manga-translator

Manga Translator

Batch manga page translator (JP→EN first) with free/local models.

Stack: Python FastAPI worker + Next.js UI. Phase 1 ships detect → OCR → translate → overlay preview. Phase 2 adds inpaint + typeset for finished visual pages.

Separate from Dinamus/ml-danbooru-onnx-repo (image tagging / folder routing).

Pipeline

1. Detect text regions (heuristic contours; CTD optional later)
2. OCR source text (manga-ocr when MANGA_TRANSLATOR_REAL_MODELS=1, else stub)
3. Translate JP→EN (opus-mt-ja-en when real models, else stub)
4. Overlay preview (Phase 1) → Inpaint + typeset (Phase 2)
5. Export side-by-side PNG + zip / CBZ

Quick start (Phase 1 stubs — no GPU/model download)

Backend

cd backend
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"
pytest
uvicorn manga_translator.api.app:app --reload --port 8000

CLI:

manga-translator translate-batch .\path\to\pages

Web UI

cd web
npm install
$env:NEXT_PUBLIC_API_BASE="http://127.0.0.1:8000"
npm run dev

Open http://localhost:3000 — upload pages, poll job progress, preview side-by-side, download zip.

Optional real models (Windows note)

Torch installs into the long project .venv path can fail with WinError 206. Prefer a short venv:

python -m venv C:\mtvenv
C:\mtvenv\Scripts\python -m pip install -e ".[ml]"
C:\mtvenv\Scripts\python -m pip install -e .
$env:MANGA_TRANSLATOR_REAL_MODELS="1"
$env:MANGA_TRANSLATOR_OCR_BACKEND="manga_ocr"
$env:MANGA_TRANSLATOR_TRANSLATE_BACKEND="opus_mt"
C:\mtvenv\Scripts\python -m uvicorn manga_translator.api.app:app --port 8010

Checklist: docs/CHECKLIST-p1-ui-local-models.md

Translation quality (glossary + replacement dictionaries)

Optional, fully local quality controls modeled on production scanlation tools:

# Glossary: preserve names/terms (only entries present in the page are sent to the LLM)
$env:MANGA_TRANSLATOR_GLOSSARY_PATH="data/dict/glossary.json"       # {"太郎":"Taro"}
# Pre-dict: fix OCR artefacts / strip SFX BEFORE translation (ordered regex rules)
$env:MANGA_TRANSLATOR_PRE_DICT_PATH="data/dict/pre_dict.example.json"
# Post-dict: fix recurring MT phrasing AFTER translation
$env:MANGA_TRANSLATOR_POST_DICT_PATH="data/dict/post_dict.example.json"

Dictionaries are JSON — either {"pattern": "replacement"} or ordered [["pattern","replacement"], ...]; keys are regular expressions. Lines already in the target language are passed through untranslated, and identical bubbles are translated once. Reading order follows manga convention (right-to-left, panel-banded).

Measure accuracy against known translations:

C:\mtvenv\Scripts\python backend/scripts/eval_accuracy.py ollama   # CER (OCR) + chrF (translation)

Layout

backend/          Python package, FastAPI, CLI, tests
web/              Next.js job UI
data/jobs/        Runtime job store (gitignored)
docs/superpowers/ Design spec + implementation plan

License

MIT — compose Apache/MIT model weights; do not embed GPL engines (manga-image-translator / BallonsTranslator) as libraries.