[Paper Review] How Generative Adversarial Networks and its variants Work: An Overview of GAN

Yongjun Hong, Uiwon Hwang+2 more|arXiv (Cornell University)|Nov 16, 2017

Generative Adversarial Networks and Image Synthesis14 references22 citations

TL;DR

This paper provides a comprehensive overview of Generative Adversarial Networks (GANs) and their variants, explaining their architecture, training dynamics, and objective functions. It details how GANs learn complex data distributions without strong assumptions and explores their integration with autoencoders to enable controllable generation in latent space, supporting applications in image synthesis, editing, and translation.

ABSTRACT

Generative adversarial networks (GANs) have received wide attention in the machine learning field because of their potential to learn high-dimensional, complex real data. Specifically, they do not perform distribution assumptions and can simply infer real-like samples from latent space. This powerful property leads GANs to be applied to various applications such as image synthesis, image attribute editing, image translation, domain adaptation and other academic fields. In this review, we aim to discuss details of GANs for those readers who are familiar but do not comprehend GANs deeply, or who wish to evaluate GANs from various perspectives. We discuss how a GAN operates and the fundamental meaning of various objective functions suggested recently. We then focus on how the GAN can be combined with an auto-encoder framework which makes it possible handle the latent space. As an extension, we also discuss GAN variants that are applied to various tasks and other fields.

Motivation & Objective

• To clarify the fundamental mechanisms of GANs for readers with basic familiarity but limited in-depth understanding.
• To analyze the design and significance of various objective functions used in GAN training.
• To investigate the integration of GANs with autoencoder frameworks for improved control over latent space representations.
• To survey GAN variants and their applications across diverse academic and practical domains.
• To provide a unified perspective on GANs for researchers evaluating or applying them in new contexts.

Proposed method

• The paper explains the adversarial training framework where a generator and discriminator are trained in a minimax game to produce realistic samples.
• It analyzes key objective functions such as the original GAN loss, WGAN, and GAN variants that improve training stability and sample quality.
• The integration of GANs with autoencoders is examined, enabling disentangled and controllable generation through learned latent codes.
• The paper discusses architectural modifications in GAN variants like conditional GANs, cycle-GANs, and StarGANs for tasks such as image translation and attribute editing.
• It reviews training techniques such as gradient penalties and spectral normalization that enhance GAN convergence and performance.
• The analysis includes theoretical and empirical insights into how latent space disentanglement improves generation quality and interpretability.

Experimental results

Research questions

• RQ1 How do GANs learn complex, high-dimensional data distributions without strong distributional assumptions?
• RQ2 What is the role and impact of different objective functions in GAN training and sample quality?
• RQ3 How can GANs be effectively combined with autoencoder architectures to enable meaningful manipulation of latent representations?
• RQ4 What are the key architectural and training innovations in GAN variants that enable applications like image translation and domain adaptation?
• RQ5 How do GANs compare in performance and stability across different objective functions and training strategies?

Key findings

• GANs can generate high-fidelity, realistic samples from random noise in latent space without requiring explicit distributional assumptions.
• The introduction of alternative objective functions such as the Wasserstein GAN loss significantly improves training stability and sample quality.
• Combining GANs with autoencoders enables disentangled and interpretable generation, allowing for precise control over attributes in generated images.
• GAN variants like conditional GANs and cycle-GANs successfully enable image-to-image translation and attribute editing with high fidelity.
• Techniques such as gradient penalty and spectral normalization are critical for stabilizing GAN training and preventing mode collapse.
• The integration of GANs with autoencoders enables more structured and controllable generation, enhancing their utility in applications like domain adaptation and data augmentation.