4. Projects: Practical Applications

The preliminaries chapter presents a methodical exploration of the literature, assembling a comprehensive database of font generation projects. This systematic exploration distils key topics that have shaped the discourse since the 2010s. The subsequent surveys scrutinise three foundational topics that emerge at the convergence of type design and machine learning domains – a meeting of centuries-old craft and cutting-edge computation. Contemporary font technology operates in the realm of vector graphics (Wright April-June/1998; Korpela 2006; Yannis Haralambous 2007), though one observes occasional forays into the domain of raster graphics (PeterCon 2020). These exceptions merely facilitate the embedding of a raster image file within a standard font file – a somewhat ungainly marriage of technological approaches that serve specific use cases.

The examination of the literature database reveals a fascinating pattern in the topic of data representation. The majority of projects opted to represent fonts through images of glyphs or images of the alphabet – an approach that, while pragmatic, discards the inherent vector paths carefully crafted in the original font file. The reasons for this sacrifice of vector data remain unclear.

Only eight projects (Campbell and Kautz 2014; Ha 2015; Zhong 2018; Lopes et al. 2019; Aoki and Aizawa 2022; Lian and Gao 2022) achieved end-to-end vector graphics generation, each employing sequential data representation. However, the sequential approach introduces its own set of challenges. The number of vector points used to draw identical letterforms can vary significantly – a seemingly arbitrary reflection of individual designer preferences. The complexity intensifies when considering the diverse topologies of letterforms, from the deceptively simple sans serif to the intricate details of serif typefaces and even more complicated scripts or display fonts. Furthermore, fonts differ not only in their number of glyphs but also in their variant forms, adding another layer of variability to the sequential - but not only - representation challenge.

The datasets survey reveals another inexplicable omission – the absence of data that captures the authentic nature of type designers’ creative process. A stark contrast emerges between the final font file and the font editor file, where the construction methods diverge non-trivially from the ultimate presentation. Consider the letter H, which rarely materialises as a single outline shape in the design phase. Instead, it typically comprises two vertical stems bridged by one horizontal stem – each defined as a distinct element. In serif typefaces, this modular approach extends further, with serifs existing as separate shapes as well.

The final font file, however, presents a unified outline – a deliberate transformation where overlapping elements merge into a single shape, primarily serving to optimise the computational complexity of the font file. This consolidation, while practical for distribution, obscures the intricate geometrical relationships between component shapes. The loss of these spatial and proportional signals – the very essence of type design practice – represents a noteworthy gap in current datasets.

These findings illuminate a compelling opportunity to address the identified challenges through a novel dataset. The project builds upon a fundamental premise: the craft of type design operates within established conventions and best practices, as exemplified in resources like “type basics” (“I ♥ Typeworkshop.com,” n.d.) by Underwear type foundry (“Underware Type Foundry” 1999). Professional type designers, through years of accumulated expertise, follow implicit rules that manifest in their work. The number of sequences used to construct letterforms, rather than being arbitrary, reflects this embedded knowledge. Therefore, the number of variances of the sequential length is limited.

The geometrical relationships between component shapes emerge as another critical aspect of type design practice. These spatial arrangements and proportional systems constitute essential patterns that warrant preservation in any meaningful dataset. This study proposes a two-pronged approach: first, capturing regularised drawings that yield consistent sequential representations, and second, preserving type designers’ original path constructions alongside their final forms.

This investigation culminates in the development of the Regularised Type Designers’ Dataset – LTTR/SET. The dataset presents an intriguing solution to the representation challenges that have historically plagued font generation projects.

This chapter unfolds across six interconnected sections, each addressing a distinct aspect of the dataset development. The Proposal section establishes foundational requirements spanning scope, quality standards, and diversity considerations. A comprehensive Framework follows, detailing the methodological approach and tooling infrastructure. The Presentation section examines the dataset’s digital interface, while the subsequent Demonstration illustrates its practical application through a font generation model case study. The Evaluation section scrutinises the empirical results, leading to the Conclusion, where findings are synthesised, and future research directions are contemplated. Through this structured progression, the chapter presents a thorough examination of the dataset’s conception, implementation, and implications.