Fork of Koharu a ML-powered manga translator, written in Rust.
This Fork contains experimental features not implemented, planned or rejected by the original Koharu team. The list of added features can be found here. And refer the Fork Docs for documentation of the added features.
The current release is based on the major version v0.83.X
Important
If you have an issue that happens with my fork, provided from this repository, that does not happen on the original koharu app, please do not open a bug report it on their bug trackers.
Instead, open an issue on this repository's issue tracker: https://github.com/gray-rice/koharu-ex/issues
I am not affiliated with koharu-rs and the original creator Mayo Takanashi
· Official Docs · Fork Docs ·
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. It also supports third party providers for translation, which are subject to their terms and conditions.
- Better error handling and reporting for Translation: Prevent complete failure of translation pipeline by skipping the specific page. The Old implementation would fail without any indication of progress before failure.
- Translation pipeline summary: Pipeline summary with errors for failed pages, success/failed counts and retry for failed pages.
- Translation Import/Export: JSON-based translation exchange with page-level granularity for collaborative workflows.
- Project Archiving: Organise projects easily with archives.
- Rectangle Selection for Eraser & Remove Tools: Precise region selection for targeted cleanup and inpainting.
- Multi Layer select: Drag to select multiple layers.
- Fine grained export options: Options for individual/group exports of selected pages without having to export entire project.
- 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
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 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 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 is available on Apple silicon Macs.
Vulkan is available on Windows and Linux as an alternative to CUDA and ROCm / HIP.
The editor canvas uses WebGPU and requires a current graphics driver even when inference runs on the CPU.
CPU inference is available for supported workloads but is substantially slower.
Koharu uses separate models for detection, OCR, inpainting, and translation. Vision and inpainting and translation and generation have separate model settings.
Detection, OCR, and inpainting models are selected separately.
The detection model finds text regions, speech bubbles, and segmentation masks.
OCR reads source text from detected regions.
Inpainting reconstructs the image behind source text before the translation is rendered.
Translation can use a local language model or a remote API.
- 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
- 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
Hosted LLM providers: OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek, and OpenRouter.
Machine-translation providers: DeepL, Google Cloud Translation, and Caiyun.
OpenAI-compatible endpoints are also supported.
Builds are only available for Linux and Windows.
Note
I have no plans to create prebuilt binaries for MacOS.
Download release builds from the releases page. Installation requirements and first launch (official docs) vary by operating system.
MacOS users are requested to refer Development section and build their packages.
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.exePlatform dependencies and validation commands for local builds are listed in Development Setup.
bun installbun devbun run buildThe executable is written to target/release.
If Koharu is useful in your workflow, consider sponsoring the original project and its creators.
Thanks to the original creator Mayo Takanashi and contributors !
The original project was dual-licensed under MIT and Apache 2.0. This fork exercises the option to distribute the software and all subsequent modifications strictly under the MIT License."
