5. Conclusions
This thesis examines the convergence of type design and machine learning through a tripartite structure. The preliminary chapter establishes the taxonomic foundations – a cartography of concepts that maps the territory where pixels meet vectors. The survey sections chart the landscape of applications, datasets and representations, whilst the project chapters traverse from theory to practice through the LTTR/SET dataset and its framework, through the demonstration of the dataset and evaluation of the results – the generated fonts.
This thesis presents contributions that serve distinct yet interconnected audiences in the realm of AI-driven type design. For type design practitioners, the study delivers a meticulously crafted regularised font dataset – LTTR/SET – that maintains typographic standards whilst ensuring vector path consistency. The accompanying Font Tester tool enables visual evaluation of generated fonts, whilst the systematic methodology integrates seamlessly with industry-standard tools like the LTTR/INK plugin.
Machine learning researchers benefit from the first comprehensive analysis of machine learning font generation projects, complete with standardised categorisation. The research challenges prevailing assumptions about dataset size, demonstrating how smaller, well-structured datasets can yield superior results to their larger, more diverse counterparts.
The study bridges these disciplines through shared contributions: a unified technical vocabulary that spans both domains, empirical evidence of skeleton type design principles enhancing ML model performance, and evaluation heuristics that merge traditional typographic criteria with machine learning metrics.
The findings reveal a curious paradox in the current landscape of machine learning font generation. While the field advances with impressive velocity, it simultaneously moves away from industry requirements. The prevalent practice of discarding vector data in favour of raster images creates an expanding chasm between academic innovation and practical application, rather like building a bridge that starts from both sides but meets in the middle with mismatched heights.
The sequential representation of vector data presents particular challenges, with merely eight projects achieving end-to-end vector graphics generation. This limited success suggests the need for architectural approaches that transcend traditional CNN-based solutions. The complexity of vector paths, coupled with the absence of geometric relationships in current datasets, results in a peculiar situation where valuable design information dissipates into the digital ether.
Perhaps most intriguingly, the research challenges regarding the dataset size and diversity in font synthesis using deep learning methods appear to have been addressed.
The superior performance of regularised datasets over their larger, more diverse counterparts suggests a more nuanced approach to data curation. Fine-tuning on regularised data demonstrates enhanced style matching compared to extended diverse training, indicating that the path to improvement lies not in accumulating more data, but in structuring it more thoughtfully.
The future trajectory of this research extends in three distinct directions. The LTTR/SET dataset presents opportunities for expansion beyond its current scope, encompassing serif and script styles whilst establishing rigorous benchmarks for vector graphics learning across diverse writing systems. The technical landscape beckons with possibilities in continual learning applications and transfer learning methodologies, where the systematic nature of LTTR/SET may serve as a foundation for improved sequential encoding methods. The research horizon suggests deeper investigations into regularisation benefits, whilst the development of sophisticated evaluation metrics promises to bridge the peculiar gap between machine learning metrics and typographic quality assessment, rather like teaching a computer to appreciate the subtle difference between Helvetica and Arial. The integration with professional type design workflows emerges as a critical path forward, where theoretical advances must navigate the practical demands of type design practitioners.
