Objective quality assessment of generated high-resolution human face images

Human face image generation can be used in media, marketing, education, creativity, research and privacy protection. Many deep learning networks have been proposed for generating high-resolution human face images, while the quality of generated human face images are not fully investigated. To bridge this gap, 11 StyleGAN-based networks are firstly trained via the database Flickr-Faces-HQ (FFHQ). Then, each network produces 10,000 high-resolution (1024 × 1024) human face images. After that, the quality of these generated images are evaluated with 6 Face Image Quality Assessment (FIQA) metrics. The average metric values suggest that the generated human face images are with higher quality than the realistic images in the FFHQ database. However, the failure cases of generated images remind us that artifacts and distortions exist, and unreasonable combinations and meaningless contents are produced. Therefore, significant progress is still required for achieving reliable quality assessment and high-accuracy generation of human faces.