Generative font shape modeling
Year: 2025
Type: dissertation
Abstract: This thesis addresses the problem of generating high-quality vector font glyphs in SVG format using deep generative models. The ultimate goal is to develop a system capable of synthesizing unseen letters in a consistent font style, giving only a subset of example glyphs from the same typeface. To approach this challenge, we first explore a supervised transformation model that converts regular font glyphs into their bold counterparts using a Transformer architecture. This preliminary step highlights the challenges of sequence-based modeling of SVG data and motivates the need for a more flexible generative approach. We then build upon the DeepSVG architecture to design a hierarchical Variational AutoEncoder (VAE) capable of modeling complex distributions over SVG glyphs. Our model processes each path within a glyph independently through a Transformer-based encoder and subsequently captures higher-level relationships between paths at the glyph level via a second encoder stage. A corresponding hierarchical decoder reconstructs full glyphs while predicting visibility and continuity information for improved path alignment. To further enhance output quality, we integrate a self-refinement module that adjusts control points to enforce predicted geometric continuity constraints. Finally, we extend this framework into a generative model of font styles. By encoding a subset of letters from a given font and aggregating their latent representations, we define a style embedding from which new, stylistically coherent glyphs can be generated.
Source: Hispana
Author: Guija Valiente, Tomás
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7113523955