Scanlation pipelines, now automated

What it does

This project ingests an image—typically manga or comics—finds the text bubbles, runs OCR, translates the content, removes the original lettering via inpainting, and renders new text back in place. It supports Japanese, Chinese, English, and roughly twenty other languages. The output is a finished image, not a subtitle file.

The interesting bit

The pipeline bundles several distinct computer-vision stages (detection, OCR, translation, inpainting, rendering) into a single pass. The README is admirably honest about its target audience: “Japanese novices like me” trying to parse untranslated group-chat comics. There is also a separate Rust rewrite for anyone who prefers a compiled binary over wrestling with PyTorch and Python 3.10 compatibility.

Key highlights

• Runs as CLI batch processor, web server, or API endpoint
• GPU acceleration supported via CUDA; CPU fallback available
• Docker image ships with all dependencies and models (~15 GB)
• Optional GPT-based translators configurable via environment variables
• Browser userscript exists for inline translation workflows

Caveats

• The project self-describes as “still in the early stages of development and has many shortcomings”
• Showcase examples may not reflect current main-branch quality
• Online demo at cotrans.touhou.ai is noted as no longer working; backup demo at touhou.ai/imgtrans/ is hosted on Google GCP and may restart unpredictably

Verdict

Worth a look if you run a scanlation group, archive raw manga, or just want to read untranslated image threads without learning Japanese. Skip it if you need production-grade reliability or are allergic to 15 GB Docker images.