Using Fully Connected and Convolutional Net for GAN-Based Face Swapping

Abstract

The lifelike results of using face swapping have contributed greatly to the research in computer vision. In this work, we extend the architecture of faceswap-GAN in order to obtain more natural results compared to the original framework. In the original architecture, the self-attention module usually converts the facial features from a source face to the target face with artificial distortion around the facial features. We use a structure of fully connected convolutional layers as a discriminator to approach the problem. The outcome can be smoother and more natural perceptually compared to the results using the original faceswap-GAN.