kha-white/manga-ocr

Built on Transformers' Vision Encoder Decoder architecture, it handles multi-line text recognition in a single forward pass—enabling entire manga speech bubbles to be processed without line splitting. The model is specifically trained to robustly handle manga-specific challenges including vertical/horizontal text, furigana annotations, image overlays, and low-quality scans. Integrates with clipboard and directory monitoring for background processing, enabling workflows with screenshot tools (ShareX, Flameshot) and dictionary lookup applications like Yomitan.

2,582 stars and 17,983 monthly downloads. No commits in the last 6 months. Available on PyPI.