Learning to draw vector graphics : applying generative modeling to font glyphs

Abstract

Today, designers work in tandem with computerized tools to create stylized graphic designs, diagrams, and icons. In this work, we explore the applications of generative modeling to the design of vectorized drawings, with a focus on font glyphs. We establish a data-driven approach for creating preliminary graphics upon which designers can iterate. To accomplish this, we present an end-to-end pipeline for a supervised training system on Scalable Vector Graphics (SVGs) that learns to reconstruct training data and produce similar but novel examples. We demonstrate its results on selected characters using a Google Fonts dataset of 2552 font faces. Our approach uses a variational autoencoder to learn sequences of SVG drawing commands and is capable of both recreating ground truth inputs and generating unseen, editable SVG outputs. To investigate improvements to model performance, we perform two experiments: one on the effects of various SVG feature encodings on generated outputs, and one on a modified architecture that explicitly encodes style and class separately for multi-class generation.