Generating human face by generative adversarial networks

Abstract

The Generative Adversarial Network (GAN) method is widely used for image generation, especially human face generation and modification. The recent GANs model can generate diverse photorealistic images and this relies on the powerful capabilities of semantic latent representation and feature disentanglement. Those capabilities are also fully demonstrated in this project. This project focuses on utilising GAN for generating human face images to train a visual alignment model, GANgealing, and utilising GAN for modifying human facial attributes on images that have been aligned by GANgealing. To achieve this, I first study the latest facial image generation methods, the StyleGAN series, which possess the capabilities of semantic latent representation and feature disentanglement. After that, I examine a GAN-supervised visual alignment method called GANgealing, which relies on StyleGAN2's feature disentanglement ability to generate the training data and the corresponding labels, in which human facial features are the same but the poses are diverse. The GANgealing aligns facial features to a unique central mode. With this alignment, a simple GAN approach, AttGAN, can semantically modify the facial attributes of face images. Undoubtedly, the editing must be correctly transferred to the original images. After all the models were trained, I integrated AttGAN into the GANgealing algorithm, allowing for alignment, editing, and editing propagation to be performed in a single application. Lastly, I investigated the performance of this facial attribute editing pipeline.