Automate manga translation using local machine learning for text detection, OCR, inpainting, and text rendering.

Koharu is an open-source desktop application for translating manga that automates the process of text detection, removal, and replacement. It provides a local-first workflow for users to translate Japanese comics into other languages while keeping data on their own hardware. The software is available for Windows, macOS, and Linux, offering a comprehensive pipeline for the scanlation process.
The application combines several machine learning stages to handle the complexities of comic layouts. It identifies speech bubbles, performs optical character recognition, and uses inpainting to clear original text before rendering translations. Users can interact with the software through a graphical user interface or run it in a headless mode via a local HTTP API, allowing for integration into broader automation workflows.
Built with Rust and Tauri, Koharu uses candle and llama.cpp for high-performance inference. The tool is designed for translators and manga enthusiasts who require precise control over typography, offering a dedicated text renderer that handles Unicode-aware shaping and script-aware line breaking. This ensures that translated text fits naturally inside irregular speech bubbles and captions. It supports a wide array of vision models for layout analysis and inpainting, such as anime-text-yolo and lama-manga, as well as various large language models for the translation phase. Users can either use built-in Google Fonts or provide their own OpenType and TrueType fonts, including variable fonts.
To ensure accessibility across different hardware, Koharu includes a CPU fallback and supports multiple GPU backends. On Windows and Linux, it bundles the CUDA Toolkit for NVIDIA users, while Apple Silicon users can leverage Metal. The software also supports experimental ZLUDA acceleration for AMD GPUs on Windows.
Koharu serves as a specialized utility for the manga scanlation community, prioritizing privacy and local execution.
A cross-platform utility for text translation and recognition using optical character recognition and selected text.
A self-hosted machine translation API that operates offline to provide translation services with complete privacy control.
Join our newsletter to get shiny new open source software delivered to your inbox. Unsubscribe anytime.