Definition
Comic-Translate is a desktop app that automatically translates comics: BDs, manga, manhwa, fumetti and many other formats (Image, PDF, Epub, CBR, CBZ). It uses SOTA LLMs such as GPT-4, Claude and Gemini for the highest translation quality.
The pipeline is detection, then OCR, translation, inpainting, and a final text render. The detector is RT-DETR-v2, trained on 11k images of manga, webtoons, and Western comics. OCR then selects manga-ocr, Pororo, or PPOCRv5 from the language. GPT-4.1, Claude-4.5, and Gemini-2.5 translate from the full page text. LaMa or AOT-GAN does the inpainting. The app is Python with a PySide6 GUI and an Apache-2.0 license.
- Repository: https://github.com/ogkalu2/comic-translate
- Website: https://www.comic-translate.com
- License: Apache-2.0
Key Metrics
- Stars: 2,500
- Forks: 277
- Language: Python 100%
- GUI: PySide6
Pipeline
┌──────────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌───────────┐
│ Detection │→ │ OCR │→ │ Translation│→ │ Inpainting│→ │ Render │
│ (RT-DETR-v2) │ │(Multi) │ │ (LLM) │ │ (LaMa/ │ │ (Text) │
│ │ │ │ │ │ │ AOT-GAN) │ │ │
└──────────────┘ └───────────┘ └───────────┘ └───────────┘ └───────────┘
Detection & OCR
Detection
- comic-text-and-bubble-detector (RT-DETR-v2): trained on 11k images (manga, webtoons, Western comics)
- Algorithmic segmentation based on the detected boxes
OCR (Auto-selected by language)
| Language | OCR Engine |
|---|---|
| Japanese | manga-ocr (kha-white) |
| Korean | Pororo |
| Other | PPOCRv5 (PaddleOCR) |
OCR (Optional - any language)
- Gemini 2.0 Flash
- Microsoft Azure Vision
Translation
Uses SOTA LLMs for better translation quality than Google Translate/DeepL:
- GPT-4.1 (OpenAI)
- Claude-4.5 (Anthropic)
- Gemini-2.5 (Google)
All LLMs receive the full page text to aid translation. An option supplies the image as well for extra context.
Inpainting
- LaMa (Anime/Manga finetuned) - from dreMaz
- AOT-GAN - from zyddnys
Input Formats
| Format | Support |
|---|---|
| Images | PNG, JPG, WebP, etc. |
| Yes | |
| Epub | Yes |
| CBR | Yes (needs WinRAR/7-Zip) |
| CBZ | Yes |
Source Languages
English, Korean, Japanese, French, Simplified Chinese, Traditional Chinese, Russian, German, Dutch, Spanish, Italian
Installation
Download (Recommended)
Download the Windows/macOS installer from comic-translate.com
From Source
git clone https://github.com/ogkalu2/comic-translate
cd comic-translate
uv init --python 3.12
uv add -r requirements.txt --compile-bytecode
# NVIDIA GPU
uv pip install onnxruntime-gpu
# Run
uv run comic.py
Features
| Feature | Description |
|---|---|
| Automatic mode | One-click translation |
| Manual mode | Corrections for detection/OCR issues |
| Zoom | Ctrl + Mouse Wheel |
| Navigation | Arrow keys, trackpad gestures |
| Batch processing | Multiple images/files |
| Undo | Undo individual image corrections |
Tech Stack
| Component | Technology |
|---|---|
| Language | Python |
| GUI | PySide6 |
| Detection | RT-DETR-v2 |
| OCR | manga-ocr, Pororo, PaddleOCR |
| Translation | GPT-4.1, Claude-4.5, Gemini-2.5 |
| Inpainting | LaMa, AOT-GAN |
Pros
| Pro | Description |
|---|---|
| Best translation quality | SOTA LLMs for translation |
| Multi-format | Image, PDF, Epub, CBR, CBZ |
| Desktop app | Nice, easy-to-use GUI |
| Manual mode | Can fix automatic mistakes |
| Multi-comic types | BD, Manga, Manhwa, Fumetti |
| Pre-built installer | No complicated setup |
| Apache-2.0 | Permissive license |
Cons
| Con | Description |
|---|---|
| LLM API required | Needs an API key (OpenAI/Anthropic/Google) |
| API costs | Cost of API calls |
| GPU only from source | GPU acceleration only when running from source |
| Python 3.12 required | Specific Python version |
When to Use
- Best translation quality: when you need the best translation quality (LLM-based)
- Multi-format comics: when you need to handle many formats (PDF, Epub, CBR, CBZ)
- Desktop workflow: when you want an easy-to-use GUI app
- Non-manga comics: BDs, Western comics, fumetti
- Manual correction: when you want control over the translation process
Use this desktop app when LLM quality matters more than Google Translate or DeepL. Automatic mode translates in one click. Manual mode corrects detection or OCR mistakes. Zoom is Ctrl plus the mouse wheel, and arrow keys or trackpad gestures move between pages. Batch mode takes several images or files. Undo rolls back one image correction. Inputs include PNG, JPG, WebP, PDF, Epub, CBR, and CBZ, and CBR needs WinRAR or 7-Zip. Source languages include English, Korean, Japanese, French, Simplified Chinese, Traditional Chinese, Russian, German, Dutch, Spanish, and Italian. An OpenAI, Anthropic, or Google API key is required, and the calls have a cost. GPU support appears only from source on Python 3.12.
References:
See also: Manga Image Translator · Koharu · Translator Comparison