4.1. Proposal: Down of Regularised Font Datasets
Achieving a regularised number of sequences per letter shape and maintaining the quality of the drawings is not a trivial effort. Employing automation to add missing or remove redundant vertices makes the examples inaccurate. Reparameterising outlines of existing fonts on the other side results in rusty, appealing details. For instance, Font-MF (Campbell and Kautz 2014) employed reparametrisation to match splines of each individual character (glyph) across all fonts. Due to topological and letter representation differences, reparametrisation and interpolating such shapes often resulted in undesired shapes. The reconstruction of the required amount of vertices following the best practices of typography is akin to compromising the quality of curves. Therefore, instead of regulating the existing font libraries, the aim is to develop a set of custom fonts and augment the dataset by using methods similar to multiple master fonts (Adobe Systems 1997).
Scope
LTTR/SET aims to cover the design space of the most popular typeface styles following the CEDARS+ classification (Boardley 2021). The preliminary dataset focuses on sans serif styles whose shapes range between three extremes: oval loops with static energy, rounded rectangle loops with static energy and superellipse loops with lively energy (Fig. CLS). On top of every shape extreme, LTTR/SET provides stroke contrast variations that range between low contrast to medium contrast, including contrast tilts of 270 degrees. To offer seemingly fluent transitions between extremes, we have provided a high granularity of instances on the design space continuum. The glyphs coverage offers a basic Latin alphabet with basic punctuation glyphs (Fig. GCV).
At the current stage, we are omitting serif styles, script styles, and styles that don’t use typographic stroke aesthetics by design due to the methodology of building a dataset and narrowing the task to the necessary input of testing the hypothesis of regularisation.
Further is discussed quality requirements.
Quality
Aesthetical attributes: The least tangible to evaluate are aesthetical attributes. It is nearly impossible to set objective judgments of aesthetics. For further research on aesthetics and its evaluation, the great resources of works of MacCormacks and his students(McCormack 2005; Colin G. Johnson et al. 2019; Jon McCormack and Andy Lomas 2021). Since the goal of dataset examples is to provide a representation of the real world, we aim to cover typical representatives of the aesthetics in the selected genres 1.
Manufacturing attributes: The high-quality fonts are not only aesthetically pleasing, but they are well shaped in manufacturing details such as design consistency, harmonious spacing that fits the design, well-designed diacritical characters and auxiliary characters like punctuation, numerals, currency symbols, etc. But most important to our dataset are properly placed nodes of Bézier curves.
Technical attributes: The technical attributes follow the standard requirements of running and displaying fonts on various software and operational systems. For deep learning purposes, attributes such as hinting, open-type features, or font metadata.
To achieve the aforementioned quality standards, the project involved recognised professional type designers.
Diversity
The LTTR/SET aims for a diversity of styles and letter variants. The diversity of styles aspires to cover the taxonomies of the known typographic styles 2, while the diversity of letter variants aims to cover variations of letter constructions within a style 3
While there are plenty of known taxonomies, even widely adopted, no matter how revered, it seems to come with their own personal fan club of critics. (Laudon 2021). Fortunately, the effort of the typographic initiative I Love Typography (ILT) introduced the CEDARS+ (Boardley 2021) classification. The aim of the LTTR/SET is to exploit the benefits of CEDARS+ classification as much as possible while introducing continuums instead of categories 4.
The diversity of letter representations covers different structures of the same letter in the same style. For instance, in the Latin alphabet, there are two most common representations of the letter “a” single or double-storey. The list of letter variants is infinite. However, in the beginning, LTTR/SET aims to cover the most common use cases.
One of the design methods is the exploitation of aesthetical attributes of existing genres or subcultures. By leveraging this method, we can capture representation or interpretation of the perceived world. This can happen by our team or by inviting other designers to contribute to the aesthetical space of the dataset↩︎
(e.g., sans serif, serif and their genre appearances)↩︎
e.g., single or double storey; and the shape of the middle stroke↩︎
CEDARS+ provides categories for specific stylistic topologies. However, those topologies are rather parts of the design space continuum rather than isolated categories↩︎
