Training Conditional GANs on Limited and Long-Tailed Data: a Survey and Comparative Analysis

TMLR Paper7481 Authors

Abstract: Generative adversarial networks (GANs) are a generative model framework that is competitive with state-of-the-art autoencoders and diffusion models in many tasks. While the latter have achieved impressive generation capabilities, mostly through large-scale, general-purpose text-to-image models, their computational requirements place them out of reach for practitioners. On the other hand, as GAN architectures mature and new developments allow for more stable training, interest in their application has grown across diverse domains. However, real-world data are often hard to deal with due to limited amount of samples or long-tailed distributions. Furthermore, previous works addressing these issues lack guidance regarding their applicability and have not been compared through appropriately diverse benchmarks nor assessed using the same metrics. In this article, we conduct a survey on methods for training GANs on limited and long-tailed data and conduct an extensive comparative analysis of existing methods. Our results allow us to draw conclusions about the advantages, disadvantages, and practical applicability of these methods, hopefully making GANs more accessible to practitioners in diverse fields. The code will be made available as soon as deanonymization is allowed.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Jakub_Mikolaj_Tomczak1

Submission Number: 7481