Novel GAN Inversion Model with Latent Space Constraints for Face Reconstruction

This paper considers how to encode a target face image into its StyleGAN latent space accurately and efficiently, with applications to allow the various image editing method being used on the real images. Compared with optimization-based methods using gradient descent on latent code iteratively, the learning-based method we adopt can encode target images with one forward propagation, which is better suited for real-world application. The key advances in this paper are: adopting the face recognition model as a constraint to keep the identity information intact and adding a classifier to encourage latent code to retain more attributes possessed in the original image. Experiments show our method can achieve an excellent reconstruction effect. The ablation study indicates the proposed design advances the GAN Inversion task qualitatively and quantitatively. However, the method may fail when there are other objects around the target face and generate a blurry patch around that object.