Abstract
As digital typing supplanted handwriting as the mode of expression, a paradox emerged in written communication. Where once each individual used a distinctive handwriting style – their visual fingerprint – the standardisation of digital typography has confined typographic expression to a set of operating system fonts. The doctrine of perfect type set block dominating the typographic landscape enforces a rigid view of communicational design. This standardisation has defeated the personal expression that characterised handwriting. The implications extend beyond aesthetics – where love letters once carried the tremor of handwriting, dating app messages now arrive in the same system font used for quarterly reports.
Font production demands expertise and collaborative teamwork – a complexity that limits the delivery of personalised typography on demand. AI font generation shows promise, yet current solutions fall short of industry requirements. This thesis addresses three questions: What is the current landscape of AI font generation? What limitations prevent the adoption of AI solutions in professional type design? How might these limitations be overcome through methodological improvements? These questions illuminate the path from theoretical possibility to practical implementation, where craft meets computation.
The first question finds its answer through a literature review, culminating in a taxonomy and database of AI font generation projects – a cartography of concepts mapping the territory where pixels meet vectors. The second question reveals itself through surveys of applications, datasets, and data representation methods, uncovering two limitations: most projects discard vector data in favour of raster images, while the few vector-based approaches struggle with irregular datasets and the absence of original type design drawings. This dual misalignment suggests a solution to the third question: the need for regularised datasets that maintain the vector-based nature of type design and typographic quality standards.
The doctoral thesis hypothesises that AI font generation improves through data regularisation and representation of raw type design drawings. The methodology employs a skeleton type design system to create LTTR/SET – a dataset encoding the anatomical structure of letterforms. The resulting dataset maintains vector path consistency while preserving typographic quality standards. Comparative evaluation shows that DeepVecFont-2, when trained on LTTR/SET, generates more precise drawings than its original version. These results demonstrate that the path to better AI font generation lies in structured data representation rather than dataset size.
Keywords
artificial intelligence, machine learning, type design, typography, font generation, font completion, type design datasets, skeleton type design, parametric type design
