WhatFontIs-Bench: A Synthetic Benchmark for Font Family Identification in Real-Looking Images

Identifying the typeface used in a photograph is a common practical task, served by several commercial and open tools, yet there is no public test set on which such tools can be compared: photographs of text rarely come with a reliable font label, and near-identical typefaces published under different names make manual labelling error-prone. We introduce WhatFontIs-Bench, a synthetic benchmark in which the ground truth is known by construction. Version 1.0 contains 11,995 JPEG images of a single word set in one of 600 fonts (200 sans-serif, 200 serif, 100 slab serif, 100 monospaced), rendered onto CC0 photographs of real surfaces, inside real scenes and on printed objects, at three controlled difficulty levels. Every image records the exact font, the text, the word outline, the outline of every letter and all rendering parameters, in JSONL and COCO format. We define a family-level top-k evaluation protocol and report a first baseline: a production font-identification API that searches a catalogue of over 1.2 million fonts reaches 83.7% top-1, 93.3% top-5 and 96.5% top-20 accuracy, with sans-serif typefaces markedly harder (75.7% top-1) than slab serifs (95.0%). The dataset is released under CC BY 4.0.