Semantic Images Editing by Operations on Latent Space of Deep Generative Models

抄録

The creation of images and other data is one of the ultimate goals of computer vision research. For this purpose, various deep learning methods have been proposed, such as variational autoencoders, adversarial networks, and diffusion models. These methods learn the distributions of photographs and illustrations and reproduce them. The generated image is determined using the coordinates provided in the latent space. Therefore, several studies have been conducted to manipulate these coordinates to edit the generated images. However, existing methods frequently provide unintended or low-quality editing results because the coordinate system in the latent space is not properly learned, among other reasons. In this study, we focus on the coordinate system in the representation space and introduce deep curvilinear editing. In particular, we propose a method for the representation vectors using representation space with a curvilinear coordinate system. The method was also combined with generative adversarial networks, whose results demonstrated that the proposed method enables the high-quality editing of generated images.