Abstract
Scalable Vector Graphics (SVG) are ubiquitous in modern 2D interfaces due totheir ability to scale to different resolutions. However, despite the successof deep learning-based models applied to rasterized images, the problem ofvector graphics representation learning and generation remains largelyunexplored. In this work, we propose a novel hierarchical generative network,called DeepSVG, for complex SVG icons generation and interpolation. Ourarchitecture effectively disentangles high-level shapes from the low-levelcommands that encode the shape itself. The network directly predicts a set ofshapes in a non-autoregressive fashion. We introduce the task of complex SVGicons generation by releasing a new large-scale dataset along with anopen-source library for SVG manipulation. We demonstrate that our networklearns to accurately reconstruct diverse vector graphics, and can serve as apowerful animation tool by performing interpolations and other latent spaceoperations. Our code is available at https://github.com/alexandre01/deepsvg.