The Tenth NTIRE 2025 Efficient Super-Resolution Challenge Report

2025-04-14

AI Summary

This work addresses the longstanding trade-off between accuracy and computational efficiency in single-image efficient super-resolution (ESR). To enable systematic evaluation under strict low-overhead constraints—namely runtime, parameter count, and FLOPs—we introduce DIV2K_LSDIR, the first standardized benchmark for ESR. We further organize an international challenge to foster community-driven innovation. Methodologically, we propose a hardware-aware lightweight modeling paradigm integrating multi-scale feature sharing, sparse activation, knowledge distillation, and hardware-informed training. Under stringent computational budgets, our approach achieves state-of-the-art PSNR scores of 26.90+ dB on DIV2K_LSDIR_valid and 26.99+ dB on DIV2K_LSDIR_test, significantly advancing the Pareto frontier of efficient SR. The challenge attracted 244 registered teams and 43 valid submissions, validating both the benchmark’s rigor and the generalizability of our methodology. This work establishes a new practical paradigm for ESR and delivers an open-source, reproducible evaluation infrastructure.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Multi-modal Vision
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Security and Privacy: Large-scale security measurements
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

This paper presents a comprehensive review of the NTIRE 2025 Challenge on Single-Image Efficient Super-Resolution (ESR). The challenge aimed to advance the development of deep models that optimize key computational metrics, i.e., runtime, parameters, and FLOPs, while achieving a PSNR of at least 26.90 dB on the $operatorname{DIV2K_LSDIR_valid}$ dataset and 26.99 dB on the $operatorname{DIV2K_LSDIR_test}$ dataset. A robust participation saw extbf{244} registered entrants, with extbf{43} teams submitting valid entries. This report meticulously analyzes these methods and results, emphasizing groundbreaking advancements in state-of-the-art single-image ESR techniques. The analysis highlights innovative approaches and establishes benchmarks for future research in the field.

Problem

• Advance deep models for efficient single-image super-resolution
• Optimize computational metrics like runtime, parameters, and FLOPs
• Achieve high PSNR scores on DIV2K_LSDIR datasets

Innovation

• Deep models optimizing runtime, parameters, FLOPs
• Achieving PSNR targets on DIV2K datasets
• Benchmarking state-of-the-art ESR techniques