SDXL LoRA fine-tune on Google Fonts glyphs. Phase 0 of a longer typography-generation project.
uv sync --extra dev
cp .env.example .env
# fill in HF_TOKEN, WANDB_API_KEY, HF_USER| Command | Purpose |
|---|---|
typeface fetch-fonts |
Clone github.com/google/fonts into fonts/. |
typeface render |
Render TTFs → PNG corpus in data/glyphs/. |
typeface push-dataset --repo-id USER/NAME |
Upload corpus to HF Hub. |
typeface train --dataset ... --output ... |
Launch LoRA training on Modal. |
typeface infer --prompt ... --lora ... |
Generate one image on local MPS. |
typeface compare --prompts ... --lora ... |
Vanilla-vs-LoRA grid. |
See docs/superpowers/specs/2026-04-19-typeface-model-phase-0-design.md.
- What worked:
- What broke:
- What surprised me:
- Phase 1 candidates: