Font.Garden
← Guide index

nerdModels

The collection trains two kinds of model. Glyph language models draw outlines as token sequences and run inside Runebender. NeuralType models are fonts: the weights draw letterforms in context. Both run locally on a laptop or a gaming GPU. Weights are released under Apache-2.0. The first public release is in preparation.

Background reading:
nerd The collection is the dataset
start The NeuralType format

Virtua-12M

Virtua-12M-v0.1
architecturedecoder-only transformer, 6 layers, 384 dimensions, 8 heads
parameters12.54 million
context1,368 tokens
vocabularyabout 1,700 tokens: commands, names, coordinates, deltas
tasksweight transfer (bolden), completion, sketch to glyph
trainingone M4 Pro laptop in 3 hours 20 minutes with MLX, or 20 minutes on an RTX 2060 SUPER
checkpoint50 MB, float32 safetensors
releaseplanned: Hugging Face fontgarden/virtua, Apache-2.0. Not yet uploaded.

Weight transfer is delta prediction. Given the base master, the model predicts, one point at a time, the offset from each point to its position in the heavier master. Each delta is a small bounded move, which is what lets a small model do the job, and the point structure is preserved by construction, so the masters still interpolate. The output is a draft, marked purple. A person corrects it on the 2-unit grid.

Two implementations, MLX and PyTorch, are held to identical output by a parity script. Weights will ship as float32 so that others can continue training. The model card will record the baseline, the model score across seeds, and the corpus manifest.

Running models

Task models run through font-ml, a Rust crate built on Candle with Metal or CUDA. A bolden pass takes about half a second per glyph on an M4 Pro. A model package is a directory with config.json, weights.safetensors and vocab.txt. Packages live in ~/Runebender/models/<kind>/<name> and, per font, in MyFont.models/.

The Shapes tool trains a separate 5.5 million parameter model from scratch on a UFO of named shapes. A quick run takes about 12 minutes and a thorough run about 30 on an M4 Pro. Adding a shape to a trained base takes about 2 minutes. This is the smallest demonstration of the whole loop: draw, train, propose, grade.

Chat models are open-weight GGUF files run through llama-server. They answer questions about a font by calling editor tools that return measurements. A 4B model gives a correct stem width only with a tool.

NeuralType models

A NeuralType model is distilled from a teacher. For the Nastaliq model the teacher is Gulzar shaped with HarfRust: for every one, two and three letter string, the shaper returns glyph ids, advances and offsets, and the outlines are read from the font. The student is a feed-forward network that predicts a signed-distance field for each cluster and its position. See The NeuralType format.

The next model is a font designed for NeuralType from the start, on the collection’s grid, rather than distilled from an OpenType font.

Releases

modeltaskwherelicensestatus
Virtua-12M-v0.1weight transferHugging Face fontgarden/virtuaApache-2.0in preparation
gulzar-int8.ntfNastaliq NeuralType fontpost-opentype repositoryOFL, with Gulzar’s copyright noticereleased

The Virtua-12M-v0.1 upload is the next release task. Until it is done, the only public model file is the Nastaliq font. What stays private: the training code, the corpus tools and the plans. The editor, the model runtime, the sources and the released weights are public.

Model file safety

Published model files are safetensors or GGUF, which hold numbers rather than code. Check the publisher’s account name letter by letter, compare the SHA-256 checksum, and read the license field inside the file before you load it.

Next reading:
team Contributing to the collection