Generating Face Images Using VQGAN and Sparse Transformer

In this paper, we present a system for face image generation using VQGAN and sparse transformer. We explore the use of VQGAN models to learn visual tokens of image constituents and enhance the autoregressive priors to generate synthetic samples. We demonstrate that the routing transformer which learns sparse attention patterns over the visual tokens can generate samples with high-quality on face image datasets such as FFHQ and CelebA-HQ, while not suffering from mode collapse and lack of diversity.