Generative Models Unveiled: Evaluating VAEs and GANs for Image Creation

This paper presents a comparative analysis of four models applied to an anime face dataset which contains 63,632 images belonging to a single class(face). The models evaluated include Generative Adversarial Networks (GAN) without CNN, Deep Convolutional GAN (DC-GAN), and Variational Autoencoders (VAE) and VAE without CNN. GAN models are the traditional generative models. They use adversarial training to generate new realistic data which is similar to input data. On the other hand, VAEs take a probabilistic approach by learning a latent space representation of input data to generate new data. Each model was trained on the anime face dataset. The performance was compared based on the quality of generated images, and analyzing the training loss graphs for all the models. Also, the quality of the generated is analyzed manually. The results demonstrate that the DC-GAN outperforms the standard GAN in terms of image quality and stability, while even if the quality of images generated by VAE with and without CNN appeared to be similar the VAE with CNN was found to be more stable by observing the training-loss graph. This study highlights the importance of convolutional layers in improving both GAN and VAE performance in image generation tasks, providing insights for selecting appropriate models for high-quality anime face generation.