[ CODE ] PROJECT
GlyphOCR
Drop a manga page, a menu or a screenshot. GlyphOCR boxes every line, translates it and letters the translation back onto the image.
Origin
It started as a manga translator I built at the InnovArt 2026 hackathon: detect speech bubbles, OCR them, translate, and typeset the result back onto the page. GlyphOCR is that prototype turned into a product.
What it does
- Auto-detects 11 languages, including vertical Japanese
- Boxes every line in about a second, then streams translations in
- Seven tools from one scan: image to text, summaries, quizzes, flashcards, pronunciation guides and more
How it works
- Front end: Next.js, TypeScript, Tailwind. I designed the whole UI
- Back end: Python and FastAPI, streaming NDJSON to the page
- OCR on the GPU: CRAFT text detection, manga-ocr for Japanese, per-script readers for other languages, optional SAM 2.1 refinement
- AI: an LLM for translation and study tools, with reasoning disabled to cut scan time from 9–31 s to about 3 s
- Business: free tier, Pro plan and a pay-per-page API, with Stripe billing and API keys
Status: in progress, launching soon.
