- Rust 72.8%
- TypeScript 12.4%
- C 7.6%
- MDX 4.1%
- C++ 2.8%
- Other 0.2%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| .agents | ||
| .cargo | ||
| .github | ||
| .vscode | ||
| crates | ||
| packages | ||
| scripts | ||
| .editorconfig | ||
| .gitignore | ||
| .oxfmtrc.json | ||
| AGENTS.md | ||
| bun.lock | ||
| Cargo.lock | ||
| Cargo.toml | ||
| CHANGELOG.md | ||
| cliff.toml | ||
| CONTRIBUTING.md | ||
| LICENSE-APACHE | ||
| LICENSE-MIT | ||
| package.json | ||
| README.md | ||
Koharu
ML-powered manga translator, written in Rust.
Getting Started · Docs · Bug reports · Discord
Koharu introduces a local-first workflow for manga translation, utilizing the power of ML to automate the process. It combines the capabilities of object detection, OCR, inpainting, and LLMs to create a seamless translation experience.
Note
Koharu runs its vision models and LLMs locally on your machine to keep your data private and secure.
Note
Join our Discord server for support and discussion.
Features
- Multi-format project management for raster images, archives, and PDFs with page sequencing
- Selective pipeline for detection, OCR, translation, and inpainting at page or project scope
- Detection and segmentation for text regions, speech bubbles, and cleanup regions
- Multimodal OCR for dialogue, captions, and general page text
- Local GGUF inference and hosted providers for LLM and machine-translation workflows
- Generative inpainting for source-text removal and artwork reconstruction
- Proofreading for correcting OCR and translation output
- WebGPU-based canvas for manual cleanup, text placement, and page composition
- Multilingual text shaping and layout with automatic fitting, font fallback, vertical CJK, and right-to-left text
- Layered PSD export for flattened delivery and layered editing
- Agent-based workflow for project inspection, editing, and pipeline control
Hardware Acceleration
Koharu supports CUDA and ROCm / HIP on Windows and Linux, Metal on Apple silicon, and Vulkan on Windows and Linux. Keep your graphics driver current; a full CUDA or ROCm SDK installation is not required. See Runtime and hardware requirements for model-specific guidance.
CUDA
CUDA 13.3 requires an NVIDIA Turing-class or newer GPU and an R610 or newer driver. Check NVIDIA's official CUDA toolkit, driver, and architecture matrix and install the latest NVIDIA driver.
ROCm / HIP
ROCm 10.0 support depends on the exact AMD GPU, operating system, and driver combination. Check AMD's official ROCm 10.0.0 compatibility matrix and install a compatible AMD driver.
Metal
Metal is available on Apple silicon Macs.
Vulkan
Vulkan is available on Windows and Linux as an alternative to CUDA and ROCm / HIP.
WebGPU
The editor canvas uses WebGPU and requires a current graphics driver even when inference runs on the CPU.
CPU
CPU inference is available for supported workloads but is substantially slower.
Machine Learning Models
Koharu uses separate models for detection, OCR, inpainting, and translation. Vision and inpainting and translation and generation have separate model settings.
Computer Vision Models
Detection, OCR, and inpainting models are selected separately.
Detection and Layout
The detection model finds text regions, speech bubbles, and segmentation masks.
OCR
OCR reads source text from detected regions.
Inpainting
Inpainting reconstructs the image behind source text before the translation is rendered.
Large Language Models
Translation can use a local language model or a remote API.
General-Purpose Local Models
- LFM 2.5: lfm2.5-1.2b-instruct
- Ministral 3: ministral-3-8b-instruct
- Gemma 4: gemma4-e2b-it, gemma4-e4b-it, gemma4-12b-it, gemma4-26b-a4b-it, gemma4-31b-it
- Qwen 3.5: qwen3.5-0.8b, qwen3.5-2b, qwen3.5-4b, qwen3.5-9b, qwen3.5-27b, qwen3.5-35b-a3b
- Qwen 3.6: qwen3.6-27b, qwen3.6-35b-a3b
- Qwen 3.8: qwen3.8-27b
Uncensored Local Models
- Gemma 4 uncensored: gemma4-e2b-uncensored, gemma4-e4b-uncensored, gemma4-12b-uncensored, gemma4-26b-a4b-uncensored, gemma4-31b-uncensored
- Qwen 3.5 uncensored: qwen3.5-2b-uncensored, qwen3.5-4b-uncensored, qwen3.5-9b-uncensored
- Qwen 3.6 uncensored: qwen3.6-27b-uncensored, qwen3.6-35b-a3b-uncensored
- Qwen 3.8 uncensored: qwen3.8-27b-uncensored
Cloud Providers
Hosted LLM providers: OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek, and OpenRouter.
Machine Translation Providers
Machine-translation providers: DeepL, Google Cloud Translation, and Caiyun.
OpenAI-Compatible Providers
OpenAI-compatible endpoints are also supported.
Installation
Download release builds from the releases page. Installation requirements and first launch vary by operating system.
Builds are available for Windows, macOS, and Linux.
WinGet
Install on Windows with winget:
winget install koharu
Homebrew
Install on macOS with Homebrew:
brew install --cask koharu
Troubleshooting
Startup, runtime, model, and provider errors are covered in Troubleshooting. Set RUST_LOG to debug or trace for verbose logs:
# macOS / Linux
RUST_LOG=debug koharu
# Windows (PowerShell)
$env:RUST_LOG="debug"; koharu.exe
Development
Platform dependencies and validation commands for local builds are listed in Development Setup.
Prerequisites
Install dependencies
bun install
Development
bun dev
Build
bun run build
The executable is written to target/release.
Sponsorship
If Koharu is useful in your workflow, consider sponsoring the project.
Contributors ❤️
Thanks to all the contributors who have helped make Koharu better!
License
Copyright 2025-2026 Mayo Takanashi and Koharu contributors.
Koharu is dual-licensed under the MIT License or the Apache License, Version 2.0, at your option.