EvenFormer: Dynamic Even Transformer for Real-World Image Restoration

Xin Lu, Yuanfei Bao, Jiarong Yang, Anya Hu, Jie Xiao, Kunyu Wang, Dong Li, Senyan Xu, Kean Liu, Xueyang Fu, Zheng-Jun Zha; Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 1081-1091

Ambient Lighting Normalization aims to restore clear image information in complex real-world scenarios. To address the intricate image restoration challenges in real-world high-resolution images, this paper proposes a novel two-stage network Dynamic Even Transformer for Real-World Image Restoration (EvenFormer). The first-stage network employs a Transformer architecture to model long sequential information. To tackle the unevenness and randomness in image degradation, we utilize a pixel-wise Gaussian shuffling method to aid in global interaction modeling, effectively restoring authentic background information under complex degradation conditions. In the second stage, we introduce the NAFNet network based on a CNN architecture to further refine large-scale images while eliminating blocky interference artifacts caused by the window-based modeling in the Transformer. Experimental results on official datasets validate the superiority of EvenFormer compared to existing approaches.