Minimising Distortion for GAN-Based Facial Attribute Manipulation

Facial Attribute Manipulation (FAM) through GAN-based methods has been an active topic in computer graphics. Existing works show high editing fidelity on randomly generated faces but suffer from distortion on embedded real faces. We alleviate this issue by dividing it into two sub-problems. First, we minimize embedding distortion by introducing a pre-trained Salient Object Detection (SOD) network. Second, we propose a nonlinear transformation network to minimize editing distortion. As a result, our framework, Character Centered Facial Attribute Manipulation (CCFAM), exhibits more disentangled edits on real faces. Moreover, CCFAM is computationally efficient by integrating image complexity into our embedding process. Evaluations demonstrate that our method performs better than the state-of-the-art in terms of both accuracy and fidelity.