Exploiting Language Models for Vector-Style Images Synthesis

Abstract

Deep learning generative models have been successfully applied to synthesize images from various sources, like human faces and natural images, with impressive and realistic results. Nonetheless, not much work has been done for generating icons and vector-style images since synthesizing them requires precision and high-frequency details. Such images are essential for modern software and web development since they communicate concepts faster and more universally. We try to fill the gap by proposing an explicit density conditional generative model that can yield high-resolution samples when trained on rasterized vector-style images. Our novel architecture can solve conditional and unconditional image generation tasks, and it is easier to train than current adversarial approaches. Moreover, we compare our work with the current state-of-the-art generative models, highlighting their strengths and weakness. Finally, we introduce a new dataset containing high-quality icons from 11 different styles to test the quality of our model when performing conditional random sampling and style transfer between icons.