Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GANs Conditioning Methods: A Survey

Anis Bourou, Valérie Mezger, Auguste Genovesio

TL;DR

This survey analyzes how conditioning can be incorporated into GANs, focusing on discriminators (auxiliary classifiers, projection-based, and contrastive discriminators) and generator conditioning techniques. It introduces ECGAN, a unifying energy-based framework that connects classifier-based and classifier-free approaches, and discusses how different formulations approximate the joint distribution p(x,y). Through experiments on CIFAR-10 and a Carnivores subset, the study highlights trade-offs in stability, diversity, and quality across methods, showing that projection- and contrastive-based conditioning often offer robustness where auxiliary classifiers struggle. The findings provide practical guidance for selecting conditioning strategies and motivate future research toward unified, stable, and scalable conditional generative models.

Abstract

In recent years, Generative Adversarial Networks (GANs) have seen significant advancements, leading to their widespread adoption across various fields. The original GAN architecture enables the generation of images without any specific control over the content, making it an unconditional generation process. However, many practical applications require precise control over the generated output, which has led to the development of conditional GANs (cGANs) that incorporate explicit conditioning to guide the generation process. cGANs extend the original framework by incorporating additional information (conditions), enabling the generation of samples that adhere to that specific criteria. Various conditioning methods have been proposed, each differing in how they integrate the conditioning information into both the generator and the discriminator networks. In this work, we review the conditioning methods proposed for GANs, exploring the characteristics of each method and highlighting their unique mechanisms and theoretical foundations. Furthermore, we conduct a comparative analysis of these methods, evaluating their performance on various image datasets. Through these analyses, we aim to provide insights into the strengths and limitations of various conditioning techniques, guiding future research and application in generative modeling.

GANs Conditioning Methods: A Survey

TL;DR

This survey analyzes how conditioning can be incorporated into GANs, focusing on discriminators (auxiliary classifiers, projection-based, and contrastive discriminators) and generator conditioning techniques. It introduces ECGAN, a unifying energy-based framework that connects classifier-based and classifier-free approaches, and discusses how different formulations approximate the joint distribution p(x,y). Through experiments on CIFAR-10 and a Carnivores subset, the study highlights trade-offs in stability, diversity, and quality across methods, showing that projection- and contrastive-based conditioning often offer robustness where auxiliary classifiers struggle. The findings provide practical guidance for selecting conditioning strategies and motivate future research toward unified, stable, and scalable conditional generative models.

Abstract

In recent years, Generative Adversarial Networks (GANs) have seen significant advancements, leading to their widespread adoption across various fields. The original GAN architecture enables the generation of images without any specific control over the content, making it an unconditional generation process. However, many practical applications require precise control over the generated output, which has led to the development of conditional GANs (cGANs) that incorporate explicit conditioning to guide the generation process. cGANs extend the original framework by incorporating additional information (conditions), enabling the generation of samples that adhere to that specific criteria. Various conditioning methods have been proposed, each differing in how they integrate the conditioning information into both the generator and the discriminator networks. In this work, we review the conditioning methods proposed for GANs, exploring the characteristics of each method and highlighting their unique mechanisms and theoretical foundations. Furthermore, we conduct a comparative analysis of these methods, evaluating their performance on various image datasets. Through these analyses, we aim to provide insights into the strengths and limitations of various conditioning techniques, guiding future research and application in generative modeling.
Paper Structure (21 sections, 37 equations, 7 figures, 4 tables)

This paper contains 21 sections, 37 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Discriminator conditioning approaches
  3. Auxiliary-classifier based discriminators
  4. Auxiliary classifier GAN (AC-GAN)
  5. Twin Auxiliary Classifier GAN (TAC-GAN)
  6. Unbiased Auxiliary Classifier GAN (UAC-GAN)
  7. Auxiliary Discriminative Classifier GAN (ADC-GAN)
  8. Projection-based discriminators
  9. Contrastive learning based discriminators
  10. Contrastive learning GAN (ContraGAN)
  11. Rebooting Auxiliary Classifier GAN (ReACGAN)
  12. Towards a unified framework for conditioning the discriminator
  13. Approaching the joint distribution from the conditional discriminator perspective
  14. Approaching the joint distribution from the classifier perspective
  15. Unifying cGANs discriminators
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Conditioning by label embedding concatenation at different levels of the discriminator, (a) was proposed in mirza_conditional_gan and (b) in reed_text_to_image
  • Figure 2: Auxiliary classifier based architectures: (a) AC-GAN, (b) TAC-GAN, (c) UAC-GAN, (d) ADC-GAN
  • Figure 3: The architecture of the projection-based discriminator as proposed in projectioncgans.
  • Figure 4: Contrastive learning based architectures. On the left: the discriminator architecture of ContraGAN. On the right: the Reboot AC-GAN architecture.
  • Figure 5: Training StyleGAN2 on AFHQ and a subset of ImageNet with and without conditioning
  • ...and 2 more figures