How to install Koharu on Windows with NVIDIA GPU support?

Download the latest release, ensure current NVIDIA drivers are installed, and Koharu will automatically bundle CUDA Toolkit 13.0 on first run. Check the GPU acceleration section for compatibility with compute capability 7.5 or higher.

Koharu vs online manga translators: which is better for speed and privacy?

Koharu prioritizes privacy by running locally, but it may be slower due to model inference on your hardware. Online tools are faster but send data to the cloud, so choose based on your privacy and performance needs.

Can Koharu translate manga from multiple source languages?

Yes, Koharu supports various languages through its LLM backends, including local models like Sakura-GalTransl and cloud APIs. Configure the provider in settings for different language pairs, as mentioned in the LLM section.

How to use Koharu with my own custom LLM models?

You can integrate self-hosted LLMs via the OpenAI-compatible API support, such as LM Studio, by setting a custom base URL in the configuration. Local GGUF models are also downloadable on demand from Hugging Face.

Does Koharu support batch processing for multiple manga pages?

While the GUI focuses on single pages, the headless mode and HTTP API allow automation for batch workflows. Use the --headless flag and API endpoints to process multiple pages sequentially, as described in the headless mode section.

What are the minimum system requirements to run Koharu smoothly?

A modern CPU and GPU with at least 8GB of RAM are recommended for basic models. For best performance, use an NVIDIA GPU with compute capability 7.5+, Apple Silicon, or AMD with HIP SDK, as detailed in the acceleration guides.

Open-Awesome

koharu

Apache-2.0Rust0.83.5

A local-first, ML-powered desktop application for translating manga, built in Rust with automated text detection, OCR, inpainting, and LLM translation.

Visit Website GitHub

6.4k stars463 forks0 contributors

What is koharu?

Koharu is an open-source desktop application that automates manga translation using machine learning. It combines object detection, OCR, inpainting, and LLM-based translation into a single local-first workflow, allowing users to translate manga pages while keeping their data private. The application is built in Rust for performance and safety.

Target Audience

Manga translators, scanlation groups, and enthusiasts who need an efficient, privacy-focused tool for translating manga without relying on cloud services. It's also suitable for developers interested in ML applications for creative workflows.

Value Proposition

Developers choose Koharu for its fully local execution, which ensures data privacy, and its comprehensive pipeline that handles the entire translation process automatically. Its use of Rust and modern ML frameworks provides high performance and reliability compared to manual or cloud-dependent alternatives.

Overview

ML-powered manga translator, written in Rust.

Use Cases

Best For

Translating manga pages with complex layouts and speech bubbles
Maintaining data privacy by running all ML models locally
Automating the text detection and inpainting steps in scanlation
Integrating manga translation into automated workflows via its API
Exporting translated pages as editable PSD files for further refinement
Using local LLMs for translation without internet dependency

Not Ideal For

Projects requiring cloud-based translation for accessibility across multiple devices
Users with low-end or GPU-less hardware unable to run local ML models efficiently
Teams needing real-time collaborative editing features for scanlation workflows

Pros & Cons

Pros

Local Privacy Focus

All vision and language models run locally on the user's machine, ensuring data never leaves the device, as emphasized in the philosophy and note sections.

End-to-End Automation

Combines text detection, OCR, inpainting, and LLM translation into a single pipeline, streamlining the entire manga translation process as described in the key features.

Advanced Text Rendering

Supports vertical CJK layout, RTL scripts, and precise font fallback, making translated text fit naturally into speech bubbles, detailed in the text rendering section.

Editable PSD Export

Exports pages as layered Photoshop files with editable text layers, facilitating manual refinement, which is highlighted in the export and features sections.

Cons

High Hardware Demands

Requires powerful GPUs and significant memory for local model inference, with dependencies on CUDA, Metal, or Vulkan, limiting accessibility on older systems as noted in the GPU acceleration section.

Complex Setup Process

Involves downloading large models, configuring GPU drivers, and managing dependencies like CUDA Toolkit or AMD HIP SDK, making initial installation cumbersome, as outlined in the installation prerequisites.

Limited Collaborative Features

Lacks built-in cloud sync or team collaboration tools, focusing on local-first workflows, which may hinder use in group scanlation projects despite the API for automation.

Frequently Asked Questions

Related Projects

clash-verge-rev

A modern GUI client based on Tauri, designed to run in Windows, macOS and Linux for tailored proxy experience

A modern runtime for JavaScript and TypeScript.

Stars108,721

Forks6,458