Source-linked AI summary
NTIRE 2026 Challenge on Single Image Reflection Removal in the Wild: Datasets, Results, and Methods
Jie Cai, Kangning Yang, Zhiyuan Li, Florin-Alexandru Vasluianu, Radu Timofte, Jinlong Li, Jinglin Shen, Zibo Meng, Junyan Cao, Lu Zhao, Pengwei Liu, Yuyi Zhang, Fengjun Guo, Jiagao Hu, Zepeng Wang, Fei Wang, Daiguo Zhou, Yi'ang Chen, Honghui Zhu, Mengru Yang, Yan Luo, Kui Jiang, Jin Guo, Jonghyuk Park, Jae-Young Sim, Wei Zhou, Hongyu Huang, Linfeng Li, Lindong Kong, Saiprasad Meesiyawar, Misbha Falak Khanpagadi, Nikhil Akalwadi, Ramesh Ashok Tabib, Uma Mudenagudi, Bilel Benjdira, Anas M. Ali, Wadii Boulila, Kosuke Shigematsu, Hiroto Shirono, Asuka Shin, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Jiachen Tu, Shreeniketh Joshi, Jin-Hui Jiang, Yu-Fan Lin, Yu-Jou Hsiao, Chia-Ming Lee, Fu-En Yang, Yu-Chiang Frank Wang, Chih-Chung Hsu
TL;DR
Real-world SIRR research is constrained by scarce, precisely aligned data and limited evaluation diversity. The paper reviews a challenge built around OpenRR-5k, a real-world benchmark and evaluation protocol, finding strong participation and competitive top-ranked methods while exposing persistent perceptual-quality challenges.
Problem
High-quality aligned real-world data are scarce, limiting evaluation of reflection removal beyond synthetic or restricted scenarios.
Method
The paper reviews an NTIRE challenge using OpenRR-5k and evaluates submitted methods with objective metrics and expert subjective assessment.
Results
Top methods achieved strong objective and subjective performance, with RRay reaching subjective score 4.45 and PSNR 36.1688.
Takeaways & Limitations
The results indicate that real-world reflection removal still requires perceptual-quality consistency across diverse scenes.
Abstract
from arXiv · showhide
In this paper, we review the NTIRE 2026 challenge on single-image reflection removal (SIRR) in the wild. SIRR is a fundamental task in image restoration. Despite progress in academic research, most methods are tested on synthetic images or limited real-world images, creating a gap in real-world applications. In this challenge, we provide participants with the OpenRR-5k dataset. This dataset requires participants to process real-world images covering a range of reflection scenarios and intensities, aiming to generate clean images without reflections. The challenge attracted more than 100 registrations, with eleven of them participating in the final testing phase. The top-ranked methods advanced the state-of-the-art reflection removal performance and earned unanimous recognition from five experts in the field. The proposed OpenRR-5k dataset is available at https://huggingface.co/datasets/qiuzhangTiTi/OpenRR-5k, and the homepage of this challenge is at https://github.com/caijie0620/OpenRR-5k.
1. Introduction
SIRR seeks to recover the transmission layer from reflection-contaminated images, but real-world generalization remains limited by handcrafted priors and scarce aligned data. The challenge addresses this gap by providing real-world evaluation toward practical photography.
- SIRR recovers the transmission layer T from input images I contaminated by reflections R from reflective surfaces.
- Traditional methods use handcrafted priors, while deep learning models the uncertainty of transmission estimation for broader scenarios.
- High-quality real-world data are scarce because precise pixel-level alignment is labor-intensive and technically difficult.
- The challenge evaluates models in real-world scenarios to narrow the gap between academic research and practical photography.
2. NTIRE 2026 SIRR Challenge
The challenge introduces a real-world, diverse, pixel-aligned benchmark and evaluates reflection-removal systems through objective metrics and expert perceptual assessment. Its protocol combines AI-assisted paired-data generation, blind testing, and subjective-score-based ranking.
- 2.1. Overview: The challenge targets diverse real-world SIRR scenarios and aims to bridge laboratory research with consumer-grade imaging solutions.
- 2.2. OpenRR-5k Dataset: Environmental wind and vibration can cause misalignment and color shifts when constructing paired reflection-removal data with physical setups.
- 2.2. OpenRR-5k Dataset: The data protocol uses an OPPO AI reflection-removal editor to generate initial transmission images, replacing manual glass removal or black-cloth blocking.
- 2.2. OpenRR-5k Dataset: The dataset contains 5,000 training pairs, 300 validation pairs, and 100 input-only test images for blind evaluation.
- 2.2. OpenRR-5k Dataset: OpenRR-5k covers diverse subjects, lighting conditions, and genuine real-world reflections rather than artificial or simulated setups.
- 2.3. Challenge Phases: Participants use paired training data, input-only validation data, and blind test images across development and testing phases.
- 2.4. Evaluation: Evaluation combines PSNR, SSIM, LPIPS, DISTS, NIQE, and visual examination by five experienced practitioners.
- 2.4. Evaluation: Final ranking uses subjective test-set scores, although objective validation metrics are also collected for all eleven teams.
3. Challenge Results and Analysis
The challenge attracted broad participation and evaluated eleven final teams using both objective metrics and subjective human assessment. RRay led the rankings, while the analysis shows that perceptual quality, progressive restoration, and diverse data matter beyond pixel-level fidelity.
- RRay won first prize, followed by Xreflect Master and AIIALab, while VIP Lab, YuFans, and KLETech-CEVI received third prizes.
- Over 100 registrations produced 1,000 submissions, with eleven teams completing the final testing phase.
- The evaluation combined objective metrics on validation data with subjective human scores on test data, but official ranking used only subjective scores.
- RRay achieved the highest subjective score of 4.45 and the best PSNR of 36.1688, while Xreflect Master scored 4.31 subjectively and obtained the lowest DISTS of 0.0127.
- Comparable PSNR values can yield different subjective scores, showing that objective metrics incompletely capture perceptual quality.
- Top methods commonly use multi-stage or refinement-based designs and external data, while pixel-level optimization alone is insufficient for high perceptual quality.
4. Challenge Methods
The competing methods use varied restoration architectures and training strategies, including cascaded refinement, larger backbones, diffusion-based distillation, and adversarial or depth-aware optimization. Across teams, methods supplement OpenRR-5k with additional data or specialized training procedures to address complex reflections.
- The participating methods are consolidated in Table 3 with technical specifications and computational characteristics, followed by detailed methodology descriptions.
- RdNafNet uses a two-stage cascaded pipeline to remove prominent reflections first and recover details lost by single-stage restoration.
- RdNafNet supplements challenge data with real-world and synthetic samples, including 29,771 paired real-world images from RRW.
- RDNet-XL replaces FocalNet-L with FocalNet-XL to strengthen multi-scale representation and global contextual modeling under complex reflections.
- RDNet-XL combines OpenRR-5k with multiple paired reflection-removal datasets and a 1,000-pair hard-case set generated using diffusion-based methods.
- MS-RDNet applies three-stage refinement with adversarial perception, perceptual fine-tuning, depth-consistency scoring, and model merging.
4.4. VIP Lab – Complementary Mixture-of-Experts and Complementary Cross-Attention for Single Image Reflection Separation in the Wild
VIP Lab builds on complementary mixture-of-experts and cross-attention, trains in two stages from mixed synthetic and real data, and uses test-time and post-training ensembling. The method also compares alternative data and augmentation choices, including a domain-gap failure from synthetic SIRS data.
- VIP Lab’s architecture is a dual-stream framework combining Complementary Mixture-of-Experts and Complementary Cross-Attention.
- Training first uses synthetic and real open datasets, then fine-tunes on OpenRR-5k at increased resolution.
- RDNet uses a FocalNet-Large backbone, four hierarchical RevCol SubNets, three NAFBlock decoders, and prompt-based modulation from a frozen ConvNext classifier.
- SWA improved performance by +0.05 dB over the best checkpoint, while 8× geometric TTA increased PSNR by approximately 1.8 dB.
- The XReflection-based encoder–decoder extracts hierarchical features and uses skip connections to reconstruct the clean transmission layer while preserving spatial information.
- The network estimates reflection residuals rather than transmission directly, then subtracts the prediction from the input; multi-scale aggregation handles varied reflection sizes and intensities.
4.7. PSU – DUSKAN: Dual Spectral Kolmogorov-Arnold Network
DUSKAN is a dual-path U-Net designed to balance global degradation removal with local texture preservation. Its spectral-spatial and adaptive paths are dynamically blended, and the model is trained from scratch using only official challenge data.
- DUSKAN uses a symmetric four-level U-Net whose DUSKANBlocks combine spectral-spatial and adaptive processing paths.
- Path A models global frequency shifts using FFT magnitude modulation, spatial positional encoding, and multi-scale depthwise convolutions.
- Path B uses Kolmogorov-Arnold polynomial-basis activations to handle spatially varying degradations.
- A learned per-stage gating weight dynamically blends both paths at each scale.
- DUSKAN uses only official training and validation data, with geometric rotations and flips for augmentation.
- The model trains from scratch for 500 epochs and combines L1, VGG perceptual, and focal frequency losses; inference uses one full-resolution forward pass without TTA.
4.8. SiGMoid – Two-head Restormer with checkpoint soup and horizontal flip TTA
SiGMoid uses a two-head Restormer to separately predict transmission and reflection, with decomposition-based supervision. Its final system combines low-learning-rate checkpoint averaging with horizontal-flip test-time augmentation and achieves strong validation metrics without extra data.
- Architecture: The 4-stage Restormer predicts six channels: three for transmission and three for reflection, explicitly decomposing the blended input.The two-head design replaces single-output restoration with separate transmission and reflection predictions.
- Training: Training supervises transmission reconstruction, proxy reflection estimation, and transmission-reflection consistency using λr = 0.2 and λc = 0.05.The consistency term encourages the two predicted components to reconstruct the observed blended image.
- Model selection: The final submission averages three stable checkpoints: the best low-learning-rate model and continuation snapshots from epochs 40 and 48.The checkpoint soup was reported as a reliable practical gain over selecting one late epoch.
- Inference: Horizontal-flip TTA was retained because heavier D4 and multi-scale variants did not improve validation performance.Inference pads images to a multiple of 8, restores them, crops to the original size, and averages original and flipped predictions.
- Results: PSNR 34.246, LPIPS 0.0296, and SSIM 0.9747 were obtained on the CodeBench validation setting without extra data.The same checkpoint and inference configuration produced both validation and test outputs.
4.10. refineX – Progressive Restormer U-Net with Multi-Loss Training for Single Image Reflection Removal
refineX develops a hierarchical Restormer-style U-Net with multiscale feature processing, skip connections, and progressive coarse-to-fine training. The broader design builds on RDNet while adding a frozen DINOv2 semantic prior to reduce blur under strong reflections.
- SIRR-Net architecture: SIRR-Net uses a four-scale symmetric encoder-decoder with pixel-unshuffle downsampling, pixel-shuffle upsampling, skip fusion, and a global residual connection.The encoder reaches a 512-channel bottleneck before the decoder reconstructs the clean image.
- Transformer blocks: Each Transformer Block combines MDTA attention, GDFN nonlinear modulation, and SE channel recalibration.MDTA uses channel-wise attention with O(C^2) complexity instead of O((HW)^2).
- Training: Progressive training moves from 128-pixel coarse patches to 256-pixel intermediate training and 384-pixel perceptual fine-tuning.The schedule is intended to stabilize optimization from coarse structures to finer detail.
- Semantic prior: The RDNet-based design adds a frozen DINOv2 ViT-B/14 branch whose adapted semantic features are fused into four encoder levels.Only four lightweight adapters are trained, while the frozen semantic branch is used to reduce blurring artifacts.
- Training: The enhanced model is fine-tuned on OpenRR-1k and OpenRR-5k with the original RDNet combination of MSE, gradient-domain L1, and VGG-19 perceptual losses.Training uses RDNet pretrained weights, DINOv2 adapters, AdamW, and validation-PSNR-based learning-rate reduction.
5. Comparison with Generative Methods
The comparison examines generative models alongside discriminative challenge methods for reflection removal. Generative approaches show stronger detail restoration and semantic consistency in difficult, heavily occluded regions, but can introduce stochastic deviations.
- Evaluation setup: Generative baselines, including gpt-image and Nano Banana series, are evaluated alongside discriminative challenge methods.The comparison explores whether generative models can contribute useful priors for SIRR.
- Observed limitations: Competition methods such as RRay and OPPO-baseline can struggle when complex nonlinear reflections overlap high-frequency background textures.This difficulty is described for challenging reflection patterns rather than all scenes.
- Generative behavior: Generative methods use prior knowledge to reconstruct the underlying transmission layer in heavily corrupted regions.The paper characterizes this behavior as hallucinating plausible underlying content.
- Comparison: Nano Banana and gpt-image-2 restore sharper background detail and suppress residual ghosting more effectively than standard CNN or Transformer pipelines.Generative models also maintain better semantic integrity in heavily occluded regions, although slight stochastic variations may occur.
6. Conclusion
The challenge advances real-world single-image reflection removal through the OpenRR-5k benchmark and broad participation. Its results demonstrate effective deep-learning approaches while highlighting persistent difficulty in maintaining consistent perceptual quality across diverse scenes.
- Conclusion: OpenRR-5k contains 5,000 training pairs, 300 validation pairs, and 100 test images captured across diverse real-world environments.The test images are used to evaluate generalization, while the paper reports remaining perceptual-quality challenges across scenes.
- Conclusion: The challenge attracted strong research and industry participation and demonstrated the effectiveness of recent deep-learning methods for reflection removal.The authors position the dataset, benchmark, and analysis as resources for further work on practical reflection removal.
- Conclusion: Consistent perceptual quality across diverse scenes remains a central unresolved challenge.The conclusion identifies robustness across varied real-world conditions as an important direction for future methods.
Appendix A. Teams and Affiliations
The appendix lists participating teams’ affiliations across universities, laboratories, and companies in multiple countries and regions. These include organizations in the United States, Europe, Asia, and the Middle East.
- Affiliations include OPPO, Meta, Intsig Information Co., Ltd., and Xiaomi Inc.
- Additional university and laboratory affiliations span Germany, India, Taiwan, China, South Korea, Saudi Arabia, and Japan.
- Academic affiliations include institutions such as the University of Illinois Urbana-Champaign, Harbin Institute of Technology, and the National University of Singapore.
- KLE Technological University affiliations cover computer applications, computer science and engineering, electronics and communication engineering, and visual intelligence.
- Other listed affiliations include robotics, Internet-of-Things, computational intelligence, intelligent systems, data science, and life science and systems engineering laboratories.
Appendix B. Extended Visual Comparisons
The appendix provides extended visual comparisons for seven test images, covering test0074, test0011, test0015, test0044, test0069, test0084, and test0090.
- Visual comparisons are reported for test0074, test0011, and test0015.
- Additional visual comparisons cover test0044, test0069, test0084, and test0090.