Review
Koharu manga translator review: what it does and how to use it
Koharu is a free, open source desktop app that finds, reads, translates and cleans manga text on your own computer, then exports PNG, CBZ or layered PSD files. It fits people with a recent graphics card who want control and local models. The limits: you install and run the models yourself, its docs do not cover sound effects, and translation instructions apply to every project.
Table of contents
At a glance
| Developer | Mayo Takanashi and Koharu contributors. Code on GitHub under koharu-rssource |
|---|---|
| Price | Free. Dual licence, MIT or Apache 2.0. Hosted translation services you connect bill you on their own termssource |
| Platforms | Desktop app for Windows, macOS on Apple silicon and Linux. Install with WinGet, Homebrew or a release file. No web app, no phone appsource |
| Hardware | CUDA on NVIDIA Turing or newer, ROCm on some AMD cards, Metal on Apple silicon, Vulkan, or a slower CPU mode. The editor needs a working WebGPU driversource |
| Languages | You pick the target language in the app. The docs do not list them. Two small local models (LFM 2.5, Ministral 3) support only some targetssource |
| Files | Import PNG, JPEG, WebP, CBZ, ZIP, RAR and PDF. Export flat PNG, CBZ or layered PSDsource |
| Sign-in | None for the app. Hosted models need your own API key, or a ChatGPT sign-in, or a Grok sign-in with SuperGrok or X Premiumsource |
| Data | Local models keep text on your machine. Hosted providers receive the request content. Error reports go to Sentrysource |
| Last release | Version 0.87.0, 20 November 2026, after 5 releases on 19 and 20 Novembersource |
Koharu is the open source manga translator that comes up most in search after the big reader tools. Its GitHub repository had about 6,750 stars on 20 November 2026, and it ships releases often: 5 between 19 and 20 November 2026. This review covers what Koharu does, how to use it step by step, what it needs from your computer, and where a different tool is the better choice. Every fact links to Koharu's own README or docs, opened on 8 November 2026.
What is Koharu?
Koharu calls itself an "ML-powered manga translator, written in Rust." It is a desktop app. It runs four stages on a page: detection (text regions, speech bubbles and panels), OCR (reading the source text), translation, and inpainting (removing the source lettering and repairing the art). Then it sets the translation back into the page, and you export a flat PNG or a layered PSD.
The main idea is "local-first." The README says Koharu "runs its vision models and LLMs locally on your machine." You can also connect a hosted model. The provider docs list OpenAI, Gemini, Claude, Grok, MiniMax, DeepSeek and OpenRouter, plus DeepL, Google Cloud Translation and Caiyun for machine translation, and LM Studio or any OpenAI-compatible server.
How do you use the Koharu manga translator?
These steps follow Koharu's own quickstart, with its menu names in bold.
- Install Koharu. Download the release for your system from GitHub, or run
winget install koharuon Windows orbrew install --cask koharuon macOS. The first launch downloads runtime packages, and each model downloads the first time you use it. - Create a project. On the start screen, type a name under New project and choose Create. Koharu makes a
.khrprojfolder under Documents/Koharu. - Import pages. Choose File > Import Pages > Files, or Folder. Images, CBZ, ZIP, RAR and PDF files work. Rename a .cbr file to .rar first.
- Pick the output. Open the processing selector above the canvas and choose a translation model, a target language and any instructions. For a hosted model, save its key first in Settings > Providers.
- Run the pipeline. Choose Current page, select all four stages and click Run.
- Review the text. Select a text layer and compare the source with the translation in the inspector. Paint over missed lettering with Remove, and use Brush, Eraser or Text for touch-ups.
- Fit the lettering. Set the font, size, automatic fitting, alignment, writing direction, fill and stroke.
- Export. Choose File > Export PNG or File > Export PSD and pick a folder.
Test on one page before a whole chapter. Koharu's docs give the same advice for translation settings: change one thing at a time.
What hardware does Koharu need?
Koharu picks a device at startup. Per its hardware page, CUDA needs an NVIDIA Turing-class or newer card with an R610 or newer driver, ROCm depends on the exact AMD card and driver, Metal works on Apple silicon Macs, and Vulkan is an option on Windows and Linux. On a Vulkan-only machine, detection, most OCR models, LaMa and AOT run on the CPU. A CPU-only setup works but is slower. You do not need to install the full CUDA or ROCm SDK. Model files can be large, and a model that fits on disk can still be too big for your graphics memory.
What Koharu does well
- It costs nothing and you can read the code. MIT or Apache 2.0, with a public issue tracker and a Discord server.
- Pages can stay on your computer. With local models, prompts and text stay on your machine after the download.
- You choose every model. Detection, OCR, inpainting and translation each have their own setting. OCR options include PaddleOCR-VL 1.6 (the default), Manga OCR and others. Inpainting options include LaMa and FLUX.2 Klein.
- Layered PSD export. The PSD holds the original image, a hidden text-removal mask, cleanup layers and editable text layers. That is what a letterer wants for final work in Photoshop.
- Real typesetting controls. Vertical CJK text uses OpenType features, and right-to-left text is supported. Fonts fall back in an order you set.
- Archives and PDFs in one project. Pages keep their order, and PDF pages render at 300 DPI.
What Koharu does not do
- No sound effect step. The docs describe text, bubbles and cleanup, not drawn sound effects. You redraw those by hand. Our post on why reader tools skip sound effects explains why this is the hard part.
- No per-series memory. Translation instructions "apply to every project" and "there is no per-project field." Names and terms for one series must be retyped when you switch to another.
- No team features in the docs. There are no roles, comments or review steps. It is a tool for one person on one computer.
- Export limits. PSD export does not write per-layer visibility and opacity, pages over 30,000 pixels on a side cannot be saved as PSD, and export overwrites files that already exist in the folder. Long webtoon strips can pass 30,000 pixels.
- Setup is on you. Drivers, model downloads, memory and provider keys are your job. The docs have a troubleshooting page for startup, model and provider errors.
Koharu compared with a hosted tool
Bublate is the hosted tool we make, so read this table with that in mind. Each Koharu cell links its source.
| Koharu | Bublate | |
|---|---|---|
| Where it runs | Desktop app on your computer (install) | Web app in the browser |
| Price | Free, open source (README) | $3 for a webtoon chapter up to 20 segments, up to $0.60 a manga page, per target language, 30 free credits |
| Hardware | A recent GPU is best, CPU mode is slower (hardware) | Any computer with a current browser |
| Models | You choose local or hosted models (providers) | Chosen and tested by Bublate |
| Sound effects | Not covered in the docs | Redrawn in the art, or kept with a small caption |
| Names across chapters | One global instruction field (translation) | A series bible per series and language |
| Teams | Not covered in the docs | Roles, comments and mentions |
| Export | PNG, CBZ or layered PSD (export) | ZIP with platform presets, or layered PSD per page |
Pick Koharu if you want free, local and fully in your control. Pick a hosted tool if setup time costs you more than $3 a webtoon chapter (up to 20 segments), if your pages have many sound effects, or if a team works on one series. See how to typeset manga and how to clean manga for the manual steps either way, and manga typesetting for how Bublate sets lettering. Prices are on the pricing page. You can start with 30 free credits on pages you own.
Who it fits
Koharu fits a translator or letterer who has a recent graphics card, likes to control each model, and wants pages to stay on their own computer when they use local models. It also fits anyone who wants a free tool and editable PSD files. Pick a hosted tool instead if you have no suitable GPU, work in a team, need sound effects redrawn in the art, or want one series memory that carries from chapter to chapter.
Frequently asked questions
How do I use the Koharu manga translator?
Install Koharu, create a project, then use File, Import Pages, Files to add pages. Pick a translation model and target language in the processing selector, choose Current page with all four stages, and click Run. Check the text in the inspector, fix the fit, then use File, Export PNG or Export PSD.
Is Koharu free?
Yes. Koharu is open source under the MIT or Apache 2.0 licence, and the release builds cost nothing. You can run local models at no charge. If you connect a hosted service such as OpenAI, Gemini, Claude, DeepL or OpenRouter, that service bills you for its use.
Does Koharu need a GPU?
Not strictly. Koharu has a CPU mode, but its README says CPU inference is substantially slower. For speed it uses CUDA on NVIDIA Turing or newer cards, ROCm on supported AMD cards, Metal on Apple silicon Macs, or Vulkan. The editor canvas needs a working WebGPU driver in all cases.
Is Koharu safe and private?
With local models, Koharu keeps prompts and text on your machine after the model download. With a hosted model, that provider receives the request content under its own retention terms. Koharu sends error reports to Sentry, and its privacy notice says these reports may contain file paths and should not be assumed anonymous.
Does Koharu translate sound effects?
The Koharu docs do not mention sound effects. They describe detection of text regions and speech bubbles, OCR, translation, inpainting and typesetting. Plan to handle drawn sound effects by hand, with the Remove, Brush and Text tools, if your pages have many of them.
Does Koharu work on a phone or in the browser?
No. Koharu is a desktop app for Windows, macOS on Apple silicon and Linux. There is no web version and no phone app. The docs site at koharu.rs is documentation only. You download a release from GitHub, or install it with WinGet on Windows or Homebrew on macOS.
Sources
- Koharu README on GitHub (features, hardware, models, licence) · checked
- Koharu releases on GitHub (version 0.87.0) · checked
- Koharu docs: Install Koharu · checked
- Koharu docs: Translate your first page · checked
- Koharu docs: Hardware and runtimes · checked
- Koharu docs: Projects and imports · checked
- Koharu docs: Formats and data · checked
- Koharu docs: Export PNG and PSD · checked
- Koharu docs: Typeset a translation · checked
- Koharu docs: Configure translation · checked
- Koharu docs: Connect a translation provider · checked
- Koharu docs: Set up Koharu Agent · checked
- Koharu docs: Privacy · checked