SAC-GAN: Face Image Inpainting with Spatial-Aware Attribute Controllable GAN

The objective of image inpainting is refilling the masked area with semantically appropriate pixels and producing visually realistic images as an output. After the introduction of generative adversarial networks (GAN), many inpainting approaches are showing promising development. Several attempts have been recently made to control reconstructed output with the desired attribute on face images using exemplar images and style vectors. Nevertheless, conventional style vector has the limitation that to project style attribute representation onto linear vector without preserving dimensional information. We introduce spatial-aware attribute controllable GAN (SAC-GAN) for face image inpainting, which is effective for reconstructing masked images with desired controllable facial attributes with advantage of utilizing style tensors as spatial forms. Various experiments to control over facial characteristics demonstrate the superiority of our method compared with previous image inpainting methods.