Efficient Perceptual Image Super Resolution: AIM 2025 Study and Benchmark

TL;DR

The paper presents EPSR, a benchmark and study for efficient perceptual image super-resolution under strict resource limits (5M parameters, 2000 GFLOPs for 960×540 inputs). It introduces PSR4K as a diverse 4K-scale testbed and uses a multi-metric score to rank methods relative to Real-ESRGAN, balancing perceptual quality with efficiency. Among the submitted methods, VPEG achieves the best perceptual metrics while remaining highly compact, with MiAlgo and IPIU offering competitive gains under the same constraints; across standard benchmarks, perceptual improvements often accompany artifacts, underscoring limitations in current NR-IQA metrics. The work establishes modern baselines for efficient perceptual SR and highlights the potential for deployment-ready solutions, while signaling a need for improved perceptual evaluation in the presence of hallucinations. Overall, the study demonstrates that meaningful perceptual gains are achievable within stringent computational budgets, paving the way for real-time, on-device perceptual SR systems.

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.

Figures (5)

• Figure 1: We show two samples of high-resolution (HR) and low-resolution (LR) images. For memory and layout considerations in this document, the HR images have been down-scaled to match the spatial dimensions of the LR images. Consequently, the visual differences between HR and LR examples may appear less pronounced in this figure. It should be noted, however, that the original LR images in the dataset are relatively large in resolution.
• Figure 2: Qualitative comparison of super-resolution results across multiple datasets. For each dataset, cropped image regions are arranged from left to right in approximate order of increasing perceptual quality, as indicated by the corresponding quantitative metrics. The corresponding datasets for each crop, from top to bottom, are: PSR4K, DIV2K-LSDIR validation set, RealSR validation set, PIPAL validation set, Real47, and RealSRSet. Images have been downscaled from their original upscaled resolution and compressed to meet compilation constraints, which may slightly affect their visual fidelity. Although the image displays 'ESRGAN', we are in fact referring to 'Real-ESRGAN'; the omission of 'Real' is purely for aesthetic purposes.
• Figure 3: An overview of the proposed SAFMN-L in \ref{['subsec:VPEG']} by the VPEG team. SAFMN employs a series of feature mixing modules (FMMs) to process deep-level features. The FMM block is composed of a spatially-adaptive feature modulation (SAFM) and a convolutional channel mixer (CCM).
• Figure 4: Overview of TinyESRGAN proposed in \ref{['subsec:MiAlgo']} by MiAlgo team.
• Figure 5: Overview of EFDN architecture presented by IPIU team in \ref{['subsec:IPIU']}.