Real-World Adaptation of Retinexformer for Low-Light Image Enhancement Using Unpaired Data

Abstract

Low-light image enhancement remains a significant challenge in real-world computer vision applications, especially where lighting conditions vary drastically and paired training data is unavailable. While transformer-based models have shown promise in controlled environments, their ef-fectiveness often diminishes when applied to naturally degraded images. This paper presents a novel approach for adapting a transformer-based enhancement model to realworld low-light scenarios using unpaired datasets. We utilize real low-light images captured under uncontrolled conditions and propose a domain adaptation framework that enables effective transfer learning from synthetic to real domains. Our method integrates unsupervised reconstruction loss, perceptual optimization, and domain-invariant feature alignment to refine the model’s performance without requiring paired supervision. Experimental evaluations reveal notable improvements in both visual quality and quantitative metrics on real-world benchmarks. Compared to existing enhancement methods, our approach offers superior generalization, robustness to noise, and high-fidelity out-put. This demonstrates the potential of our domain-adapted transformer model in practical low-light imaging applications, including night photography, surveillance, and mobile vision systems.