When a designer chooses a font, they are not selecting a shape. They are selecting a social signal. A serif font on a wedding invitation means something different from the same serif font on a parking ticket. Xin, Endo, and Kanamori formalize this observation by building FontUse — a dataset of 70,000 images annotated not just with font style (serif, sans-serif, script) but with use case (wedding invitation, coffee-shop menu, protest sign).
The key move is treating typography as jointly conditioned on style and context. Previous work on in-image text generation could match visual properties — weight, slant, decoration — but ignored the pragmatic dimension: what is this text for, and does the font choice make sense in that situation? FontUse provides the supervision to learn this joint distribution. Existing image generators can then be fine-tuned to interpret prompts like “elegant script for a formal event” without architectural changes. The knowledge lives in the data, not the model.
For evaluation, the team introduces a Long-CLIP-based metric that measures alignment between the generated typography and the requested attributes, capturing both visual fidelity and contextual appropriateness in a single score.
A font is fully described by its geometric properties — every curve, every kerning pair, every weight variant is specified. But this complete formal description contains zero information about when to use it. The mapping from form to appropriateness is not derivable from form alone; it must be learned from examples of use. This is true of any designed artifact where function is socially constructed rather than physically determined: architectural styles, color palettes, interface layouts, even code formatting conventions. The specification of the object and the specification of its proper deployment are two entirely separate bodies of knowledge, and mastering one provides no access to the other.