SemanticStyleGAN: Learning Compositional Generative Priors for Controllable Image Synthesis and Editing

Recent studies have shown that StyleGANs provide promising prior models for\ndownstream tasks on image synthesis and editing. However, since the latent\ncodes of StyleGANs are designed to control global styles, it is hard to achieve\na fine-grained control over synthesized images. We present SemanticStyleGAN,\nwhere a generator is trained to model local semantic parts separately and\nsynthesizes images in a compositional way. The structure and texture of\ndifferent local parts are controlled by corresponding latent codes.\nExperimental results demonstrate that our model provides a strong\ndisentanglement between different spatial areas. When combined with editing\nmethods designed for StyleGANs, it can achieve a more fine-grained control to\nedit synthesized or real images. The model can also be extended to other\ndomains via transfer learning. Thus, as a generic prior model with built-in\ndisentanglement, it could facilitate the development of GAN-based applications\nand enable more potential downstream tasks.\n