Source-linked AI summary
The First Challenge on Mobile Real-World Image Super-Resolution at NTIRE 2026: Benchmark Results and Method Overview
Jiatong Li, Zheng Chen, Kai Liu, Jingkai Wang, Zihan Zhou, Xiaoyang Liu, Libo Zhu, Jue Gong, Radu Timofte, Yulun Zhang, Congyu Wang, Zihao Wang, Ke Wu, Xinzhe Zhu, Fengkai Zhang, Zhongbao Yang, Long Sun, Jiangxin Dong, Jinshan Pan, Jiachen Tu, Yaokun Shi, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Renyuan Situ, Yixin Yang, Zhaorun Zhou, Junyang Chen, Yuqi Li, Chuanguang Yang, Weilun Feng, Chuanyue Yan, Yuedong Tan, Yingli Tian, Zhenzhong Chen, Tongqi Guo, Ruhan Liu, Sangzi Shi, Huazhang Deng, Jie Yang, Wenzhuo Ma, Yuantong Zhang, Daiqin Yang, Tianrun Chen, Deyi Ji, Yuxiao Jiang, Qi Zhu, Lanyun Zhu, Yuwen Pan, Runze Tian, Mingyu Shi, Zhanfeng Feng, Yuanfei Bao, Jiaming Guo, Renjing Pei, Xin Di, Long Peng, Linfeng Jiang, Xueyang Fu, Yang Cao, Zhengjun Zha, Choulhyouc Lee, Shyang-En Weng, Yi-Cheng Liao, Jorge Tyrakowski, Yu-Syuan Xu, Wei-Chen Chiu, Ching-Chun Huang, Yoonjin Im, Jihye Park, Hyungju Chun, Hyunhee Park, MinKyu Park, Xiaoxuan Yu, Jianxing Zhang, Yuxuan Jiang, Chengxi Zeng, Tianhao Peng, Fan Zhang, David Bull, Watchara Ruangsang, Supavadee Aramvith, JiaHao Deng, Wei Zhou, Hongyu Huang, Shaohui Lin, Zihan Wang, Yilin Chen, Yunchen Li, Junbo Qiao, Wei Li, Jiao Xie, Gaoqi He, Wenxi Li
TL;DR
Mobile real-world super-resolution must recover x4 high-resolution images from low-resolution inputs under unknown degradations while remaining deployable on mobile hardware. This paper reviews the NTIRE 2026 challenge, its quality-speed evaluation, and participating solutions, finding that generative priors, compression, and adaptive latent strategies were central to competitive results.
Problem
The challenge addresses the need for perceptually strong real-world super-resolution models that also satisfy mobile deployment constraints.
Method
The paper reviews a single-track challenge that combines six IQA metrics with measured inference speed and analyzes submitted solution strategies.
Results
108 participants registered and 16 teams submitted valid models, with competitive solutions leveraging diffusion priors, knowledge distillation, architectural compression, and latent-space or timestep adaptation.
Takeaways & Limitations
The challenge results highlight generative priors, compression, and latent or timestep optimization as prominent directions for mobile real-world super-resolution.
Takeaways & Limitations
The competition restricts training by prohibiting use of DIV2K validation high-resolution images, while allowing additional datasets and no-reference IQA or simulated degradation pipelines.
Abstract
from arXiv · showhide
This paper provides a review of the NTIRE 2026 challenge on mobile real-world image super-resolution, highlighting the proposed solutions and the resulting outcomes. The challenge aims to recover high-resolution (HR) images from low-resolution (LR) counterparts generated through unknown degradations with a x4 scaling factor while ensuring the models remain executable on mobile devices. The objective is to develop effective and efficient network designs or solutions that achieve state-of-the-art real-world image super-resolution performance. The track of the challenge evaluates performance using a weighted combination of image quality assessment (IQA) score and speedup ratios. The competition attracted 108 registrants, with 16 teams achieving a valid score in the final ranking. This collaborative effort advances the performance of mobile real-world image super-resolution while offering an in-depth overview of the latest trends in the field.
1. Introduction
Mobile real-world super-resolution must recover high-resolution images under complex, unknown degradations while meeting strict deployment constraints. The NTIRE 2026 challenge addresses this gap by jointly emphasizing perceptual quality, efficiency, and mobile hardware compatibility.
- Single-image super-resolution reconstructs high-resolution images from low-resolution counterparts, but the inverse problem is ill-posed.
- Real-world degradations such as sensor noise, compression artifacts, and unknown blur are more complex than the bicubic assumptions used by classical benchmarks.
- Modern SR methods have progressed from interpolation and shallow CNNs to Transformers, state-space models, GANs, and diffusion models.
- Advanced generative models can require billions of parameters, massive computation, seconds per image, and hundreds of megabytes of memory, limiting real-time mobile inference.
- The challenge targets x4 real-world SR while imposing strict parameter, FLOP, and mobile-operator constraints.
- Its evaluation jointly rewards perceptual quality and relative inference speed, favoring solutions that combine visual quality with low latency.
2. NTIRE 2026 Challenge on Mobile RealWorld Image Super-Resolution
The NTIRE 2026 challenge evaluates mobile real-world super-resolution through perceptual quality and hardware-measured speed, using standardized datasets, submissions, and verification. Its final rankings compare teams using a weighted perceptual score and speedup relative to OSEDiff.
- The challenge seeks state-of-the-art real-world super-resolution under unknown degradations while maintaining efficiency for mobile deployment.
- Participants use official DIV2K and LSDIR data with x4 LR-HR construction and may incorporate additional external training data.
- The single track evaluates six IQA metrics, including LPIPS, DISTS, CLIP-IQA, MANIQA, MUSIQ, and NIQE, to measure perceptual quality.
- Inference speed is measured with FP16 precision on the MediaTek Dimensity 8400 platform.
- Final rankings weight perceptual score and speedup relative to OSEDiff on the MediaTek Dimensity 8400, using 100 DIV2K validation images.
- The benchmark uses 128 × 128 inputs and 512 × 512 outputs for its x4 super-resolution setting.
- The process includes development uploads, hidden-HR metric feedback, final outputs, code, checkpoints, reports, and organizer verification of speed.
3. Challenge Results
The challenge elicited diverse mobile super-resolution strategies, with teams combining generative priors, compression, latent alignment, hybrid architectures, and staged training to balance perceptual quality and efficiency. The final ranking included 16 valid teams from 108 registrants under rules designed to support fair comparison.
- Key technical trends: One-step diffusion models and LoRA adaptation formed a common foundation for competitive solutions.
- Key technical trends: Knowledge distillation and aggressive pruning transferred capabilities from heavy generative teachers to lightweight mobile students.
- Key technical trends: Latent inversion and timestep alignment anchored low-quality inputs to diffusion trajectories, reducing hallucinated artifacts and structural distortions.
- Key technical trends: Hybrid GAN-diffusion architectures and refinement modules combined structural consistency with perceptual texture generation.
- Participation: 108 participants registered, while 16 teams submitted valid models for the final ranking.
- Evaluation rules: Competition rules prohibited DIV2K validation-image training while permitting additional datasets and IQA-based training or testing practices.
4. Challenge Methods and Teams
The challenge methods combine compact restoration architectures, perceptual-quality-aware optimization, diffusion distillation, latent alignment, and hybrid ensembles to balance image quality with mobile efficiency. The described teams adapt these strategies through deployment-aware training and perceptual supervision.
- VIPSL: VIPSL combines a compact PLKSR-Rep backbone with multi-stage IQA-aware fine-tuning for mobile real-world super-resolution.The method uses an L1-plus-IQA composite loss, followed by score-oriented tuning emphasizing LPIPS and DISTS with additional IQA constraints.
- Antman: Antman retains a simple fully convolutional RRDBNet architecture while redesigning training around no-reference and full-reference perceptual metrics.Its objective uses MANIQA, MUSIQ, CLIP-IQA, LPIPS, and DISTS to improve perceptual realism without adding deployment-heavy architectural components.
- SamsungAICamera: SamsungAICamera distills a one-step diffusion teacher into a pruned U-Net and lightweight TAESD decoder, with a shallow DEM for latent detail refinement.The design targets perceptual richness while reducing runtime overhead relative to heavier diffusion systems.
- TODSR: TODSR improves one-step diffusion super-resolution through latent statistical-level alignment and asynchronous conditional score distillation.These components align low-quality latent features with diffusion states and exploit timestep-aware guidance during training.
- YuFans: YuFans blends DiffBIR v2.1 and Real-ESRGAN outputs with fixed pixel-level ratios and lightweight post-processing, avoiding additional model retraining.The team reports that blending retains most perceptual gains with better practical robustness because pure diffusion outputs may exceed platform time limits.
- IMAG2006: Team IMAG2006 builds on OMGSR, replacing its DISTS supervision with LPIPS and DISTS losses and adding differentiable NR-IQA supervision.The training objective combines low-quality-input alignment, pixel fidelity, adversarial realism, and perceptual quality signals.
5. Methods of the Remaining Teams
The remaining valid teams explored diverse efficient restoration and perceptual-enhancement strategies. Their methods reflect tradeoffs among perceptual quality, robustness, and inference efficiency, with fuller details provided in supplementary materials.
- Remaining teams: Remaining teams used efficient CNN or Transformer backbones, diffusion-prior enhancement, perceptual-loss reweighting, and data-degradation strategies.The main manuscript summarizes these approaches rather than detailing each team individually.
- Remaining teams: These solutions provide technical diversity across perceptual quality, robustness, and inference-efficiency tradeoffs.Complete method descriptions, implementation settings, and additional ablations are included in the supplementary materials.