Skip to main content
Koharu is a local-first desktop application that takes manga from raw scan to finished translation without sending your images to a cloud pipeline. It runs ML models directly on your hardware to detect text regions, read source text with OCR, generate translations through local GGUF models or your chosen hosted provider, and reconstruct the artwork beneath each text block with generative inpainting. Every stage of that pipeline stays inspectable and editable — you can revise OCR output, swap a translation, or repaint a single region without reprocessing the whole project.

Install Koharu

Download and set up Koharu on Windows, macOS, or Linux.

Translate Your First Project

Walk through the complete pipeline from import to export.

Workflow Overview

Learn how projects, pages, and the editing canvas fit together.

Koharu Agent

Use the agent interface to inspect and drive projects programmatically.

​
What Koharu Does

Koharu organises manga translation into a sequential pipeline. Each stage produces output you can review and edit before the next stage begins. You can run the full pipeline in one pass or open any individual stage to revise its results.
1

Import and organise

Create a project and import pages as individual PNG, JPEG, or WebP images; as CBZ, ZIP, or RAR archives; or as PDFs. Koharu sequences the pages and tracks all project data locally in a single project file.
2

Detect text regions

The Koharu Layout RF-DETR Seg 2XL model scans each page and identifies text regions, speech bubbles, and segmentation masks. Detection runs at page scope or across the full project in one action.
3

Read source text with OCR

An OCR model reads the source text from each detected region. Koharu supports four OCR engines — PaddleOCR VL 1.6, Manga OCR, Baberu OCR, and Hayai OCR — so you can choose the engine that performs best on your source material. The OCR output is fully editable before translation begins.
4

Translate

Koharu sends the recognised text to your configured translation engine. Run inference locally with any GGUF-format language model, connect a hosted LLM provider (OpenAI, Gemini, Claude, Grok, DeepSeek, and others), or use a dedicated machine-translation service such as DeepL or Google Cloud Translation.
5

Inpaint the artwork

Before rendered translations appear, Koharu removes the source lettering and reconstructs the artwork behind it. Four inpainting models are available — FLUX.2 Klein, RORem Mixed, LaMa, and AOT GAN — ranging from fast structural fill to high-quality generative reconstruction.
6

Typeset and export

Place translated text on the WebGPU-based canvas with automatic fitting, font fallback, vertical CJK support, and right-to-left layout. Export finished pages as flattened PNGs or as layered PSDs for further editing in any image editor.

​
Hardware Support

Koharu accelerates ML inference using the GPU backend available on your machine. Keep your graphics driver current; a full CUDA or ROCm SDK installation is not required.
CUDA 13.0 is supported on NVIDIA Turing-class or newer GPUs with an R580 or newer driver on Windows and Linux. Install the latest NVIDIA driver and verify your GPU against NVIDIA’s CUDA toolkit, driver, and architecture matrix.

​
Privacy and Local-First Design

Koharu runs all vision models and — when you choose a local GGUF model — all LLM inference directly on your machine. Your page images and project data never leave your computer unless you explicitly configure a hosted provider. When you do connect a hosted provider, it receives only the data required for that specific request (for example, recognised text sent for translation), not your raw images or full project.
Koharu is available on Windows, macOS (Apple Silicon), and Linux. Download the latest release from the GitHub releases page, or install with your system package manager — winget install --id mayocream.koharu on Windows, brew install --cask koharu on macOS.