#2020th of 24 in Desktop AI apps

koharu

Desktop manga translator with local detection, OCR, inpainting and LLM translation

Stars
6.6k
License
Apache-2.0
Last commit
Oct 2026
Last release
Sep 2026
Language
Rust
Stars, last 29 weeks+4,151
10 Apr 2026: 2,409 stars17 Apr 2026: 3,611 stars24 Apr 2026: 3,997 stars1 May 2026: 4,169 stars8 May 2026: 4,252 stars15 May 2026: 4,331 stars22 May 2026: 4,393 stars29 May 2026: 4,438 stars5 Jun 2026: 4,502 stars12 Jun 2026: 4,583 stars19 Jun 2026: 4,640 stars26 Jun 2026: 4,720 stars3 Jul 2026: 4,763 stars10 Jul 2026: 4,814 stars17 Jul 2026: 4,844 stars24 Jul 2026: 4,891 stars31 Jul 2026: 4,947 stars7 Aug 2026: 4,981 stars14 Aug 2026: 5,161 stars21 Aug 2026: 5,310 stars28 Aug 2026: 5,402 stars4 Sept 2026: 5,464 stars11 Sept 2026: 5,522 stars18 Sept 2026: 5,572 stars25 Sept 2026: 5,647 stars2 Oct 2026: 5,699 stars9 Oct 2026: 5,930 stars16 Oct 2026: 6,361 stars23 Oct 2026: 6,479 stars30 Oct 2026: 6,560 stars

2,409 on 10 Apr 20266,560 on 30 Oct 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

Koharu is a Rust desktop app that translates manga pages in a pipeline of text and bubble detection, OCR, inpainting to remove source text, and LLM or machine translation. Vision models and GGUF LLMs run locally, with optional hosted providers such as OpenAI, Gemini, Claude, DeepL and OpenAI-compatible endpoints. It includes a WebGPU canvas editor for manual cleanup and typesetting, and exports layered PSD.

Who it is for: Manga scanlators and readers who want private, local page translation

Strengths

  • Full pipeline in one app: detection, OCR, inpainting, translation, typesetting and layered PSD export
  • Runs vision models and GGUF LLMs locally; hosted LLM and MT providers are optional
  • GPU backends: CUDA, ROCm/HIP, Metal and Vulkan; no full CUDA SDK install needed
  • Handles images, archives and PDFs, with vertical CJK and right-to-left text layout

Weaknesses

  • CPU inference is documented as substantially slower than GPU
  • Editor canvas needs WebGPU and a current graphics driver even with CPU inference
  • Narrow use case: manga translation only, not a general chat or assistant app
  • README gives no RAM or VRAM figures; larger local models need extra hardware

What it needs

  • GPU optional
  • Needs WebGPU-capable graphics driver, Hugging Face model downloads, Rust, Bun and LLVM (build from source only)
  • Models: GGUF, Gemma 4, Qwen 3.5/3.6/3.8, Ministral 3, LFM 2.5

Quick answers

Is koharu open source?

Yes. Its code is on GitHub under the Apache-2.0 license.

Does koharu need a GPU?

No. It runs on a CPU and can use a GPU when you have one.

Also in Desktop AI apps

See all 24
Also in Desktop AI apps
RankProjectScore
1cherry-studioDesktop client for chatting with many LLM providers, with agents and knowledge bases53.2k starsAGPL-3.092 out of 100
2dsh-desktopNative Windows and macOS desktop client for DeepSeek Harness, built on plugins34.4k starsMIT88 out of 100
3screenpipeRecords screen and audio locally, makes it searchable for you and your agents22.2k starscustom license88 out of 100
#4prompt-optimizerPrompt rewriting and testing tool for web, desktop, Chrome and MCP Live demo ↗ (opens in a new tab)40.2k starscustom license85 out of 100
#5deep-live-camReal-time webcam face swap and video deepfake from a single source image97.5k starsAGPL-3.084 out of 100