Stable Conditional Image Generation with WGAN-GP and DCGAN Architectures

Generative Adversarial Networks (GANs) have made great strides in image generation, but output control is an issue with vanilla GANs. Conditional GANs (cGANs) solve this by enabling input-directed generation, but, training stability is a pervasive problem in GANs. This paper explores the combination of Conditional GAN concepts with the added stability of Wasserstein Loss with Gradient Penalty (WGAN-GP). We use a deep-convolutional GAN (DCGAN) architecture for both the generator and the discriminator with Layer Normalization and the WGAN-GP loss function. The model is tested on the CIFAR-10 dataset. Our results show successful condition image generation, generating unique images for class labels. We examine the training dynamics, such as loss convergence, and discuss findings from hyperparameter tuning, providing information on utilizing this combined method for stable conditional image synthesis.