Source-linked AI summary
The Third Challenge on Image Denoising at NTIRE 2026: Methods and Results
Lei Sun, Hang Guo, Bin Ren, Shaolin Su, Xian Wang, Danda Pani Paudel, Luc Van Gool, Radu Timofte, Yawei Li
TL;DR
Image denoising remains an ill-posed reconstruction problem, and the paper benchmarks modern neural architectures under standardized AWGN at σ = 50. It synthesizes the NTIRE 2026 challenge results, where BuptMM achieved the top PSNR of 29.90 dB and the leading teams were closely matched. The benchmark permits broad architectural and computational choices while enforcing data-partition safeguards.
Problem
Image denoising must recover noise-free images from corrupted observations, while the challenge seeks high-fidelity restoration under severe AWGN corruption.
Method
The paper reports a challenge benchmark that evaluates unconstrained denoising architectures using PSNR under AWGN at σ = 50 and synthesizes methods from participating teams.
Results
BuptMM ranked first with 29.90 dB PSNR, while the leading four teams were separated by less than 0.03 dB.
Takeaways & Limitations
The challenge places the current high-noise image-denoising frontier near 30 dB PSNR and provides a comparison point for diverse restoration architectures.
Takeaways & Limitations
Validation and test distributions differed by 0.92 dB, and validation improvements from tiling and model averaging did not transfer to test.
Abstract
from arXiv · showhide
This paper reports on the NTIRE 2026 Challenge on Image Denoising, specifically focusing on the high-noise regime ($σ= 50$). The competition investigates advanced neural architectures designed to restore high-fidelity details from images corrupted by additive white Gaussian noise (AWGN). Unlike constrained benchmarks, this track emphasizes peak quantitative performance, measured by Peak Signal-to-Noise Ratio (PSNR), without limitations on parameter count or computational overhead. By synthesizing contributions from 20 finalist teams out of 116 registrants, this report benchmarks the latest technical innovations and provides a comprehensive snapshot of the current state-of-the-art in unconstrained image restoration.
1. Introduction
Image denoising seeks to invert corrupted observations into noise-free images, but diverse acquisition noise makes this restoration ill-posed. The NTIRE 2026 challenge therefore evaluates methods under standardized AWGN at high noise.
- Image denoising reconstructs latent noise-free images from corrupted observations.
- Real acquisition pipelines combine Gaussian, Poisson, and structured quantization noise, making restoration inherently ill-posed.
- The challenge uses AWGN as a standardized benchmark for quantitatively comparing denoising architectures.
- NTIRE 2026 targets denoising in a high-noise regime with σ = 50 to assess modern architectures under extreme restoration conditions.
- The broader challenge context includes related NTIRE 2026 image-restoration competitions across multiple tasks.
2. NTIRE 2026 Image Denoising Challenge
The challenge is designed to advance image restoration through rigorous benchmarking and collaboration. It combines methodological evaluation with a unified reporting framework covering data, protocols, rankings, and participating architectures.
- The challenge pursues methodological breakthroughs in image restoration, equitable benchmarking, and stronger academia–industry collaboration.
- Its unified benchmark organizes dataset characteristics, evaluation protocols, final rankings, and architectural innovations from participating teams.
3. Datasets
The challenge evaluates high-noise denoising using DIV2K and LSDIR data, AWGN corruption at σ = 50, and PSNR- and SSIM-based assessment on a 200-image test ensemble.
- DIV2K provides 1,000 RGB images, while LSDIR provides 86,991 high-quality high-resolution samples with separate training, validation, and test splits.
- The competition’s core objective is high-fidelity denoising, with PSNR serving as the primary performance metric.
- Training data are corrupted by injecting AWGN at a consistent noise level of σ = 50.
- Final evaluation uses 100 noisy DIV2K images and 100 noisy LSDIR images, with ground-truth references withheld from participants.
- The leaderboard computes PSNR and SSIM over the 200-image test ensemble.
4. Challenge Results
The challenge drew 116 registrants and 20 valid finalist teams, whose methods combine transformer, Mamba, progressive-learning, frequency-domain, and test-time techniques. BuptMM led with 29.90 dB PSNR, while the top four teams were separated by less than 0.03 dB.
- Challenge Results: BuptMM ranked first with 29.90 dB PSNR, followed by I2WM&JNU at 29.89 dB, MIIE at 29.873 dB, and Titans at 29.871 dB.
- Challenge Results: The leading four teams were separated by less than 0.03 dB, indicating exceptionally close competition.
- Participants: The challenge attracted 116 registered participants, with 20 teams submitting valid results.
- Technical Strategies: Leading approaches used Mamba and transformer backbones, including Restormer’s MDTA for global context and UnifyFormer’s multi-scale local aggregation.
- Technical Strategies: Progressive training, frequency-domain losses, and inference-time enhancement were recurring strategies among leading teams.
- Competition Integrity: Participants could use external data and augmentation, but training on the DIV2K validation partition was prohibited to preserve evaluation validity.
5. Challenge Methods and Teams
Finalist teams explored diverse high-noise denoising strategies, including complementary ensembles, large-scale training, inference-time refinement, attention innovations, adaptive experts, and parameter-efficient adaptation. Reported results include strong PSNR/SSIM performance and measurable gains from refinement and data scaling.
- DualExNet: DualExNet uses parallel local and global branches to model fine-grained details and global context, with bidirectional multi-scale feature exchange at skip connections.Residual Swin Transformer Blocks process local features, while a U-Net branch captures global context; learnable gating supports cross-scale fusion and residual enhancement.
- Ensembling and data scaling: +3.2 dB improvement resulted from fine-tuning on 85K LSDIR images, whereas fine-tuning on only 800 DIV2K images degraded pretrained performance by 0.06–0.19 dB.This approach fine-tuned Restormer and SCUNet on large-scale LSDIR data.
- Ensembling and data scaling: A Restormer–SCUNet ensemble combines complementary error profiles, with Restormer restoring global structure and SCUNet preserving fine-grained details.The ensemble consistently outperforms either constituent model alone.
- Inference-time refinement: I2WM&JNU combined Restormer with large-scale training, test-time augmentation, and Test-time Local Converter refinement to improve reconstruction quality.The method is framed around large-scale data and inference-time enhancement strategies.
- Inference-time refinement: SRX fused Restormer, XFormer, and SCUNet through weighted ensembling, then applied tiled test-time augmentation for high-resolution inputs.XFormer contributes spatial window-based and channel-wise attention alongside the complementary receptive fields of Restormer.
- Attention and architecture design: UnifyFormer augments self-attention with explicit multi-scale local aggregation and complements it with Restormer-style channel attention.Its four-level encoder-decoder uses UnifyBlocks, concatenation-based skip fusion, and output refinement; channel attention computes relationships across channels and is described as less sensitive to localized noise.
- Adaptive experts and efficient adaptation: MoCE-IR routes inputs among experts with different computational capacities and receptive fields using complexity-aware allocation that favors simpler experts when possible.The proposed strategy is intended to improve computational efficiency while inducing task-discriminative learning.
- Reported results: MoCE-IR achieved 29 dB PSNR and 0.84 SSIM on the provided NTIRE testing set.These results were reported for the severe noise level 50 setting.
NTIRE 2026 team
The NTIRE 2026 team includes researchers from multiple institutions and proposes DUSKAN, a four-level U-Net architecture with specialized processing paths.
- The team includes Lei Sun, Hang Guo, Bin Ren, Shaolin Su, Xian Wang, Danda Pani Paudel, Luc Van Gool, Radu Timofte, and Yawei Li.
- Its affiliations span INSAIT, Tsinghua University, the Universities of Pisa and Trento, Zhejiang University, the Computer Vision Center, the University of Würzburg, and ETH Zürich.
- DUSKAN uses a symmetric 4-level U-Net backbone with global residual learning, parallel Spectral-Spatial and Kolmogorov-Arnold Adaptive Processing paths, and learned-logit fusion.
Perceptual Vision Team
The Perceptual Vision Team is affiliated with Northwestern Polytechnical University in Xi’an, Shaanxi, China.
- The Perceptual Vision Team is affiliated with Northwestern Polytechnical University.
- The listed affiliation is located in Xi’an, Shaanxi, China.
- The affiliation identifies China as the team’s country location.
BuptMM
BuptMM includes researchers from Beijing University of Posts and Telecommunications, China University of Petroleum (East China), and JD Inc., China.
- BuptMM includes researchers affiliated with Beijing University of Posts and Telecommunications.
- The team also includes researchers from China University of Petroleum (East China).
- JD Inc., China, is listed among the team’s affiliations.
YuFans
YuFans is associated with the National University of Singapore and Zhejiang University.
- YuFans includes an affiliation with the National University of Singapore.
- YuFans also includes an affiliation with Zhejiang University.
- The listed affiliations span institutions in Singapore and China.
I2WM&JNU
The entry lists affiliations spanning Jinan University, Hangzhou Dianzi University, and Hefei University of Technology.
- Jinan University is listed as affiliation 1.
- Hangzhou Dianzi University is listed as affiliation 2.
- Hefei University of Technology is listed as affiliation 3.
Noice
The entries list multiple institutional affiliations and include a contribution titled “Two-Stage MDAE Pretraining and Supervised Fine-Tuning for Image Denoising.”
- Affiliations: The listed affiliations include Indian Institute of Information Technology Una, University of Beira Interior, UPES, and Indian Institute of Technology Mandi.
- Affiliations: The affiliations also include the Centre for Visual Information Technology at IIIT-H and TU Darmstadt.
- Affiliations: Additional affiliations include INSAIT, The City University of New York, Jiangnan University, and Zhejiang University.
- Affiliations: Other listed institutions include KLE Technological University, Shiv Nadar University, and Sardar Vallabhbhai National Institute of Technology.
- Affiliations: The affiliations further include Norwegian University of Science and Technology, University of Illinois Urbana-Champaign, and National University of Defense Technology.
- Contribution: One contribution is titled “Two-Stage MDAE Pretraining and Supervised Fine-Tuning for Image Denoising.”