koharu-rs/koharu

What it solves

Koharu automates the complex, multi-step process of translating manga. It streamlines the workflow of detecting text in images, extracting that text, translating it, and reconstructing the original artwork so the new text can be placed naturally.

How it works

Koharu uses a local-first pipeline that runs vision and language models directly on your machine to ensure data privacy. It coordinates several specialized AI tasks: it uses detection models to identify speech bubbles and text regions, OCR models to read the source text, and generative inpainting models to remove the original text and rebuild the background. Finally, it uses LLMs (either local GGUF models or remote APIs) to perform the translation, providing a web-based canvas for final proofreading and typesetting.

Who it’s for

This tool is designed for manga readers, translators, and editors who want a professional-grade, local-first application to handle the technical heavy lifting of manga localization.

Highlights

• Local-first architecture that keeps all data processing on your machine.
• Supports a wide range of hardware acceleration including CUDA, ROCm, Metal, and Vulkan.
• Integrated generative inpainting to reconstruct artwork after removing source text.
• Flexible pipeline that allows users to choose specific models for detection, OCR, inpainting, and translation.
• Built-in web-based editor for manual cleanup, text placement, and final review.