torchfont 0.13.12

TorchFont: A Machine Learning Library for Vector Fonts

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 CodepointDataset or GlyphIdDataset at a font directory or a repository checkout that already exists on disk, and TorchFont turns font files into lightweight glyph references. CodepointDataset indexes one sample per face and codepoint, while GlyphIdDataset indexes one sample per face and glyph, reaching ligatures and alternates no codepoint maps to. Load outlines explicitly with LoadGlyph in your transform pipeline when tensors are needed.

Installation

The package requires Python 3.10+ and PyTorch 2.5+.

Install TorchFont with uv:

uv add torchfont

Or with pip:

pip install torchfont

Quickstart

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

from torchfont import CodepointData, ElementType, Outline
from torchfont.datasets import CodepointDataset
from torchfont.transforms import LoadGlyph


def collate_fn(samples: list[CodepointData[Outline]]):
    outlines = [sample.data for sample in samples]
    return {
        "types": pad_sequence(
            [outline.types for outline in outlines],
            batch_first=True,
            padding_value=ElementType.PAD,
        ),
        "coords": pad_sequence(
            [outline.coords for outline in outlines], batch_first=True
        ),
    }


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

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

print(batch["types"].shape)  # (8, L)
print(batch["coords"].shape)  # (8, L, 6)

What TorchFont Focuses On

• local font directories and repository checkouts as the input boundary
• local font indexing by codepoint or by glyph id, plus explicit outline loading through lightweight glyph references
• torchvision.transforms.v2-style semantic pipelines for adapting glyph samples
• PyTorch DataLoader integration through an explicit, customizable collate_fn

Manage font repository synchronization with Git or another tool, then point CodepointDataset(root=...) or GlyphIdDataset(root=...) at the resulting directory.

Citing TorchFont

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

@software{fujioka2025torchfont,
    author = {Fujioka, Takumu},
    title  = {{TorchFont}: A Machine Learning Library for Vector Fonts},
    year   = {2025},
    url    = {https://github.com/torchfont/torchfont}
}