Currently, most popular generative models mainly create high-quality images that are very similar to the original human face. However, few people have studied how to obtain more face images based on exist- ing face datasets, which is a more significant direction. By generating diverse realistic face data images, on the one hand, it can effectively expand the data set because face images are more and more difficult to obtain. On the other hand, it could effectively protect the privacy of the original face. Most of the existing generative models start training from random noise. Although high-definition images can be obtained in the early stage, the diversity of generated faces will be greatly reduced with the progress of training. In this thesis, We propose a simple yet powerful model that can generate from picture to picture. We incorporate ideas from the Vector Quantised-Variational AutoEncoder (VQ-VAE). Since the element-wise L2 norm could not fully reflect the human perception loss. We provide a more advance discriminator network to optimize our model in an adversarial minmax manner. Moreover, for generating different faces, we extract the identity feature with the pre-trained Resnet model from FaceNet, and calculate their embedding feature distance between the original and reconstructed images measured by L2 norm. The results measured by the Fid score have further proved that our model is effective in producing same/various images while maintaining the relatively high quality.