This paper proposes a framework called InterfaceGAN for semantic face editing by interpreting the latent semantics learned by GANs. The framework analyzes how different semantics are encoded in the latent space of GANs for face synthesis and shows that well-trained generative models learn a disentangled representation after linear transformations. The study explores the disentanglement between various semantics and manages to decouple some entangled semantics using subspace projection, enabling more precise control of facial attributes. The framework allows for manipulating gender, age, expression, and the presence of eyeglasses, as well as varying face pose and fixing artifacts generated by GANs. The proposed method is further applied to real image manipulation when combined with GAN inversion methods or encoder-involved models. Extensive results suggest that learning to synthesize faces spontaneously brings a disentangled and controllable facial attribute representation. The framework provides both theoretical analysis and experimental results to verify that linear subspaces align with different true-or-false semantics emerging in the latent space. The study also shows that the disentangled semantics enable precise control of facial attributes with any given GAN model without retraining. The contributions include proposing InterFaceGAN to explore how a single or multiple semantics are encoded in the latent space of GANs, showing that InterFaceGAN enables semantic face editing with any fixed pre-trained GAN model, and extending InterFaceGAN to real image editing with GAN inversion methods and encoder-involved models. The paper also discusses related work on GANs, latent space of GANs, and semantic face editing with GANs. The framework of InterfaceGAN is introduced, which first provides a rigorous analysis of the semantic attributes emerging in the latent space of well-trained GAN models and then constructs a manipulation pipeline of leveraging the semantics in the latent code for facial attribute editing. The paper presents experiments on PGGAN and StyleGAN, showing that the framework can be applied to real image editing. The results demonstrate that the framework can manipulate real faces with the proposed InterFaceGAN, verifying that the semantic attributes learned by GANs can be applied to real faces. The paper also provides a detailed proof of Property 2 in the main paper and discusses the theoretical foundations of the framework.This paper proposes a framework called InterfaceGAN for semantic face editing by interpreting the latent semantics learned by GANs. The framework analyzes how different semantics are encoded in the latent space of GANs for face synthesis and shows that well-trained generative models learn a disentangled representation after linear transformations. The study explores the disentanglement between various semantics and manages to decouple some entangled semantics using subspace projection, enabling more precise control of facial attributes. The framework allows for manipulating gender, age, expression, and the presence of eyeglasses, as well as varying face pose and fixing artifacts generated by GANs. The proposed method is further applied to real image manipulation when combined with GAN inversion methods or encoder-involved models. Extensive results suggest that learning to synthesize faces spontaneously brings a disentangled and controllable facial attribute representation. The framework provides both theoretical analysis and experimental results to verify that linear subspaces align with different true-or-false semantics emerging in the latent space. The study also shows that the disentangled semantics enable precise control of facial attributes with any given GAN model without retraining. The contributions include proposing InterFaceGAN to explore how a single or multiple semantics are encoded in the latent space of GANs, showing that InterFaceGAN enables semantic face editing with any fixed pre-trained GAN model, and extending InterFaceGAN to real image editing with GAN inversion methods and encoder-involved models. The paper also discusses related work on GANs, latent space of GANs, and semantic face editing with GANs. The framework of InterfaceGAN is introduced, which first provides a rigorous analysis of the semantic attributes emerging in the latent space of well-trained GAN models and then constructs a manipulation pipeline of leveraging the semantics in the latent code for facial attribute editing. The paper presents experiments on PGGAN and StyleGAN, showing that the framework can be applied to real image editing. The results demonstrate that the framework can manipulate real faces with the proposed InterFaceGAN, verifying that the semantic attributes learned by GANs can be applied to real faces. The paper also provides a detailed proof of Property 2 in the main paper and discusses the theoretical foundations of the framework.