Abstract
Disentanglement research is a critical and important issue in the field of image editing. In order to perform disentangled editing on images generated by generative models, this paper presents an unsupervised, model-agnostic, two-stage trained editing framework. This work addresses the problem of discovering interpretable, disentangled directions of edited image attributes in the latent space of generative models. This effort's primary objective was to address the limitations discovered in previous research, mainly (a) the discovered editing directions are interpretable but significantly entangled, i.e., changes to one attribute affect the others and (b) Prior research has utilized direction discovery and direction disentanglement separately, and they can't work synergistically. More specifically, this paper proposes a two-stage training method that discovers the editing direction with semantics, perturbs the dimension of the direction vector, adjusts it with a penalty mechanism, and makes the editing direction more disentangled. This allows easy distinguishable image editing, such as age and facial expressions in facial images. Experimentally compared to other methods, the proposed method outperforms them both qualitatively and quantitatively in terms of interpretability, disentanglement, and distinguishability of the generated images. The implementation of our method is available at https://github.com/ydniuyongjie/twoStageForFaceEdit.