Efficient Perceptual Image Super Resolution: AIM 2025 Study and Benchmark

2025-10-14

AI Summary

Perception-oriented super-resolution (SR) models suffer from low efficiency, hindering real-world deployment. Method: Under strict constraints (≤5M parameters, ≤2000 GFLOPs, 960×540 input), we construct the first practical 4K degradation test set comprising 500 images and establish a comprehensive efficient perceptual SR benchmark. We propose a lightweight network architecture, multi-degradation modeling, and perceptual loss co-optimization, integrated with a no-reference evaluation strategy to achieve high-fidelity reconstruction without ground-truth supervision. Contribution/Results: Our best-performing method surpasses Real-ESRGAN in visual quality across the entire test set, significantly breaking the conventional trade-off between perceptual performance and computational cost. It sets a new state-of-the-art benchmark for efficient perceptual SR, enabling scalable deployment while preserving fidelity.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Machine Learning: Learning on the Edge & Model Compression
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Systems and Infrastructure for Web, Mobile and WoT: Web performance, measurement, and characterization
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

This paper presents a comprehensive study and benchmark on Efficient Perceptual Super-Resolution (EPSR). While significant progress has been made in efficient PSNR-oriented super resolution, approaches focusing on perceptual quality metrics remain relatively inefficient. Motivated by this gap, we aim to replicate or improve the perceptual results of Real-ESRGAN while meeting strict efficiency constraints: a maximum of 5M parameters and 2000 GFLOPs, calculated for an input size of 960x540 pixels. The proposed solutions were evaluated on a novel dataset consisting of 500 test images of 4K resolution, each degraded using multiple degradation types, without providing the original high-quality counterparts. This design aims to reflect realistic deployment conditions and serves as a diverse and challenging benchmark. The top-performing approach manages to outperform Real-ESRGAN across all benchmark datasets, demonstrating the potential of efficient methods in the perceptual domain. This paper establishes the modern baselines for efficient perceptual super resolution.

Problem

• Improving perceptual quality in efficient super-resolution methods
• Balancing high perceptual results with strict computational constraints
• Establishing benchmarks for realistic deployment of super-resolution models

Innovation

• Efficient perceptual super-resolution with 5M parameters
• Optimized for 2000 GFLOPs computational budget
• Outperforms Real-ESRGAN across benchmark datasets