FaceGANs: Stable Generative Adversarial Networks with High-Quality Images

Generative Adversarial Networks (GANs) have shown impressive performance in producing images highly similar to original dataset under unsupervised learning. However, the losses of discriminator and generator are highly fluctuated, which affects the quality of fake images produced by the generator. In this work, we propose Face Generative Adversarial Networks (FaceGANs). Compared to the conventional GANs, our new structure can stabilize the loss fluctuation of discriminator and generator. It also improves the capabilities of generator and discriminator. In order to fully investigate FaceGANs, we compare the performance of FaceGANs with Deep Convolution Generative Adversarial Network (DCGANs) on the Celeba dataset. Experimental results show that our FaceGANs structure can fast generate images with better quality than DCGANs in a facial reconstruction.