mlfont-notosanssc: Range-Compact MLFont Slicing for Noto Sans SC (Simplified Chinese)

This is not an officially supported Google product.

Overview

Pretrained MLFont machine-learning-guided font slicing model and WOFF2 slice bundle for Noto Sans SC (Simplified Chinese) (NotoSansSC-Regular.otf, 30,759 Unicode codepoints partitioned into 102 range-compact slices).

• Paper: Learning Better Font Slicing Strategies from Data (Quan Wang and Yiran Mao)
• Open-Source Repository: https://github.com/wq2012/mlfont
• PyPI Package: https://pypi.org/project/mlfont/

Training Configuration

• Target Font: NotoSansSC-Regular.otf (30,759 Unicode codepoints)
• Training & Evaluation Corpus: wikimedia/wikipedia (20231101, languages: zh), 1,600 training articles and 400 testing articles
• Slicing Architecture: Tiered Document-Frequency Slicing + Online Variational LDA Topic Modeling + Range-Compact Genetic Algorithm (mlfont preset: chinese)
• Total Slices: 102 slices (11,741 total CSS unicode-range intervals)
• Physical WOFF2 Bundle Size: 9183.8 KB across 102 .woff2 files (companion CSS: 129.6 KB)
• Training Wall-Time: 118.6 seconds on CPU/GPU

Model Files Included

• model.safetensors: Learned topic-codepoint weight matrix (lambda_matrix), codepoints, frequency_array, and dna_assignment in SafeTensors format
• strategy.json: Complete SlicingStrategy definition and hyperparameters
• config.json: Model and font bundle summary metadata
• font_sliced.css: Production-ready @font-face stylesheet with compact unicode-range descriptors
• woff2/*.woff2: 102 physical WOFF2 font slices generated via fontTools.subset
• nam/*.nam: 102 Google Fonts .nam codepoint subset files
• evaluation_metrics.json: Full benchmark metrics across all 7 network/session profiles

Benchmark Performance (400 Wikipedia Test Pages, Session Length L = 10)

Strategy	Slices	CSS Ranges	Missing Glyphs	GIN-2g Mean / P95 (ms)	GIN-3g Mean / P95 (ms)	DSL Mean / P95 (ms)	Session Util (%)
single_slice	1	418	0	190.95 / 1046.58	47.78 / 261.86	9.54 / 52.31	5.00%
equal_bin_102	102	519	0	140.76 / 851.22	37.53 / 226.94	6.80 / 41.15	6.90%
random_slice_102	102	30457	0	202.63 / 1325.94	54.02 / 353.49	9.80 / 64.10	5.00%
language_slice	12	7087	121	68.74 / 377.15	17.36 / 95.16	3.42 / 18.77	13.95%
most_1000_500_equal_bin_100	102	2978	0	111.15 / 493.00	29.58 / 131.64	5.38 / 23.81	7.17%
online_lda_topics100_greedy	102	11848	0	57.88 / 305.81	15.36 / 81.13	2.81 / 14.82	14.99%
dnntm_topics100_greedy	102	11928	0	48.51 / 210.03	12.86 / 55.64	2.35 / 10.19	17.46%
range_compact_most1000_500_topics100 (This Model)	102	11741	0	58.10 / 307.01	15.41 / 81.43	2.82 / 14.88	14.93%

Usage

Install mlfont:

pip install mlfont huggingface_hub

Load the pretrained bundle and analyze or slice a font in Python:

from huggingface_hub import snapshot_download
import mlfont

model_dir = snapshot_download(repo_id="wq2012/mlfont-notosanssc")
bundle = mlfont.load_pretrained(model_dir)
print("Loaded strategy:", bundle.strategy.name, "slices:", bundle.strategy.num_slices)
print("CSS unicode-range count:", bundle.strategy.count_unicode_ranges())

Citation

@article{wang2017learning,
  title={Learning Better Font Slicing Strategies from Data},
  author={Wang, Quan and Mao, Yiran},
  journal={Technical Disclosure Commons, Defensive Publications Series},
  volume={906},
  year={2017}
}