AI-powered manga translator, written in Rust.
Koharu introduces a new workflow for manga translation, utilizing the power of AI to automate the process. It combines the capabilities of object detection, OCR, inpainting, and LLMs to create a seamless translation experience.
Under the hood, Koharu uses candle for high-performance inference, and uses Tauri for the GUI. All components are written in Rust, ensuring safety and speed.
Note
For help and support, please join our Discord server.
- Automatic speech bubble detection and segmentation
- OCR for manga text recognition
- Inpainting to remove original text from images
- LLM-powered translation
- Vertical text layout for CJK languages
CUDA and Metal are supported for GPU acceleration, significantly improving performance on supported hardware.
Koharu is built with CUDA support, allowing it to leverage the power of NVIDIA GPUs for faster processing.
Koharu bundles CUDA toolkit 12.x and cuDNN 9.x, dylibs will be automatically extracted to the application data directory on first run.
Koharu supports NVIDIA GPUs with compute capability 7.5 or higher.
Please make sure your GPU is supported by checking the CUDA GPU Compute Capability and the cuDNN Support Matrix.
Koharu supports Metal for GPU acceleration on macOS with Apple Silicon (M1, M2, etc.). This allows Koharu to run efficiently on a wide range of Apple devices.
Koharu relies on a mixin of computer vision and natural language processing models to perform its tasks.
Koharu uses several pre-trained models for different tasks:
The models will be automatically downloaded when you run Koharu for the first time. Windows user can download the Koharu-win-CUDA-bundled.7z release package to get those models pre-bundled.
We convert the original models to safetensors format for better performance and compatibility with Rust. The converted models are hosted on Hugging Face.
Koharu supports various quantized LLMs in GGUF format via candle, and preselect model based on system locale settings. Supported models and suggested usage:
For translating to English:
- vntl-llama3-8b-v2: ~8.5 GB Q8_0 weight size and suggests >=10 GB VRAM or plenty of system RAM for CPU inference, best when accuracy matters most.
- lfm2-350m-enjp-mt: ultra-light (≈350M, Q8_0); runs comfortably on CPUs and low-memory GPUs, ideal for quick previews or low-spec machines at the cost of quality.
For translating to Chinese:
- sakura-galtransl-7b-v3.7: ~6.3 GB and fits on 8 GB VRAM, good balance of quality and speed.
- sakura-1.5b-qwen2.5-v1.0: lightweight (≈1.5B, Q5KS); fits on mid-range GPUs (4–6 GB VRAM) or CPU-only setups with moderate RAM, faster than 7B/8B while keeping Qwen-style tokenizer behavior.
LLMs will be automatically downloaded on demand when you select a model in the settings. Choose the smallest model that meets your quality needs if you are memory-bound; prefer the 7B/8B variants when you have sufficient VRAM/RAM for better translations.
You can download the latest release of Koharu from the releases page.
We provide pre-built binaries for Windows and macOS, for other platforms, you may need to build from source, see the Development section below.
To build Koharu from source, follow the steps below.
bun installbun run buildThe built binaries will be located in the target/release directory.
If you find Koharu useful, consider sponsoring the project to support its development!
Koharu application is licensed under the GNU General Public License v3.0.
The sub-crates of Koharu are licensed under the Apache License 2.0.
![@dependabot[bot]](https://avatars.githubusercontent.com/in/29110?s=64&v=4){kind=small}
