ARIA - Local Manga Translation Tool

ARIA translates Japanese manga pages locally through a browser-based interface. It supports selecting or pasting multiple images, local OCR, optional DeepL translation, text cleanup, and server-side rendering of translated pages.

Current Workflow

1. Add one or more manga pages by selecting, pasting, or dragging images.
2. Start a local processing job.
3. Choose the local text detector and recognizer in the UI.
4. ARIA runs those providers locally and stores the detected regions.
5. ARIA optionally sends only the detected text to DeepL.
6. ARIA refines accepted OCR regions to dark glyph masks inside enclosed light speech bubbles, then removes those glyphs with OpenCV inpainting.
7. ARIA renders the translated text and exposes the original, cleaned, and final images.

Completed pages can be manually cleaned and rerendered without repeating OCR or translation. The renderer removes narrow tails and curved edges from the usable layout area, then chooses the largest wrapped font whose measured pixels fit.

The local backend deliberately stores each intermediate artifact so later work can add OCR correction, translation editing, better masks, and rerendering without repeating every stage.

Cleanup is intentionally conservative. Detector regions rejected by OCR are not cleaned, and a region must have a light, locally enclosed bubble-like background before its dark pixels are masked. This avoids treating sound effects, titles, and artwork as dialogue. A strict OCR-seeded fallback handles enclosed cream and gray bubbles while preserving their original tone. ARIA_MASK_DILATION expands the refined glyph mask; it no longer expands the entire detector polygon.

Bubbles that intentionally open into a panel frame are left alone when their light interior is indistinguishable from the page background. Automatically guessing those regions is more likely to erase side text or panel artwork; a dedicated bubble detector or manual bubble selection is needed to handle them reliably.

Quick Start

1. Install Tesseract

Install Tesseract with Japanese language data. On Windows, the UB Mannheim builds are a practical option. Ensure tesseract.exe is on PATH and that jpn.traineddata is installed.

2. Start the local backend

cd aria-local
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -e .
uvicorn aria_local.main:app --reload --port 8000

The backend API and Swagger documentation are available at:

• http://127.0.0.1:8000/health
• http://127.0.0.1:8000/docs

3. Start the frontend

cd aria-frontend
npm install
npm run dev

The frontend defaults to http://127.0.0.1:8000. To use another backend, copy aria-frontend/.env.example to .env.local and change VITE_API_URL.

Translation Providers

Without DEEPL_API_KEY, ARIA uses the Japanese source text as a placeholder translation. This allows the local OCR, cleanup, and renderer stages to be tested without an external service.

Text detection, Japanese recognition, image cleanup, and rendering stay on the local machine. Argos Translate is also local. Only DeepL translation leaves the machine when explicitly enabled.

The UI offers Argos Translate as a no-key Japanese-to-English alternative. It is expected to be less natural than DeepL for manga dialogue, and its model package is downloaded the first time it is selected.

Helsinki OPUS-MT (Helsinki-NLP/opus-mt-ja-en) is also available as a local option and usually produces higher-quality translations than Argos. When a DeepL key is configured, the frontend prefers DeepL for translation quality; Helsinki remains the preferred local option.

To use DeepL:

$env:DEEPL_API_KEY = "your-deepl-key"
$env:ARIA_TRANSLATION_PROVIDER = "deepl"
uvicorn aria_local.main:app --reload --port 8000

The translation provider is intentionally isolated from OCR and image processing so a future local translation model can be added without changing the API or frontend.

Useful Settings

$env:ARIA_OCR_LANG = "jpn"
$env:ARIA_OCR_MIN_CONFIDENCE = "25"
$env:ARIA_TEXT_DETECTOR = "tesseract"
$env:ARIA_TEXT_RECOGNIZER = "tesseract"
$env:ARIA_PADDLEOCR_ENABLE_MKLDNN = "0"
$env:ARIA_PADDLEOCR_LIMIT_SIDE_LEN = "960"
$env:ARIA_MASK_DILATION = "5"
$env:ARIA_FONT_PATH = "C:\Windows\Fonts\arial.ttf"
$env:ARIA_DATA_DIR = ".aria-data"
$env:TESSERACT_CMD = "C:\Program Files\Tesseract-OCR\tesseract.exe"

For pages that are primarily vertical Japanese, install jpn_vert.traineddata and try ARIA_OCR_LANG=jpn_vert.

Architecture

The project has one backend path:

• aria-local: the Python backend for OCR, translation, image cleanup, and rendering

The frontend talks directly to this local backend. The CPU-heavy pipeline stays in the same Python process as the provider adapters, avoiding an extra remote runtime and serialization step.

Roadmap

• Replace the default provider adapters with manga-specific detection and masks
• Use manga-ocr with PaddleOCR or a manga-specific detector for higher-quality Japanese OCR
• Improve speech-bubble expansion and inpainting
• Add editable OCR and translation regions
• Add page ordering and chapter-level translation context
• Add stronger local translation providers such as Ollama/Qwen
• Add Docker packaging for easier installation