Attribute Space Analysis for Image Editing.

Image editing is a widely studied topic in computer vision, which enables the modification of specific attributes in images without altering other crucial information. One popular unsupervised technique currently used is feature decomposition in the latent space of Generative Adversarial Networks (GANs), which provides editing directions that can control attribute changes to achieve desired image editing results. However, this method often does not allow for the direct acquisition of the desired editing direction by setting the target attribute in advance. In this work, we propose a method to finding editing directions in the attribute space by analyzing image differences. This enables users to obtain target directions by actively defining the attribute they want to change. Specifically, this method discovers semantic directions suitable for target attribute editing by applying Principal Component Analysis (PCA) on the difference of image latent codes embedded in the latent space. Through experiments, our method can effectively find the target editing direction according to user needs and achieve satisfactory editing effects at the same time.