Maskfacegan: Masked GAN Latent Code Optimization for High-Resolution Face Editing

1 Chimata Ravya, 2 ⁠⁠Gangineni Lakshmi Naga Chandrika 3 Ajmira Prem Sagar, 4  Nimmagadda Chandra Sekhar

1,2,3 B. Tech Students, Department of CSE, RVR & JC College of Engineering, Chowdavaram, Guntur, A.P, India.

4Assistant Professor, Department of CSE, RVR & JC College of Engineering, Chowdavaram, Guntur, A.P, India,

E-mail:  y22cs026@rvrjc.ac.in ,     Y22cs049@rvrjc.ac.in ,  y22cs005@rvrjc.ac.in , nimmagadda65@gmail.com

Abstract:Face editing is a key research area in computer vision, but existing methods often suffer from low-resolution outputs, visual artifacts, and limited control over specific facial attributes, leading to unintended changes. This paper introduces MaskFaceGAN, a novel approach for precise local facial attribute editing. It optimizes the latent code of a pre-trained StyleGAN2 model using multiple constraints to preserve image content, accurately generate desired attributes, and restrict modifications to selected regions. These constraints are guided by a differentiable attribute classifier and face parser. Experimental results on FRGC, SiblingsDB-HQf, and XM2VTS datasets demonstrate that MaskFaceGAN achieves high-quality, high-resolution (1024 × 1024) edits with improved control and reduced attribute entanglement compared to existing methods. Index Terms— Facial attribute editing, generative adversarial networks, GAN inversion, latent code optimization.