Comparison

Koharu vs manga-image-translator: which open source manga translator should you use?

Use Koharu if you want a desktop app with an editor: install it on Windows, macOS or Linux, translate, fix the text on a canvas and export layered PSD files. Use manga-image-translator if you want a Python program to script batch jobs from the command line or an API. Both are free and both run best on a GPU.

Table of contents
  1. Side by side
  2. What is Koharu?
  3. What is manga-image-translator?
  4. Which one is easier to install?
  5. What hardware do they need?
  6. Which one gives better translations?
  7. If you make or publish the comic
  8. Which one fits

Side by side

Koharu and manga-image-translator
Koharumanga-image-translator
What it issourceDesktop app with a project browser, canvas and editor, written in RustPython program with a command line, a local web page and an API
InstallsourceRelease builds for Windows, macOS and Linux, or winget and HomebrewPython 3.12 or later with pip, or Docker
HardwaresourceCUDA, ROCm, Metal on Apple silicon or Vulkan. CPU is "substantially slower"Runs on CPU, GPU with a PyTorch build for your CUDA version
TranslationsourceLocal models through llama.cpp, or OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek, OpenRouter, DeepL, Google Cloud, CaiyunOnline services (DeepL, ChatGPT, Gemini, DeepSeek, Papago and more) or offline models (Sugoi, NLLB, M2M100 and more)
LanguagessourceVertical Japanese and right-to-left text in typesettingMainly Japanese, plus Chinese, English and 20 other languages
EditingsourceProofreading view and a canvas for cleanup and text placementNo editor. Save the text to a file, edit it, load it back
ExportsourceFlat PNG, CBZ or layered PSDPNG, WEBP, JPG, or XCF, PSD and PDF through GIMP
LicencesourceMIT or Apache 2.0GPL-3.0
ActivitysourceVersion 0.87.0, released 20 November 2026Last commit 18 November 2026

Koharu and manga-image-translator are the two most starred open source manga translators on GitHub (6,610 and 10,578 stars on 8 November 2026). Both find the text on a page, read it, erase it, translate it and set the translation back in. They differ in how you work with them: Koharu is an app you click through, manga-image-translator is a program you configure and run. For the paid and browser tools, see the best manga translators.

What is Koharu?

Koharu calls itself an "ML-powered manga translator, written in Rust". It is a desktop app for Windows, macOS and Linux. You make a project, import images, archives or PDFs, and choose which stages to run: detection, OCR, translation and inpainting, on one page or the whole project. A proofreading view lets you correct the OCR and the translation, and a canvas lets you clean up and place text by hand. Typesetting fits the text to the bubble, falls back to other fonts, and handles vertical Japanese and right-to-left text. For what that step involves by hand, see manga typesetting.

You can run every model on your own machine, so pages never leave it, or connect a hosted provider for the translation step. Koharu stores provider keys in the system's credential store. It exports a flat PNG, a CBZ archive, or a layered PSD for further work in Photoshop. The docs warn that PSD is an interchange format and that you should keep the Koharu project folder as your editable source.

What is manga-image-translator?

manga-image-translator is a Python project that aims to translate images "that are unlikely to be professionally translated", in its own words. You run it in local batch mode on a folder, as a web server on your own machine, or as an API. Every step is a setting: the text detector, the OCR model, the inpainter (the README recommends lama_large), the translator, an upscaler and a colorizer. Its inpainter does the job a cleaner does by hand (see manga cleaning). It also reads a glossary for the OpenAI translator, and pre and post dictionaries that fix common OCR errors and mistranslations with regular expressions.

It has no editor. To fix a line you save the extracted text and the translation to a file, change it, and render again from the file. With GIMP installed it writes XCF files with the original, the inpainting and each text box on its own layer. Its PSD output also goes through GIMP, which converts the text layers to plain image layers, so the text is no longer editable in Photoshop. The README says the project is "still in the early stages of development and has many shortcomings".

Which one is easier to install?

Koharu. You download an installer, or run winget install koharu on Windows or brew install --cask koharu on macOS. The first launch downloads its runtime, and each model downloads when you first use it. manga-image-translator needs Python 3.12 or later, a pip install of its requirements (with a C++ compiler for one dependency), and the right PyTorch build if you want GPU speed. Docker is the easier path for it. A separate Rust rewrite exists as a compiled command line tool.

What hardware do they need?

Koharu picks a device at start-up: Metal on Apple silicon, then CUDA, ROCm, Vulkan or the CPU. CUDA needs an NVIDIA Turing-class or newer card. The editor canvas uses WebGPU, so it needs a current graphics driver even when the models run on the CPU. manga-image-translator runs on the CPU by default and uses an NVIDIA GPU when you install a PyTorch build for your CUDA version. If you translate with a hosted model, the heavy part of the work moves off your machine in both tools.

Which one gives better translations?

That depends on the translator you plug in, more than on the tool. Both can use the same large hosted models, such as OpenAI, Gemini or DeepSeek, and both can run local models. manga-image-translator recommends the offline Sugoi model for Japanese to English. Koharu can send the page image along with the text, so a model that reads images can use the art as context. We have not run a test of our own between the two, so we do not rank their output. Neither tool lists sound effect redraws, so a page full of drawn effects such as the ones in our Japanese sound effects guide still needs hand work.

If you make or publish the comic

Both tools are a good choice for a technical translator who wants control over each step. If you own the pages and want a release instead of a project to maintain, Bublate is a web app that letters the bubbles, redraws sound effects, keeps names consistent across chapters with a series bible, and exports a layered PSD or platform-ready ZIP. It costs $3 for a webtoon chapter of up to 20 segments, with 30 free credits. Manga and other print pages cost up to $0.60 a page. See Bublate for translators and letterers, how to typeset manga and how to clean manga.

Which one fits

Pick Koharu if you want to see and fix every page before you export it, or if you need layered PSD files for Photoshop. Pick manga-image-translator if you want to run hundreds of pages through a script, call it from your own code, or tune each step (detector, OCR, inpainter, translator) in a config file. Both are free, and both need a person to check the output.

Frequently asked questions

Is Koharu free?

Yes. Koharu is free and open source under the MIT or Apache 2.0 licence, your choice. You download a release from GitHub for Windows, macOS or Linux. The only costs are your own hardware and any hosted translation provider you connect, such as OpenAI or DeepL, which bills you directly.

How do you use the Koharu manga translator?

Install a release, wait for the first launch to download its runtime, then create a project and import images, an archive or a PDF. Run detection, OCR, translation and cleanup on a page or the whole project, check the text in the review view, and choose File, then Export PNG or Export PSD.

How do you use manga-image-translator?

Clone the repository, install the requirements with pip (Python 3.12 or later) or run the Docker image, then point it at a folder of images. Results go to a new folder next to the source. You can also start its web mode or its API mode and send images to it on your own machine.

Is there a manga-image-translator online version?

The README links a demo at touhou.ai/imgtrans, which opens Cotrans, and a browser script. Cotrans loaded on 20 November 2026, but its status API showed 0 workers and 30 images waiting. A shared demo can be slow or offline, so install it locally for real work.

Do Koharu and manga-image-translator keep my pages private?

With local models, yes: the pages and text stay on your computer. Koharu says hosted providers receive the request content and apply their own retention terms, and it sends crash reports to Sentry. manga-image-translator sends text to whichever online translator you configure.

Which is better for webtoons?

Neither project documents a special mode for long strips. Koharu works page by page inside a project. manga-image-translator has options for detection size, inpainting size and upscaling that its README says to tune for image resolution. Test a few strips of your series in both before you commit.

Sources

  1. Koharu, README on GitHub · checked
  2. Koharu, releases on GitHub · checked
  3. Koharu docs, install · checked
  4. Koharu docs, hardware and runtimes · checked
  5. Koharu docs, connect a translation provider · checked
  6. Koharu docs, export PNG and PSD · checked
  7. Koharu docs, privacy · checked
  8. manga-image-translator, README on GitHub · checked
  9. Cotrans web front end · checked

Try Bublate for free

Continue with Google

or