fontaine

Online font recognition over a stream of text-box crops.

uv sync                                             # 1. install
uv run fontaine assets fetch                        # 2. get the fonts and backgrounds
uv run fontaine generate -n 5000 -o data/streams/v1 # 3. freeze a stream
uv run fontaine recognize --stream data/streams/v1  # 4. score the baseline

Then write your model as one file in models/ implementing two methods:

# models/my_cnn.py
from fontaine.contracts import Recognizer


class MyCNN(Recognizer):
    """One line saying what this is — shows up in `fontaine models list`."""

    name = "my-cnn"  # what --model matches

    def predict(self, image):  # a PIL crop; return a face_id, or None to abstain
        ...

    def learn(self, image, label):  # label may be a font never seen before
        ...

uv run fontaine models list                                          # confirm it was found
uv run fontaine recognize --stream data/streams/v1 --model my-cnn    # score it

Nothing in src/fontaine/ needs to change, and you can copy models/last_seen.py (23 lines) as a template. Two rules that shape the design: predict is always called before learn on the same item and scored on that prediction, and the label space is never announced — a fixed output layer cannot play.

Useful while iterating:

Command	What it does
fontaine preview -n 48 -o data/preview.png	contact sheet of crops — the fastest look at your config
fontaine arrival -n 50000	simulate the font schedule with no rendering, under a second
fontaine recognize -c configs/stream.yaml -n 5000	score against a live stream instead of a saved one
fontaine recognize --stream … --limit 500	short run while debugging
pytest / ruff check / ty check	tests, lint, types

Once your model works and you have some performance metrics to show, open up a PR!