Workflow from Scientific Research

Overview of MaskFaceGAN, illustrated with the Black hair target attribute. To initialize the editing procedure, MaskFaceGAN uses a face parser to define masks that correspond to image regions that should be preserved (Sskin) and regions that should be altered (Star). Next, latent code optimization is performed in accordance with semantic and spatial constraints to generate an intermediate image IG with the targeted characteristics. Finally, blending is used to combine the generated image IG with the original I and to produce the final editing result I'. The image is best viewed electronically.

DOI