severely under-/overexposed regions. This physically grounded simulation enables scalable, realistic paired training data generation. For model design, we propose the Low-light Residual Diffusion (LoRDiff) model, which exploits the generative capacity of pre-trained text-to-image diffusion models for LLIE. Leveraging Low-Rank Adapters (LoRA), LoRDiff selectively fine-tunes only the most relevant layers, adapting the rich structural, textural, and semantic priors learned from large-scale natural image datasets to the LLIE task.This design allows the model to capture complex degradations and produce visually pleasing, perceptually consistent enhancements. Extensive experiments across multiple LLIE benchmarks demonstrate that LoRDiff achieves competitive or superior performance in both quantitative and perceptual evaluations. The model also exhibits strong generalization and stability under challenging real-world conditions, highlighting its promise for practical deployment. The source code and pretrained weights are publicly available at: https://github.com/wzhsysu/LoRDiff.

Item Type:	Article
Date Type:	Published Online
Status:	In Press
Schools:	Schools > Computer Science & Informatics
Additional Information:	License information from Publisher: LICENSE 1: Title: This article is under embargo with an end date yet to be finalised.
Publisher:	Elsevier
ISSN:	0031-3203
Last Modified:	10 Mar 2026 10:16
URI:	https://orca.cardiff.ac.uk/id/eprint/185639

Actions (repository staff only)

Edit Item