1.3. Contributions: Place for Praise

This thesis presents contributions across three main chapters: Preliminaries, Surveys, and Projects. Each chapter culminates in specific advances that bridge the gap between academic research and industry practice in AI font generation.

The preliminary research established essential frameworks for the field:

1. A literature database of AI font generation projects, categorised by key topics and including content summaries, serving as a central resource for literature review and trend identification.
2. A standardised taxonomy culminating around a machine learning pipeline covering key concepts in AI font generation, providing a common language for type design and machine learning practitioners addressing terminological inconsistencies.
3. An automated system for literature database maintenance using a set of robots and text extraction AI agents, allowing keep the database actualised in the future.

These contributions lay the groundwork for newcomers to AI font generation, sparing them the traditional academic rite of passage – months spent wading through scattered papers. With a curated knowledge base and standardised frameworks already in place, researchers can bypass the preliminary surveying phase and dive directly into their specific areas of interest. A rather cultivated approach to academic initiation, one might say.

The survey research mapped the intersection of AI font generation and type design through three key contributions:

1. A systematic analysis of AI font generation applications, covering current implementations and emerging trends, enabling researchers and practitioners to identify promising directions for future development.
2. A comprehensive catalogue of font datasets, documenting their characteristics and quality metrics.
3. A structured classification of font data representations, mapping relationships between traditional type design and machine learning approaches, facilitating more effective translation between these two domains. These contributions illuminate the uncharted territories where type design meets artificial intelligence – a landscape where traditional craft collides with algorithmic ambition, often with curious results. While the surveys offer essential navigation through this peculiar terrain, they also reveal how much remains unexplored in this rapidly evolving field.

The project research delivered practical advances through two steps:

1. Regularised Font Dataset
   1. A novel regularised font dataset engineered with consistent sequential lengths and typographic quality standards, enabling reliable training of ML models on vector fonts.
   2. A systematic font creation methodology, combining skeleton type design principles with parametric control through the LTTR/INK plugin, empowering type designers to create ML-ready datasets while maintaining typographic quality.
   3. A comprehensive evaluation framework incorporating six key assessment criteria for font generation models, enabling systematic comparison of model performance across technical and typographic dimensions.
2. Empirical Experiment
   1. An interactive web-based evaluation tool (AI Font Tester) featuring real-time comparison capabilities and OTF export functionality, facilitating practical assessment of AI-generated fonts by both researchers and type designers.
   2. A structured set of typographic heuristics, combining traditional quality criteria with AI-specific considerations, providing the first standardised framework for evaluating AI-generated fonts from a type design perspective.
   3. A set of empirical findings about font generation training, derived from a systematic comparison of dataset types and training strategies, challenging common assumptions and providing practical guidelines for optimising generation quality:
      1. Regularised datasets outperform larger, diverse datasets in consistency
      2. Fine-tuning on regularised data yields better style matching than extended diverse training
      3. Generation quality plateaus after 50 trial samples