From Diffusion to TTF: Making an AI Font and Tackling Normalization

After poor results trying to edit glyph vectors with an LLM, the author switched to a diffusion model that generated clean letter images. They converted images to SVG with PIL + Potrace and built a TTF, discovering that consistent typographic metrics (baseline, x-height, ascenders/descenders) are the real challenge. With grids, references, and scripting, results improved, pointing toward an accessible path to custom AI-made fonts.

Key Points

• LLMs struggle with precise geometric edits to glyph vectors; text-only manipulation is unreliable for visual accuracy.
• Diffusion models excel at generating legible, stylistically coherent letterforms from prompts on the first try.
• A practical pipeline is: generate raster letters → threshold to black/white → vectorize with Potrace → assemble into TTF/OTF.
• The main bottleneck is typographic normalization—maintaining consistent baseline, x-height, ascenders/descenders, and overshoot across glyphs.
• Providing traced reference letters and a typographic grid improves consistency, but careful metric management remains essential.

Sentiment

The community is mildly amused but largely skeptical. Appreciation for the accessible pipeline is overshadowed by debunking the novelty claim and critiques of output quality. The tone is light with humor, not hostile.

In Agreement

• The diffusion-to-TTF pipeline (generate → binarize → vectorize with Potrace → compile) is genuinely accessible and represents a practical democratization of custom font creation.
• The article honestly documents the main challenge of typographic normalization—baseline, x-height, ascender/descender consistency—which aligns with real font engineering difficulty.
• Design and branding agencies charge enormous fees (like the $2,000/character example), so AI tools that reduce that cost have clear value.
• The chaotic, exploratory writing style was appreciated by some commenters as an authentic record of the experimentation process.

Opposed

• The 'world's first AI-generated font' claim is false: prior examples include a June 2024 Python/SD 1.5 script, Tom7's 2021 AI fonts, gwern's dropcap workflow, and years of academic research.
• The resulting fonts are of poor quality and the article fails to show them in actual use, undermining the demonstration.
• The idea of every Substack owner making their own AI font is viewed with horror as a regression to 2000s-era web design chaos.
• Font copyright is largely irrelevant to the value of this work, since typefaces are not copyrightable in the US anyway.