Accurate face swap using cycleGAN

Abstract

Face swap is to transfer a face image of one person to another given person. It can generate false but very realistic images and video data. At present, numerous frameworks have been proposed, such as CNN, FCN, GAN. Recently, GAN-based methods are popular with satisfying performance. In this paper, we aim to improve the GAN-based face swap methods under the cycleGAN framework and achieve a more realistic face-swapping result. In this method, cycleGAN is applied to to face swap task, which mainly solves the problems of unavailable paired training images, and our model can be trained without pairing real images. More than 1,000 pictures in total have participated in the model training. The test needs to input a video and output a video of the facial changes in a few minutes. We collect Hillary and Trump videos data on the google website for testing and analysis, obtain a trained model, and accurately change the face of the video. Our method can handle face swap effectively with reasonable speed. The quality and speed of its generation are no less than that of other frameworks like GAN, which is more popular nowadays.