Content-Aware Latent Semantic Direction Fusion for Multi-Attribute Editing

Xiwen Wei, Yihan Tang, Si Wu

Abstract: For facial attribute editing, significant progress has been made in discovering semantic directions in the latent space of StyleGAN, and the manipulation is performed by mapping an input image to a latent code and then moving along a direction associated with a target attribute. In this case, multi-attribute editing typically needs a sequential transformation process, which may cause ineffective manipulation or the cumulative effect on irrelevant attribute deviation. In this work, we aim to simultaneously manipulate multiple attributes through a single transformation. Toward this end, we propose a StyleGAN-based latent semantic direction fusion model, referred to as StyleLSF. There are two learnable components: a content-aware direction predictor learns to infer the latent directions, which are associated with preset attributes. A fusion network fuses the directions with respect to target attributes and yields a single translation vector. We further ensure irrelevant attribute preservation by imposing an attribute-aware feature consistency regularization approach.