Effects of Different Generative Adversarial Networks on the Face Generation Task

The human face may be the bodily part with the greatest degree of individuality, and it may provide a lot of information. On account of its importance in a wide range of applications, human face has become a popular study topic in the domain of computer vision. Besides the well-known face detection and face recognition task, human face generation technology is also widely concerned because of its various potential applications, including privacy protection, scientific use, entertainment purposes and so on. Generative adversarial network has been a powerful tool to generate human face since its first proposition in 2014 and numerous variants based on GAN were proposed in recent years to improve its stability and capability. With the advancements in generative methods, face generation technology has seen rapid development. In this paper, the author use three representative GAN variants, DCGAN, WGAN-GP and PGGAN to generate human face on the dataset of CelebA-HQ 128x128 and analyze their actual performance. The generated image resembles real human being, however, doesn't reach the level of which the original paper demonstrated because of insufficient computing power and limited running time. In the end, the author pointed out the potential ethical problems of face generation technology and made a prospect for the future of human face generation.