Wang and Lian (2021) | DeepVecFont: Synthesizing High-quality Vector Fonts via Dual-modality Learning

Wang and Lian specifically address the problem of generating vector glyphs obtained from fonts. The authors build upon the work from Carlier et al. (2020), Ha and Eck (2017), and Lopes et al. (2019). The authors aim to tackle the problem that previous works were not able to synthesize visually pleasing vector glyphs and name two reasons for this issue:

  1. Single modality – they propose to use a dual modality with vector data representation and raster image data representation

  2. Location shift issue brought by the Mixture Distribution Network – they propose to employ a differentiable rasterizer (developed in the meantime by Li et al. (2020)) for imposing an additional restriction on the drawing commands predicted by the MDN

This paper was submitted to SIGGRAPH Asia 2021 and published in ACM Transactions on Graphics.

Wand and Lian 2021 paper

Fig. 54 Screenshot of the DeepVecFont paper by Wang and Lian (2021)

Data representation

Like most other papers, Wand and Lian only consider 4 SVG commands.

  1. Move – moving the drawing location (for starting a new path)

  2. Line – drawing a line

  3. Curve – Cubic Bézier Curve

  4. End – ending the draw-command sequence