Source-limited few-shot font generation for multi-style logotypes: A type designer-AI collaborative framework

Abstract Font generation algorithms have evolved from many-shot to few-shot methods in recent years. However, both approaches still diverge from real-world design scenarios. To provide designers with a clear type designer-AI collaboration paradigm, we propose the concept of source-limited few-shot font generation, which utilizes fewer than 10 characters commonly appearing in logotypes as experiment samples. We develop a Font Style Classification Standard (FSCS) to organize fonts in the dataset and guide training strategies. Our experiments reveal two key findings: (1) the classification standard and training strategy proposed in this study significantly enhance font generation results; (2) the design evaluation metrics provide an effective approach for assessing the quality of source-limited few-shot font generation. This exploratory study represents an initial step in applying font generation algorithms to type design. It provides valuable insights for future studies on new font generation methods and human-AI collaboration.