[Paper Review] Generative Adversarial Networks: A Survey and Taxonomy.

Zhengwei Wang, Qi She+1 more|arXiv (Cornell University)|Jun 4, 2019

Generative Adversarial Networks and Image SynthesisComputer Science49 references67 citations

TL;DR

This paper presents a comprehensive survey and taxonomy of generative adversarial networks (GANs), focusing on their progress in addressing three core challenges in computer vision: high-quality image generation, diversity in generation, and training stability. It reviews popular GAN architectures and loss functions, offering critical analysis and future research directions based on empirical performance across key application domains.

ABSTRACT

Generative adversarial networks (GANs) have been extensively studied in the past few years. Arguably their most significant impact has been in the area of computer vision where great advances have been made in challenges such as plausible image generation, image-to-image translation, facial attribute manipulation and similar domains. Despite the significant successes achieved to date, applying GANs to real-world problems still poses significant challenges, three of which we focus on here. These are: (1) the generation of high quality images, (2) diversity of image generation, and (3) stable training. Focusing on the degree to which popular GAN technologies have made progress against these challenges, we provide a detailed review of the state of the art in GAN-related research in the published scientific literature. We further structure this review through a convenient taxonomy we have adopted based on variations in GAN architectures and loss functions. While several reviews for GANs have been presented to date, none have considered the status of this field based on their progress towards addressing practical challenges relevant to computer vision. Accordingly, we review and critically discuss the most popular architecture-variant, and loss-variant GANs, for tackling these challenges. Our objective is to provide an overview as well as a critical analysis of the status of GAN research in terms of relevant progress towards important computer vision application requirements. As we do this we also discuss the most compelling applications in computer vision in which GANs have demonstrated considerable success along with some suggestions for future research directions. Code related to GAN-variants studied in this work is summarized on this https URL.

Motivation & Objective

• To address the persistent challenges of generating high-quality, diverse, and stable GAN outputs in real-world computer vision applications.
• To critically evaluate the state of the art in GAN research by focusing on practical challenges rather than theoretical novelty alone.
• To provide a structured taxonomy based on architectural and loss function variations to organize and compare existing GAN variants.
• To identify the most successful GAN approaches in key computer vision tasks such as image generation, image-to-image translation, and facial attribute manipulation.
• To suggest future research directions based on gaps in current progress toward practical deployment of GANs.

Proposed method

• The paper conducts a systematic review of published scientific literature on GANs, emphasizing architectures and loss functions that target specific challenges.
• It organizes GAN variants into a taxonomy based on architectural design (e.g., conditional GANs, StyleGAN, BigGAN) and loss function modifications (e.g., adversarial loss, perceptual loss, cycle consistency).
• The analysis evaluates each GAN variant's performance on three core challenges: image quality, diversity, and training stability using qualitative and quantitative benchmarks.
• The review includes critical discussion of trade-offs between different GAN designs, such as mode collapse mitigation and fidelity improvements.
• It evaluates the impact of architectural innovations like skip connections, normalization layers, and progressive growing in improving training dynamics and output quality.
• The paper compiles and summarizes code repositories for studied GAN variants to support reproducibility and further research.

Experimental results

Research questions

• RQ1 To what extent have recent GAN architectures improved the quality of generated images in computer vision tasks?
• RQ2 How do different loss functions contribute to enhancing the diversity and avoiding mode collapse in GAN-generated outputs?
• RQ3 What architectural and training strategies have led to more stable GAN training, and how do they compare across benchmarks?
• RQ4 Which GAN variants have demonstrated the most significant success in applications such as image-to-image translation and facial attribute manipulation?
• RQ5 What are the key limitations and open challenges in deploying GANs for real-world computer vision applications?

Key findings

• Architectural innovations such as progressive growing and style-based normalization significantly improve training stability and image quality, as seen in models like StyleGAN and BigGAN.
• Loss function modifications—particularly the integration of perceptual and cycle-consistency losses—have enhanced diversity and reduced mode collapse in generated samples.
• Conditional GANs and their variants have shown strong performance in controlled image generation tasks, including facial attribute editing and image-to-image translation.
• Despite progress, training instability and mode collapse remain persistent challenges, especially in high-resolution image generation.
• The taxonomy presented in the paper effectively categorizes GAN variants by architecture and loss, enabling clearer comparison and identification of promising research directions.
• Code repositories for the reviewed GAN variants are compiled and publicly available, supporting reproducibility and future benchmarking.