Source-linked AI summary
The Eleventh NTIRE 2026 Efficient Super-Resolution Challenge Report
Bin Ren, Hang Guo, Yan Shu, Jiaqi Ma, Ziteng Cui, Shuhong Liu, Guofeng Mei, Lei Sun, Zongwei Wu, Fahad Shahbaz Khan, Salman Khan, Radu Timofte, Yawei Li, Hongyuan Yu, Pufan Xu, Chen Wu, Long Peng, Jiaojiao Yi, Siyang Yi, Yuning Cui, Jingyuan Xia, Xing Mou, Keji He, Jinlin Wu, Zongang Gao, Sen Yang, Rui Zheng, Fengguo Li, Yecheng Lei, Wenkai Min, Jie Liu, Keye Cao, Shubham Sharma, Manish Prasad, Haobo Li, Matin Fazel, Abdelhak Bentaleb, Rui Chen, Shurui Shi, Zitao Dai, Qingliang Liu, Yang Cheng, Jing Hu, Xuan Zhang, Rui Ding, Tingyi Zhang, Hui Deng, Mengyang Wang, Fulin Liu, Jing Wei, Qian Wang, Hongying Liu, Mingyang Li, Guanglu Dong, Zheng Yang, Chao Ren, Hongbo Fang, Lingxuan Li, Lin Si, Pan Gao, Moncef Gabbouj, Watchara Ruangsang, Supavadee Aramvith
TL;DR
Efficient single-image super-resolution seeks lower runtime, parameter count, and FLOPs without sacrificing reconstruction quality. This paper reviews the NTIRE 2026 challenge, its methods and evaluation, and the resulting design patterns, including pruning with distillation and hardware-aware optimization. The challenge attracted 97 registered participants and 15 valid teams, while SPANV2 achieved the best runtime through a custom fused CUDA kernel.
Problem
Efficient super-resolution must reduce runtime, parameters, or FLOPs while preserving reconstruction quality for resource-constrained applications.
Method
The paper reviews the NTIRE 2026 Efficient Super-Resolution Challenge, its evaluation protocol, participating methods, and reported results.
Results
97 participants registered and 15 teams made valid submissions; SPANV2 achieved the best runtime using a custom fused CUDA kernel.
Takeaways & Limitations
Pruning with distillation and hardware-aware optimization emerged as effective routes to efficient super-resolution, complementing architectural efficiency improvements.
Takeaways & Limitations
The challenge’s reported models retain high runtimes, indicating scope for more holistic optimization beyond parameter and FLOP reduction.
Abstract
from arXiv · showhide
This paper reviews the NTIRE 2026 challenge on efficient single-image super-resolution with a focus on the proposed solutions and results. The aim of this challenge is to devise a network that reduces one or several aspects, such as runtime, parameters, and FLOPs, while maintaining PSNR of around 26.90 dB on the DIV2K_LSDIR_valid dataset, and 26.99 dB on the DIV2K_LSDIR_test dataset. The challenge had 95 registered participants, and 15 teams made valid submissions. They gauge the state-of-the-art results for efficient single-image super-resolution.
1. Introduction
Single-image super-resolution reconstructs high-resolution images from low-resolution observations, but practical deployment requires balancing reconstruction quality with efficiency. The challenge targets this balance by reducing runtime, parameters, or FLOPs while preserving specified PSNR thresholds.
- SR reconstructs a high-resolution image from a single low-resolution observation degraded by blurring and down-sampling.
- SR supports mobile photography, video streaming, surveillance, medical imaging, and remote sensing when hardware, bandwidth, or storage constrain high-resolution acquisition.
- 5 approaches such as pruning, quantization, neural architecture search, Mamba, and distillation reduce computational overhead while preserving reconstruction quality.
- The challenge seeks ×4 super-resolution models that improve efficiency over SPAN while maintaining at least 26.90 dB on DIV2K LSDIR valid and 26.99 dB on DIV2K LSDIR test.
- The benchmark standardizes efficient SR evaluation, encourages practically deployable solutions, and provides insights into effective efficient-network design strategies.
2. NTIRE 2026 Efficient Super-Resolution Challenge
The NTIRE 2026 Efficient Super-Resolution Challenge benchmarks methods against SPAN using reconstruction quality, runtime, parameters, and FLOPs. Its protocol combines shared data and evaluation rules with a scoring system designed to compare efficient ×4 SR models.
- The challenge uses DIV2K and LSDIR, comprising 1,000 and 86,991 high-resolution images respectively, with separate training, validation, and test splits.
- SPAN is the baseline, with 0.151 M parameters, 26.94 dB validation PSNR, 27.01 dB testing PSNR, and 7.65 ms average runtime.
- Participants must reduce runtime, parameters, or FLOPs relative to SPAN while maintaining 26.90 dB validation and 26.99 dB test PSNR.
- Evaluation measures validation and test PSNR, inference runtime, parameter count, and FLOPs, with runtime averaged across 200-image sets and FLOPs measured at 256 × 256 input size.
- The final score combines runtime, FLOPs, and parameter scores, with weights of 0.8, 0.1, and 0.1 to prioritize speed efficiency.
3. Challenge Results
The challenge evaluates efficient ×4 super-resolution through runtime, FLOPs, parameters, and reconstruction quality, with runtime emerging as the decisive factor in overall rankings. Submissions combine architectural compression, distillation, attention reuse, large-scale training, and hardware-aware optimization to balance efficiency with PSNR.
- Runtime: The top three runtime solutions averaged below 7 ms, while the first six teams averaged below 8 ms on validation and test sets.XiaomiMM achieved the shortest runtime, followed by DISP and BOE AIoT.
- Parameters: ZenoSR achieved the smallest model with 0.038M parameters, while XuptSR and XSR followed with 0.051M and 0.086M parameters.These models nevertheless exhibited relatively high runtimes, exposing a trade-off between parameter count and inference speed.
- FLOPs: ZenoSR, XuptSR, and XSR led the FLOPs sub-track with 2.68G, 3.30G, and 5.24G, respectively.Their close sub-track scores mirrored the parameter rankings, but the associated runtime results indicate that low arithmetic complexity does not guarantee fast inference.
- PSNR: HAESR achieved 27.22 dB test PSNR, followed by WMESR at 27.06 dB and Sunflower and XuptSR at 27.03 dB.The challenge required at least 26.90 dB on validation and 26.99 dB on test; 14 teams met the adjusted benchmark, while MDAP did not.
- Overall Evaluation: XiaomiMM ranked first overall, followed by BOE AIoT and PKDSR, using a unified score combining runtime, FLOPs, and parameters.The overall evaluation synthesizes the three efficiency metrics into one score.
- Evaluation Trends: Runtime regained decisive influence over rankings, with XiaomiMM placing first overall after ranking first in runtime but fifth in parameters and sixth in FLOPs.The scoring protocol used increased runtime weight, while reasonably competitive complexity metrics remained important.
4. Challenge Methods and Teams
The challenge teams explore efficient SR through architectural redesign, attention mechanisms, operator fusion, pruning, distillation, and training-aware optimization. Their methods target lower runtime, parameter count, or computation while preserving reconstruction quality.
- XiaomiMM: SPANV2 combines learned 1×1 channel mixing, fused CUDA attention, and a near-pixel upsampling branch in a compact CNN.The model uses five SPABV2 blocks, a near-pixel branch, and depthwise-separable fusion before PixelShuffle×4.
- XiaomiMM: SPANV2’s learned attention produces a full C ×C map from f3, enabling content-adaptive suppression and cross-channel gating with C2 additional parameters per block.Unlike SPAN’s parameter-free attention, the map can take negative values and mix channels without softmax or normalization.
- XiaomiMM: Fusing Conv1×1, addition, and multiplication into one CUDA kernel removes 3× redundant DRAM round-trips in the attention computation.The implementation uses shared-memory weight caching, vectorized loads, register caching, loop unrolling, and channel-specialized kernels.
- Other team methods: Other teams reduce efficiency costs through channel pruning with distillation, lightweight adaptive fusion, Mamba-based gating, shared attention maps, and parameter-free attention.Examples include BOE AIoT’s 32-to-20-channel pruning, ACN’s reduced-depth backbone, MambaGate-SR’s lightweight Mamba branch, and HAESR’s SHAB modules.
- Other team methods: Additional approaches include exact operator fusion with DSCLoRA distillation, progressive spatial training with transient FFT supervision, architecture-guided multi-head attention, and reparameterization with knowledge distillation.SAFMN-Deep15 instead prioritizes 15 sequential SAFM blocks under a 0.15M-parameter constraint using compact 40-dimensional features.
XiaomiMM
The XiaomiMM entry is titled “SPANV2: Beyond Parameter-free Attention for Efficient Super-Resolution.”
- The submission is titled “SPANV2: Beyond Parameter-free Attention for Efficient Super-Resolution.”
- The listed contributors include Hongyuan Yu, Pufan Xu, and Chen Wu.
- The contributor list includes affiliations spanning Xiaomi and several Chinese universities and research institutions.
CUIT HTT
The CUIT HTT entry identifies Chengdu University of Information Technology as an affiliation.
- The listed affiliation is Chengdu University of Information Technology.
- The affiliation is numbered as institution 1.
- The affiliation information places the team within an academic university setting.
IN2GM
The supplied passages associate this entry with the DualWave-Mamba and SAFMN-Deep15 submission titles and list several Chinese institutional affiliations.
- The listed affiliations include Xi’an University of Posts and Telecommunications, National Engineering Laboratory for Cyber Event Warning and Control Technologies, and Tianjin University Medical School.
- One listed submission title is “DualWave-Mamba: A Wavelet-Guided Dual-Branch Network for Efficient Image Super-Resolution.”