Estimating the Training Performance of Generative Adversarial Networks by Image Quality

Generative adversarial networks (GANs) can address the problem when the number of training samples is insufficient for many fields of AI models. However, we observed that in Generative Adversarial Networks, the loss cannot directly indicate the quality of generated images. A low loss does not necessarily guarantee good image quality, and a high loss does not necessarily imply poor image quality. This paper discusses the impact of various hyperparameters on the quality of generated images. We also attempt to use a small dataset to create synthetic images of a certain quality using the proposed evaluation method in GANs.