torchfont 0.9.3

Datasets, Transforms and Models specific to Vector Fonts

TorchFont

TorchFont is an unofficial library based on PyTorch for deep learning with vector fonts. It is not affiliated with or endorsed by the PyTorch project.

TorchFont is local-first: point GlyphDataset at a font directory or a repository checkout that already exists on disk, and TorchFont turns glyph outlines into GlyphSample objects for PyTorch pipelines.

Installation

The package requires Python 3.10+ and PyTorch 2.3+.

Install TorchFont with uv:

uv add torchfont

Or with pip:

pip install torchfont

Quickstart

from torch.utils.data import DataLoader
from torch.nn.utils.rnn import pad_sequence

from torchfont.datasets import GlyphDataset, GlyphSample


def transform(sample: GlyphSample):
    return sample.types, sample.coords


def collate_fn(batch):
    types = pad_sequence([types for types, _ in batch], batch_first=True)
    coords = pad_sequence([coords for _, coords in batch], batch_first=True)
    return types, coords


dataset = GlyphDataset(
    root="~/fonts",  # or "tests/fonts" in this repository
    patterns=("*.ttf",),
    codepoints=range(0x20, 0x7F),  # printable ASCII
    transform=transform,
)

loader = DataLoader(dataset, batch_size=8, shuffle=True, collate_fn=collate_fn)
types_t, coords_t = next(iter(loader))

print(types_t.shape)   # (8, L)
print(coords_t.shape)  # (8, L, 6)

What TorchFont Focuses On

• local font directories and repository checkouts as the input boundary
• Rust-backed outline decoding into GlyphSample with outline tensors, metrics, and glyph name
• transform utilities such as quad_to_cubic for adapting glyph samples
• PyTorch DataLoader integration through ordinary user-defined collate_fn functions

TorchFont does not need to own Git clone / fetch / checkout in the main workflow. Sync repositories with Git or another tool, then point GlyphDataset(root=...) at the resulting directory.

Citation

If you find TorchFont useful in your work, please consider citing the following BibTeX entry:

@software{fujioka2025torchfont,
    title        = {TorchFont: A Machine Learning library for Vector Fonts},
    author       = {Takumu Fujioka},
    year         = 2025,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/torchfont/torchfont}}
}