The Eleventh NTIRE 2026 Efficient Super-Resolution Challenge Report
| Authors | Bin Ren et al. |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.03198 |
| Download | |
| Categories | cs.CV |
Abstract
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.
Engineering Breakdown
Plain English
The NTIRE 2026 Efficient Super-Resolution Challenge tasked 95 registered teams with building single-image super-resolution networks that minimize runtime, parameters, and/or FLOPs while maintaining PSNR around 26.90 dB on validation and 26.99 dB on test sets. Fifteen teams submitted valid solutions that were evaluated on the DIV2K_LSDIR dataset, a standard benchmark for measuring super-resolution quality. This challenge report synthesizes the state-of-the-art approaches across the competition, revealing how the research community is balancing image quality against computational efficiency—a critical tradeoff for real-world deployment of super-resolution in mobile, edge, and server environments.
Core Technical Contribution
The core contribution is a comprehensive empirical evaluation of efficient super-resolution architectures submitted by leading research groups, establishing new performance baselines for trading off PSNR (image quality), FLOPs, parameter count, and runtime. Rather than proposing a single novel method, the challenge aggregates diverse approaches—likely including knowledge distillation, neural architecture search, pruning, and lightweight convolution designs—to identify which techniques most effectively reduce computational burden without sacrificing the 26.9+ dB quality threshold. This competitive benchmark establishes what 'state-of-the-art efficient super-resolution' looks like in 2026 and creates a reproducible leaderboard that future work can measure against.
How It Works
The challenge setup defines a constrained optimization problem where participants design convolutional or transformer-based super-resolution networks tasked with 4x upsampling of low-resolution DIV2K images. Teams are evaluated on Pareto-optimal tradeoffs: achieving the target PSNR while minimizing at least one of (runtime on standard hardware, total parameters, or FLOPs). The input is a low-resolution image; the network applies learned upsampling layers (typically using sub-pixel convolution or transposed convolution) to produce a high-resolution output; the output is compared pixel-wise against ground truth using PSNR, while computational cost is measured via FLOPs counters and wall-clock inference time. The winning solutions likely employ techniques such as depthwise separable convolutions to reduce parameters, progressive upsampling to reduce intermediate feature map sizes, or knowledge distillation from larger teacher models to guide lightweight student networks.
Production Impact
For production engineers, this challenge establishes practical efficiency thresholds for super-resolution: networks can achieve photo-realistic 4x upsampling (26.9 dB PSNR) with significantly reduced computational footprint compared to older methods. This directly enables deployment in bandwidth-constrained or latency-sensitive scenarios—e.g., on-device image enhancement on mobile phones, real-time video upscaling in video conferencing, or server-side batch processing where inference cost per image matters at scale. The 15 validated solutions provide reference implementations and architectural patterns that can be adapted for specific hardware targets (mobile GPUs, TPUs, CPUs). The tradeoff analysis helps teams decide: if you have 2GB of model memory and need <100ms latency per image on an edge device, which 2026 efficient-SR approach maps to your constraint? Adoption requires benchmarking these methods on your actual hardware rather than assuming published FLOPs correlate to real inference time.
Limitations and When Not to Use This
The challenge is constrained to 4x single-image super-resolution on the DIV2K_LSDIR validation/test splits, so results may not generalize to other upsampling factors (2x, 8x), other datasets (real-world degraded images), or other tasks (video SR, face SR, medical imaging). The PSNR metric itself is well-known to not correlate perfectly with perceptual quality—solutions optimized for PSNR may produce artifacts that LPIPS or human raters would penalize. The challenge doesn't systematically evaluate robustness to out-of-distribution inputs (JPEG artifacts, noise, motion blur) that real images encounter, nor does it characterize performance across diverse hardware platforms (the actual latency and memory of a winning architecture may vary 2-5x depending on whether it runs on a mobile SoC, edge TPU, or data-center GPU). Follow-up work should establish efficient SR baselines for blind upsampling (unknown degradation), 2x/8x/16x factors, and perceptual metrics beyond PSNR.
Research Context
This challenge builds on the NTIRE competition series (running annually since 2017) and extends prior efficient-vision work that combines neural architecture search, quantization, and pruning to shrink models while preserving accuracy. It directly addresses the gap between academic super-resolution methods (often measuring only PSNR with no computational reporting) and real-world deployment constraints, where FLOPs and latency are as important as image quality. The DIV2K dataset has become the de facto benchmark for SR research; adding a labeled efficient-SR leaderboard raises the bar for future SR papers to report computational cost alongside accuracy. This opens a research direction toward 'pareto-efficient' computer vision—not just achieving state-of-the-art quality, but achieving it at a cost level commensurate with deployment requirements.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::