Generating and Editing Arbitrary Facial Images by Learning Feature Axis

Abstract

There are mainly three limitations of the traditional facial attribute editing techniques: 1) incapability of generating an arbitrary facial image with high-resolution; 2) being unable to generate and edit new facial images synthesized by the computer and 3) limited diversity of edited images. This paper presents a method for generating and editing images simultaneously. It incorporates a high-resolution facial image generator, a multi-label classifier, and a Generalized Linear Model (GLM). Experimental results show that our method can generate arbitrary high-resolution facial images, edit computer-synthesized images, perform multi-attribute editing, and effectively control the intensity and style of the generated images. Besides, the approach has high efficiency and flexibility, allowing rapid migration of attribute information from the data set. We design a graphical interface program, which can be integrated as a mobile application.

A comprehensive survey on semantic facial attribute editing using generative adversarial networks

“…Finally, for each attribute, a direction vector orthogonal to the decision boundary is obtained that can be used for changing the value of that attribute. Yang et al [189] proposed another model for face editing based on the direction vector in the latent space. In the training phase of this model, first, the mapping network f generates the disentangled latent vector u from the noise latent vector z.…”