| Summary: | Since generative adversarial network (GAN) was first proposed, the processing of face images, especially the research of facial attribute editing, has attracted much interest. It not only can alleviate the problems associated with data deficiency, but also has great applications in the field of entertainment. However, existing approaches have limited scalability in the processing of newly-added face attributes, and the quality of generated images is poor. To solve these problems, FEGAN is proposed in this paper to achieve the accurate editing of multi-attribute faces by modifying feature vectors in the latent space. Firstly, a trained generator is used, which greatly reduces the training difficulty of GANs, and the inverse of the generator is used to establish the unique correspondence between the input image and the latent code. Secondly, a linear guide is applied to the latent code, and thus the same distribution as the target image in the latent space is assured. Finally, a generator is used to generate a face image from the guided latent code. The proposed method is utilized for a large number of attribute editing experiments, and the results show that FEGAN can flexibly perform accurate attribute editing while guaranteeing that other areas are not changed. Both qualitative and quantitative results demonstrate its advantages over existing methods.
|
|---|
