hangul-font-finder

Identifies the free Korean font used in an image of a single line of Hangul text.

한글 텍스트 한 줄 이미지에서 사용된 무료 한글 폰트를 식별하는 모델입니다.

• Identifies 1,279 free Korean fonts (1,393 font families, Fonts List)
• Robust to outlines, shadows, 3D extrusion, gradients, horizontal scaling, partial occlusion, low resolution and compression
• One line image in, no OCR or character segmentation
• Outputs calibrated font probabilities
• Trained on synthetic data

Input	Top-1	Prob.
	Black Han Sans	0.996
	Jua	0.996
	Do Hyeon	0.994

Usage

pip install torch transformers safetensors pillow numpy huggingface_hub

import sys
from huggingface_hub import snapshot_download
from PIL import Image

path = snapshot_download("prenny/hangul-font-finder")
sys.path.insert(0, path)
from inference import FontIdentifier

model = FontIdentifier(path)
model.predict(Image.open("line.png"), top=5)
# [{'font': 'Jua', 'same_shape_fonts': ['BM JUA_TTF', 'Jua'],
#   'korean_name': '배달의민족 주아', 'probability': 0.9955}, ...]

• Input is an image of one cropped text line. Text detection is not included.
• Fonts with identical Hangul glyphs (OTF/TTF builds, English/Korean names, Google Fonts releases) are returned as one font, with all names in same_shape_fonts.
• Runs on CUDA, MPS and CPU.

Evaluation

Test set	Lines	Top-1	Top-3	Top-5
Hard (mixed effects)	2,702	93.3%	97.9%	98.7%
Clean	1,351	97.4%	≥ 99.2%	≥ 99.8%

Predictions with probability ≥ 0.95 cover 84% of the hard set at 99.6% accuracy.

Training

• Architecture: SigLIP 2 Base vision encoder (fully fine-tuned) + 512-d projection + linear classifier
• Input encoding: greyscale, resized to 56 px height (aspect preserved), folded into 4 rows of a 224×224 image; long lines averaged over 3 windows
• Data: about 400k synthetic line images (about 300 per font)
• Schedule: 6 epochs, AdamW (lr 3e-5), batch 128, label smoothing 0.1, temperature-calibrated (T = 0.690), 1× RTX 5080 (about 85 min)

Limitations

• Fonts outside the 1,279 known fonts are answered with the most similar known font, usually with a low probability.
• All reported numbers are measured on synthetic images.
• Designed for Hangul; lines with only Latin letters or digits were not evaluated.
• The model only provides font names. Check each font's licence before use.

License

MIT. Weights are fine-tuned from SigLIP 2 (Apache-2.0); see LICENSE-SIGLIP2-APACHE-2.0 and NOTICE.