YuzuMarker.FontDetection

YuzuMarker.FontDetection earns its place in a designer's bookmarks because it solves a narrow problem that mainstream font matchers ignore. WhatTheFont and Font Squirrel are built around Latin letterforms; feed them a Japanese poster and you get noise. This Space is trained specifically on CJK typefaces, and the same image that confuses a Latin matcher comes back here as a ranked list of candidate faces with confidence scores attached. The confidence score is the feature that makes the output honest. Many matchers present a single answer with false certainty; here you can see whether the model is confident or guessing, which changes how much verification work you do afterward. The attribute extraction — weight, slant, stroke thickness — is the second genuinely useful layer. When the exact font is outside the training set, those attributes still describe the lettering precisely enough to filter a catalogue. Where it falls short is reproducibility. There is no published accuracy figure, no supported-font list, and no documentation of how many typefaces the model was trained on. You cannot predict whether your target font is even in scope, and you cannot reproduce a result across sessions. The demo processes one image at a time, so batch work isn't supported, and the Space runs on shared Hugging Face infrastructure, meaning cold-start delays and availability that depends on the Space being awake rather than on any service commitment. That combination — real capability, no operational guarantees — defines how you should use it. It fits a designer naming a typeface from a scan, a manga or comic translator identifying fonts in Japanese or Korean source pages, a localization specialist auditing CJK typography, or a digital humanities researcher cataloguing scanned East Asian printed material. It does not fit a production pipeline needing an SLA, a documented API, or batch throughput. Take the candidate names and confidence scores as your starting point, then confirm against a specimen. If you need Latin, Latin Cyrillic, or Arabic identification, go elsewhere — this model targets CJK only. One caution worth stating plainly: rare, custom, or newly released CJK fonts are unlikely to be represented in the training set, so treat a low-confidence result as a signal to stop trusting the exact-name guess and lean on the style attributes instead. The model architecture and training data are not described in the material available, so there is no way to reason about its blind spots from first principles.