What is Koharu? How AI manga translation works

Koharu combines manga translation steps in one application, from reading text to translating it and placing it back in the image. Users can review and edit the work along the way. Translation quality and whether data leaves the computer depend on the model and service selected.

A manga book with empty speech bubbles and overlapping sketch sheets behind it

Koharu translates and typesets manga

Koharu is an open-source manga translation application. It locates text, reads characters from images using OCR, translates them, removes the original lettering, and places the translation on the page. The developers state that the core is written in Rust and the desktop application uses Tauri.

Users can correct OCR errors before translating and adjust text placement before exporting. Automated results still need review. This article is based on developer documentation; we have not installed the application or tested its translation quality.

Misread text can lead to a wrong translation

The application locates text, reads the characters, translates them, then removes the original lettering and places the translation. If OCR misreads a name, the model may produce a fluent translation from that incorrect input. Compare the recognized text with the original image as well.

Inpainting fills the areas where lettering has been removed and may alter line art or backgrounds. Keep the original to compare those areas. When placing the translation in speech bubbles, check font size, line breaks, reading order, and dialogue that continues across bubbles.

Choose a method based on where your data goes

Koharu offers local models, online services, and servers you connect yourself. Before translating, check where the selected method sends text or images for processing.

Translation methodProcessing locationDoes data leave your computer?
Use a local model, such as a GGUF file run through llama.cppYour computerAfter the model is downloaded, translation runs on your computer.
Use an online service, such as OpenAI, Gemini, or DeepLThe provider's systemsText is sent to the service. Images may also be sent when Vision is enabled with a supported model.
Connect to a server, such as LM Studio or an OpenAI-compatible serverThe machine hosting the serverIt depends on the configured server address. If the server runs on another machine, data is sent there.

Even when OCR runs locally, translation may send data to an online service. Check what the application sends and how the provider stores it. When Vision is enabled, the application may include the manga page image for a model that supports images. See the connection and privacy guide

How to check Thai translation quality

Check OCR, translation, and text placement separately. A multilingual label does not tell you how well a model translates into Thai. Try a page with short dialogue, proper names, sentences that depend on a character’s gender or status, and dialogue that continues across speech bubbles.

First compare the OCR text with the original. Then check who is speaking, the reading order, and whether the translation fits the context. Finally, inspect Thai tone marks and line breaks in the exported image. This is an assessment method proposed by KrobClick, not a Koharu test result.

Start with one page and keep the original

Choose manga you created, have permission to use, or that has clear terms allowing translation. Try one page at a time and store the original separately from the translated image. Images available online may still have restrictions on translation and redistribution. Koharu’s software license does not grant rights to the manga being translated.

Before translating a whole volume, check your computer specifications, the model’s size and license, the translation service selected, and export formats.

Check the corrections needed before translating a whole volume

After trying one page, check the time taken and the corrections needed: OCR errors, inconsistent names, or image damage from removing lettering. Also check how easily you can continue editing. If the same problems keep recurring, the automated process may not suit that task.

Choose a page whose content you understand well enough to check the translation. Use the errors you find to select a model or adjust the process. Check the connected service before concluding that every step runs locally; defaults or promotional descriptions alone cannot establish this.

Sources

This article draws on documentation from koharu-rs/koharu at commit dee04d7ddaf5ba4178e2bb9e45b540ae6c8db04a, reviewed on October 30, 2026, Thailand time. We have not installed the application or tested a live service connection. Capabilities and terms may vary by version.