Research on the Application of Jigsaw Generative Adversarial Network to Face Generation

In recent years, the development and research of GAN[1] is nothing more than a change in the number or structure of the generated network. The method of discriminate network is only to map different regional features to different discriminators and integrate them. As we all know, the quality of generator can affect the effectiveness of the entire network directly, but I think the function of the discriminator is the important pillar to support the entire network. So the author proposed a face-generation network named Jigsaw Generative Adversarial Network, which uses a single generator network with a discriminator as a layer. Each layer creates different parts to enhance the authenticity of each part, and then the face is composed in a jigsaw way, and is matched with the style module of Style GAN[2] to achieve more highly build quality. The proposed system can select the desired style for data set generation. For example, you can select the adult male style to generate a data set of all adult men and use this data set for research, or use the extracted style modules to apply to other generation networks and control the direction of the generated graph of the generated network.