Source-linked AI summary
The Fourth Challenge on Image Super-Resolution ($\times$4) at NTIRE 2026: Benchmark Results and Method Overview
Zheng Chen, Kai Liu, Jingkai Wang, Xianglong Yan, Jianze Li, Ziqing Zhang, Jue Gong, Jiatong Li, Lei Sun, Xiaoyang Liu, Radu Timofte, Yulun Zhang, Jihye Park, Yoonjin Im, Hyungju Chun, Hyunhee Park, MinKyu Park, Zheng Xie, Xiangyu Kong, Weijun Yuan, Zhan Li, Qiurong Song, Luen Zhu, Fengkai Zhang, Xinzhe Zhu, Junyang Chen, Congyu Wang, Yixin Yang, Zhaorun Zhou, Jiangxin Dong, Jinshan Pan, Shengwei Wang, Jiajie Ou, Baiang Li, Sizhuo Ma, Qiang Gao, Jusheng Zhang, Jian Wang, Keze Wang, Yijiao Liu, Yingsi Chen, Hui Li, Yu Wang, Congchao Zhu, Saeed Ahmad, Ik Hyun Lee, Jun Young Park, Ji Hwan Yoon, Kainan Yan, Zian Wang, Weibo Wang, Shihao Zou, Chao Dong, Wei Zhou, Linfeng Li, Jaeseong Lee, Jaeho Chae, Jinwoo Kim, Seonjoo Kim, Yucong Hong, Zhenming Yan, Junye Chen, Ruize Han, Song Wang, Yuxuan Jiang, Chengxi Zeng, Tianhao Peng, Fan Zhang, David Bull, Tongyao Mu, Qiong Cao, Yifan Wang, Youwei Pan, Leilei Cao, Xiaoping Peng, Wei Deng, Yifei Chen, Wenbo Xiong, Xian Hu, Yuxin Zhang, Xiaoyun Cheng, Yang Ji, Zonghao Chen, Zhihao Xue, Junqin Hu, Nihal Kumar, Snehal Singh Tomar, Klaus Mueller, Surya Vashisth, Prateek Shaily, Jayant Kumar, Hardik Sharma, Ashish Negi, Sachin Chaudhary, Akshay Dudhane, Praful Hambarde, Amit Shukla, Shijun Shi, Jiangning Zhang, Yong Liu, Kai Hu, Jing Xu, Xianfang Zeng, Amitesh M, Hariharan S, Chia-Ming Lee, Yu-Fan Lin, Chih-Chung Hsu, Nishalini K, Sreenath K A, Bilel Benjdira, Anas M. Ali, Wadii Boulila, Shuling Zheng, Zhiheng Fu, Feng Zhang, Zhanglu Chen, Boyang Yao, Nikhil Pathak, Aagam Jain, Milan Kumar, Kishor Upla, Vivek Chavda, Sarang N S, Raghavendra Ramachandra, Zhipeng Zhang, Qi Wang, Shiyu Wang, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Yuqi Li, Chuanguang Yang, Weilun Feng, Zhuzhi Hong, Hao Wu, Junming Liu, Yingli Tian, Amish Bhushan Kulkarni, Tejas R R Shet, Saakshi M Vernekar, Nikhil Akalwadi, Kaushik Mallibhat, Ramesh Ashok Tabib, Uma Mudenagudi, Yuwen Pan, Tianrun Chen, Deyi Ji, Qi Zhu, Lanyun Zhu, Heyan Zhangyi
TL;DR
Image super-resolution must recover HR images from bicubic ×4 LR inputs while balancing restoration fidelity and perceptual realism. This report organizes the NTIRE 2026 benchmark around separate PSNR and perceptual tracks, analyzes participating methods, and finds strong results alongside recurring architectural trends. It also identifies metric–visual-quality trade-offs in some training choices.
Problem
The challenge addresses how to benchmark SR methods that optimize either pixel-wise restoration fidelity or perceptual realism under the classical bicubic ×4 setting.
Method
The report evaluates submissions on DIV2K using separate PSNR-based restoration and multi-metric perceptual tracks, then synthesizes methods and technical trends.
Results
SamsungAICamera led both tracks, reaching 33.73 dB PSNR in restoration and a 4.7853 perceptual score.
Takeaways & Limitations
Pretrained Transformer restoration models remain strong foundations, while two-stage fidelity-restoration and perceptual-enhancement pipelines are particularly effective for perceptual SR.
Takeaways & Limitations
Adding CLIP loss can improve no-reference metrics while introducing pseudo-textures and high-frequency artifacts, and removing it may produce more natural textures.
Abstract
from arXiv · showhide
This paper presents the NTIRE 2026 image super-resolution ($\times$4) challenge, one of the associated competitions of the NTIRE 2026 Workshop at CVPR 2026. The challenge aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs generated through bicubic downsampling with a $\times$4 scaling factor. The objective is to develop effective super-resolution solutions and analyze recent advances in the field. To reflect the evolving objectives of image super-resolution, the challenge includes two tracks: (1) a restoration track, which emphasizes pixel-wise fidelity and ranks submissions based on PSNR; and (2) a perceptual track, which focuses on visual realism and evaluates results using a perceptual score. A total of 194 participants registered for the challenge, with 31 teams submitting valid entries. This report summarizes the challenge design, datasets, evaluation protocol, main results, and methods of participating teams. The challenge provides a unified benchmark and offers insights into current progress and future directions in image super-resolution.
1. Introduction
Single-image super-resolution reconstructs HR images from information-limited LR inputs, making it a fundamental but ill-posed computer-vision problem. The challenge frames current progress around the tension between pixel fidelity and perceptual realism.
- Motivation: Single-image super-resolution reconstructs a high-resolution image from a low-resolution input degraded by information loss.The task supports applications including surveillance, medical imaging, and remote sensing.
- Classical SR setting: Bicubic downsampling removes high-frequency details, so SR must recover missing information using learned priors.The predefined classical setting also enables fair comparison and often generalizes to more complex degradations.
- Method evolution: Deep learning has shifted SR from interpolation toward convolutional, attention-based, Transformer, and sequence-model architectures.The cited progression includes deeper networks, residual connections, self-attention, and long-range dependency modeling.
- Perceptual SR: GAN and diffusion methods improve visual realism but may reduce fidelity, creating a distortion–perception trade-off.Diffusion models generate images through denoising and aim to model complex image distributions.
- Optimization strategies: Training and inference strategies such as pretraining, large-scale data, advanced losses, and self-ensemble are widely used to improve SR performance.These strategies are especially common in recent competitions.
- Challenge scope: The NTIRE 2026 challenge benchmarks bicubic ×4 SR on DIV2K and analyzes recent progress through restoration and perceptual tracks.The two-track design uses PSNR for restoration quality and multiple image-quality metrics for perceptual quality.
2. NTIRE 2026 Image Super-Resolution (×4)
The NTIRE 2026 ×4 challenge benchmarks recent SR methods using standardized datasets, hidden evaluation data, and separate restoration and perceptual tracks. It combines PSNR-based fidelity ranking with a multi-metric perceptual score.
- Goals: The challenge provides a venue for an up-to-date overview of SR progress and collaboration between academic researchers and industry practitioners.These are the two stated goals of the NTIRE 2026 SR challenge.
- Datasets: Two official datasets, DIV2K and LSDIR, supply HR images for bicubic ×4 LR-HR pair generation, with supplementary training data permitted.DIV2K contains 1,000 2K images, while LSDIR contains 86,991 high-quality images.
- Track protocols: The Restoration track ranks teams by PSNR between enhanced HR outputs and hidden DIV2K test ground truth.The protocol follows the previous challenge edition.
- Track protocols: The Perceptual track evaluates results with LPIPS, DISTS, CLIP-IQA, MANIQA, MUSIQ, and NIQE and ranks teams by a combined perceptual score.The score uses the listed metrics, including a normalized MUSIQ term and a clipped NIQE contribution.
- Challenge phases: During development, participants receive training pairs and validation images, may use external data, and obtain server feedback from submitted outputs.The final testing phase provides 100 LR images without HR ground truth and requires outputs, code, and a report.
3. Challenge Results
The challenge shows strong performance in both restoration and perceptual quality while revealing recurring design patterns across competitive submissions. Leading methods combine pretrained restoration backbones, inference optimization, and increasingly separated fidelity and perceptual enhancement stages.
- Track 2: Perception Quality: 4.7853 gave SamsungAICamera first place in the perceptual track, followed by VEPG at 4.7666 and HONORAICamera at 4.4787.Seven solutions scored above 4.0 and fourteen surpassed 3.6.
- Architectures and main ideas: Pretrained Transformer restoration backbones such as HAT, SwinIR, HMANet, and PFT-SR were widely adopted as competitive foundations.Teams generally adapted publicly available pretrained models rather than designing entirely new backbones.
- Architectures and main ideas: Inference-time strategies included geometric self-ensemble, overlapped tiled inference, Gaussian-weighted stitching, reflection padding, and checkpoint interpolation.These techniques target robustness on high-resolution test images.
- Architectures and main ideas: Two-stage pipelines first restore fidelity with a deterministic model and then enhance perceptual realism using a generative prior.This division assigns structural reconstruction and realistic detail synthesis to separate stages.
- Architectures and main ideas: Diffusion and rectified-flow generative priors were used to synthesize visually pleasing textures and natural-looking details for perceptual SR.These approaches differ from restoration-only models by emphasizing generative detail recovery.
- Architectures and main ideas: Teams also used degradation, semantic, and structural conditioning together with detail-aware losses and residual refinement modules.Examples include degradation descriptors, structure maps, semantic guidance, frequency supervision, gradient-weighted losses, and residual correction branches.
7. Training strategies increasingly emphasize staged
NTIRE 2026 submissions increasingly leverage strong pretrained foundations through staged or lightweight adaptation, while combining restoration backbones with generative priors. The challenge attracted broad participation and enforced rules governing training data and augmentation.
- Training strategies: Partial fine-tuning, staged training, sequential LoRA adaptation, and restoration-first / perception-second optimization reduce reliance on full retraining.
- 194 participants registered, and 31 teams submitted valid entries spanning Transformer restoration, diffusion, and rectified-flow perceptual enhancement.
- Competition constraints: Participants could use declared public external datasets, but training on benchmark test HR images was prohibited.
- Training strategies: Two-stage pipelines preserve structural consistency during fidelity restoration before adding perceptual detail with generative priors.
- Training strategies: Diffusion and rectified-flow backbones support perceptual super-resolution, with task-specific adaptation including LoRA tuning, conditional guidance, and structure-aware generation.
4. Challenge Methods and Teams
Participating teams largely combine pretrained restoration or generative backbones with inference-time enhancement, fusion, and perceptual optimization. Their methods span cascaded global-local reconstruction, training-free ensembles, diffusion-based objectives, and lightweight fusion of complementary models.
- SamsungAICamera: SamsungAICamera combines a HAT-based Global Optimization Module with a NAFNet-based Detail Enhancement Module and adaptive spatial fusion.Semantic intermediate features guide local texture restoration, while dynamic fusion preserves structure and enhances details.
- I2WM&JNU: I2WM&JNU uses an inference-only ensemble of pretrained Hybrid Network and MambaIRv2 models with TLC, self-ensemble, and weighted averaging.The pipeline avoids additional training while exploiting complementary pretrained priors.
- VEPG: VEPG builds on one-step diffusion-based OMGSR with FLUX.2-kleinbase and adds LPIPS, DISTS, GAN, restoration, and no-reference IQA losses.The objective combines restoration-oriented terms with perceptual supervision from MUSIQ, CLIP-IQA, and MANIQA.
- SR-Strugglers: SR-Strugglers fuses HAT-style and MSHAT-based pretrained transformers, applying 8× test-time augmentation to the second branch and setting the fusion weight to w = 0.04.The method is inference-only, with the fusion weight selected on the validation set to maximize PSNR.
- HONORAICamera: HONORAICamera adapts the Z-Image-Turbo generative prior with OMGSR-style training and an additional CLIP loss for semantic consistency.The team reports that CLIP loss can improve no-reference metrics but may introduce pseudo-textures and high-frequency artifacts.
- IK-LAB: IK-LAB trains a lightweight FusionNet to combine frozen HAT-IQCMix and DAT backbones with complementary attention mechanisms and reconstruction errors.Freezing both backbones preserves their restoration capability and reduces overfitting risk.
5. Methods of the Remaining Teams
The report directs readers to supplementary material for comprehensive descriptions of the remaining participating teams. These methods and implementation details are omitted from the main report because of space constraints.
- The remaining teams proposed diverse solutions and conducted extensive experimental studies during the competition.
- Detailed descriptions of their methods and implementation details are provided in Section A of the supplementary materials.
- Although omitted from the main report, these teams’ contributions provide valuable insights into image super-resolution.