Improved Semantic-aware StyleGAN-based Real Face Editing Model
Fangxin Wang, Xianliang Wang, Fanjun Meng · 2021
The quality of the face editing is greatly affected by the latent codes. As far as to the most popular StyleGAN series models, many researches have shown that the W space is better for reconstruction quality, while the W+ space is better for editing quality. Since editing real faces, instead of those fake faces generated by GAN model, involves both face reconstruction and face editing, so most of the previous approaches have to make a trade-off between reconstruction and editing quality. In this paper, we attribute this problem to the mismatch of the latent space in training and inference, i.e., training in W space but inference in W+ space, which leads to the semantic loss. To this end, we propose a semantic-aware StyleGAN-based real face editing model. In this model, training and inference are performed in the similar space by connecting mapping layers and synthesis layers directly. Besides, the non-linear mapping between pixel space and latent space, also leads to the semantic loss as no semantic relations are guaranteed. To solve this problem, we introduce a manifold regularization term to keep the manifold during StyleGAN training. We conduct extensive experiments of face reconstruction and face editing, and the results suggest that our method significantly outperforms the previous models.