3.4. Discussion: Interlude to the Practicalities

While the Preliminaries chapter mapped the landscape of AI font generation, this chapter, Surveys, transformed that knowledge into practical insights by examining three critical questions: the potential enhancement of type design practices through generative models, the data contributions from type design, and the effective presentation of type design data to learning algorithms.

The surveys reveal three distinct applications of generative models in type design: blending, completion, and modality translation. While blending multiple fonts into novel designs and generating typefaces from descriptions present intriguing possibilities, completion emerges as the most pragmatic application for type design practice. This utility manifests throughout the creative process – from initial sketching through iterative refinement – where completion algorithms provide consistent leverage for the designer’s workflow.

The investigation of datasets presents a curious trichotomy. Bitmap datasets, while abundant, prove questionable for practical font generation due to their inherent loss of vector information. Real font datasets, though readily available, lack the crucial original design drawings that could illuminate the creative process. The absence of raw font datasets – containing original drawings – represents a notable gap that, if filled, could provide valuable insights into the design journey.

The examination of representation types reveals an equally intricate landscape. Spatial representations, predominantly bitmap-based, offer limited utility for type design applications. Sequential representations, focusing solely on paths without contextual relationships, present an improvement over bitmaps but often lack crucial geometric information. The spatio-sequential approach emerges as the most comprehensive, providing both path data and spatial relationships between points – rather like a map that shows both the routes and the territories they connect.

The examination of font technology reveals an intriguing paradox. While fonts fundamentally operate in the domain of vector graphics (Wright April-June/1998; Korpela 2006; Yannis Haralambous 2007), with only occasional intersections into raster graphics through embedded bitmap technology (PeterCon 2020), the prevalent approach in machine learning research takes an unexpected turn down a peculiar rabbit hole.

A return to the literature database raises a compelling question: why does the majority of research represent fonts through raster images of glyphs or alphabets, effectively discarding the original vector data? This curious preference suggests either a limited understanding of type design requirements or an inclination to remain within the comfortable territory of established CNN-based approaches – rather like choosing the well-worn path instead of exploring the uncharted forest.

The vector graphics approach, while more aligned with font technology’s nature, appears in merely eight projects (Campbell and Kautz 2014; Ha 2015; Zhong 2018; Lopes et al. 2019; Aoki and Aizawa 2022; Lian and Gao 2022). Among these, only two (Campbell and Kautz 2014; Lian and Gao 2022) attempted end-to-end font generation. The remaining projects, while successfully generating vector paths in SVG format, encounter limitations as this format was not designed for font technology – though the ability to learn vector path drawing presents a promising foundation.

The preservation of original vector data suggests sequential representation as the logical path forward, potentially enhanced by spatial representation. However, rather than employing pixel-based images, auxiliary points emerge as a more efficient approach, encoding only the essential geometric information – much like a cartographer marking only the significant landmarks instead of mapping every tree in the forest.

This sequential representation introduces its own set of challenges. The same letter shape can be drawn with varying numbers of vector points, creating sequences of different lengths. Combined with diverse shape topologies, this variance presents a non-trivial problem that requires careful consideration.

The abovementioned questions and found challenges will be addressed in the next chapter. Hold on tight!