Font.Garden
← Guide index

startThe collection

The collection is a set of masters drawn on one dyadic grid system, the models trained on them, and the fonts generated from both. It grows by drawing more designs on the same grid and by grading what the models draft. This chapter describes the state of the collection and the plan for it.

Background reading:
start Post-authorship and open weights

What is in it

The seed of the collection is Virtua Grotesk, a neo-grotesk drawn on the dyadic self-labeling grid described in The dyadic grid. As of this writing it has two masters, Regular and Bold, with Latin, Arabic and Hebrew. Its sources and design documentation are public at github.com/eliheuer/virtua-grotesk.

The first model, Virtua-12M-v0.1, is a 12.54 million parameter decoder-only transformer trained on those sources. It drafts a Bold from a Regular and redraws a brush sketch in the collection’s conventions. See Models.

The next designs planned on the same grid are an oldstyle serif in the Plantin tradition, a mono, and further weights of Virtua from Thin to Black. Each adds a style region to the collection and real data to the corpus.

How it grows

  1. A person draws a core set of glyphs for a new style on the grid.
  2. A model drafts the rest: the remaining glyphs of a master, the other masters, the other scripts. Drafts are marked purple.
  3. A person corrects and grades each draft. Green glyphs enter the corpus.
  4. The model is retrained on the larger corpus and the next drafts are better.

Interpolation between masters gives volume, because every grid-snapped interpolation is a valid instance. Variety comes only from new hand-drawn designs. The corpus grows by drawing, not by collecting.

What it is not

  • It is not a catalog of families by named designers. See Post-authorship and open weights.
  • It is not trained on Google Fonts or other OFL collections. An early experiment did this, from compiled binaries, and was retired because compiled outlines carry conversion artifacts and unrelated designs add noise. See The collection is the dataset.
  • It is not a text-to-font service. Zero-shot generation from a prompt is not a goal.

The long-term aim

Fonts that are models rather than tables of outlines. An OpenType font stores one fixed shape per glyph in a box inherited from metal type. A neural font draws each glyph in context. The collection is being built so that its sources can train such fonts. See The NeuralType format.