GFRRN: Explore the Gaps in Single Image Reflection Removal

Yu Chen, Zewei He†, Xingyu Liu, Zixuan Chen, Zheming Lu
Zhejiang University, Huanjiang Lab, The Chinese University of Hong Kong
CVPR 2026
†Indicates Corresponding Author
arXiv Code Dataset
Teaser image showing our method's results

(a) Single-stream method; (b) Dual-stream method; (c) Dual-stream method with feature interaction mechanism; (d) Ours; (e) Experimental results with different settings: left side shows PSNR comparison, right side shows SSIM comparison.

Abstract

Prior dual-stream methods with the feature interaction mechanism have achieved remarkable performance in single image reflection removal (SIRR). However, they often struggle with (1) semantic understanding gap between the features of pre-trained models and those of image restoration models, and (2) reflection label inconsistencies between synthetic and real-world training data. In this paper, we first adopt the parameter efficient fine-tuning strategy by integrating several learnable Mona layers into the pre-trained model to align the training directions. Then, a label generator is designed to unify the labels for both synthetic and real-world data with an optimized reflection label. In addition, a Gaussian-based Adaptive Frequency Learning Block (G-AFLB) is proposed to adaptively learn and fuse the frequency priors, and a dynamic agent attention (DAA) is employed as an alternative to window-based attention by dynamically modeling the significance levels across windows (inter-window) and within individual windows (intra-window). The aforementioned improvements collectively constitute our proposed Gap-Free Reflection Removal Network (GFRRN). Extensive experiments demonstrate the effectiveness of our GFRRN, achieving superior performance against state-of-the-art SIRR methods.

Visual Comparison

Drag the slider to compare the input image with our result.

Processed Result 1
Input Image 1
Input Output
Processed Result 4
Input Image 4
Input Output
Processed Result 2
Input Image 2
Input Output
Processed Result 3
Input Image 3
Input Output

Visual Effects on Open Datasets

The inputs are above, and the outputs are below. These images are from the datasets 'Nature', 'Real' and 'SIR2'.

BibTeX

@article{chen2026gfrrn,
      title={GFRRN: Explore the Gaps in Single Image Reflection Removal}, 
      author={Yu Chen and Zewei He and Xingyu Liu and Zixuan Chen and Zheming Lu},
      year={2026},
      eprint={2602.22695},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.22695}, 
}