Enhancing WGAN performance by architectural and optimizer variations for image generation

Generative model has opened up the area of image generation and has become a hot topic in recent years. Among the most famous generative models, Generative Adversarial Network (GAN) is outstanding among them, offering extensive avenues for exploration. The Wasserstein GAN (WGAN), as one of the GANs, introduces an innovative framework for training GANs based on the Earth Movers (Wasserstein) distance, providing a steadier training process. The experiment tried various modifications to WGAN, including changing the optimizers and the network architecture. Specifically, this work tried replacing the original Root Mean Square Prop (RMSprop) with another optimizers. Also, this work tried to add residual blocks to the network structure. These modifications provided interesting results, providing supplementary validation of the original WGAN structure, and providing some possibilities of optimization. According to the results, it could be found that the results of some modifications are very positive. However, some of the changes presented very unsatisfactory results, which gave us some insight.