Open Access

REVIEW

Generative Adversarial Networks for Image Super-Resolution: A Survey

Ziang Wu¹, Xuanyu Zhang², Yinbo Yu³, Qi Zhu³, Jerry Chun-Wei Lin⁴, Chunwei Tian^5,*

1 School of Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, China
2 School of Software, Northwestern Polytechnical University, Xi’an, China
3 College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Nanjing, China
4 Department of Distributed Systems and IT Devices, Silesian University of Technology, Gliwice, Poland
5 School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China

* Corresponding Author: Chunwei Tian. Email:

Computers, Materials & Continua 2026, 87(3), 3 https://doi.org/10.32604/cmc.2026.078842

Received 09 January 2026; Accepted 05 March 2026; Issue published 09 April 2026

Abstract

Image super-resolution is a significant area in the field of image processing, with broad applications across multiple domains. In recent years, advancements in Generative Adversarial Networks (GANs) have led to an increased adoption of GAN-based methods in image super-resolution, yielding remarkable results. However, there is still a limited amount of research that systematically and comprehensively summarizes the various GAN-based techniques for image super-resolution. This paper provides a comparative study that elucidates the application differences of GANs in this field. We begin by reviewing the development of GANs and introducing their popular variants used in image applications. Subsequently, we systematically analyze the theoretical motivations, implementation approaches, and technical distinctions of GAN-based optimization methods and discriminative learning from three perspectives: supervised, semi-supervised, and unsupervised learning. We examine these methods concerning their integration of different network architectures, prior knowledge, loss functions, and multitask strategies. Furthermore, we conduct a systematic comparison of state-of-the-art GAN methods through quantitative and qualitative analyses using publicly available super-resolution datasets. In addition to traditional metrics such as PSNR and SSIM in our quantitative analysis, we also consider complexity and running time as reference standards to better align the evaluation with practical application demands. Finally, we identify several challenges currently faced by GANs in the domain of image super-resolution, including issues related to training stability and the need for improved evaluation metrics. We outline future research directions aimed at enhancing the robustness and efficiency of GAN-based super-resolution techniques, emphasizing the importance of integration with other machine learning frameworks to further advance this exciting field.

---

Keywords

---