Advances in fidelity-preserving GANs for small datasets: Focus on StyleGAN-ADA and its variants

Artificial Intelligence, Faculty of Science and Technology, Beijing Normal-Hong Kong Baptist University, Zhuhai, Guangdong, China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

It seems challenging to train a generative adversarial network with limited data as issues like mode collapse and discriminator overfitting occurs. In this way, the appearance of StyleGAN2-ADA had solved the problems with adaptive discriminator augmentation technique. It works by dynamically adjusting data augmentation to prevent overfitting. This review analyses StyleGAN2-ADA as well as its variants including StyleGAN3 and ViT-StyleGAN2-ADA. Furthermore, it compares models based on CIDAR- 10 dataset. The result demonstrates that the performance of StyleGAN2- ADA is better than StyleGAN2. Vit-StyleGAN2-ADA further enhances the performance by combining vision transformer into discriminator. Even though, StyleGAN3 performs well on high-resolution images, it may be not suitable for training on limited data. This review also considers some challenges such as high computational costs and limitations of evaluation metrics. Eventually, this review recommends to use StyleGAN2-ADA for cases with small dataset. StyleGAN3 is suitable for artistic image generation while ViT-StyleGAN2-ADA is suitable for natural image synthesis. Future work ought to concentrate on finding more efficient architectures and developing evaluation metrics for limited data.