Japanese manga OCR that keeps pop-up dictionaries happy

What it does

mokuro is an offline preprocessing pipeline that runs text detection and OCR over manga page images, then produces a .mokuro metadata file. Load that file alongside the raw images into its companion web reader, and the Japanese text becomes selectable—letting you hook in pop-up dictionaries like Yomitan. It is essentially glue between comic-text-detector and manga-ocr, wrapped in a format that separates the data from the reading interface.

The interesting bit

The project treats OCR as a build step rather than a runtime dependency: all the heavy inference happens before you open the book, so reading is just a lightweight web app. The newer .mokuro format deliberately decouples the text overlay data from the reader GUI, which avoids the fragility of the old self-contained HTML files where a moved image folder broke everything.

Key highlights

• Runs fully offline: all heavy processing happens before reading, leaving a lightweight browser-based reader
• Generates .mokuro files containing OCR results and metadata, keeping data separate from presentation
• Companion web reader doubles as a catalog for processed series and volumes
• Integrates with language-learning tools like Yomitan for instant dictionary pop-ups
• Optional GPU acceleration via PyTorch, with CPU fallback available

Caveats

• PyTorch dependency means the newest Python releases often break compatibility until upstream catches up
• Windows users with Microsoft Store Python may hit DLL load errors with fugashi
• The legacy HTML output is still generated but frozen; the README nudges you toward the newer .mokuro workflow

Verdict

Worth a look if you are learning Japanese and have a folder of raw manga scans. If you do not read manga or study Japanese, this is not your tool.