Local manga translator with LLMs built in
Open-source Rust app uses YOLO, LaMa, and LLMs to detect, translate, and redraw manga text automatically.
Developer Mayocream has launched Koharu, a sophisticated open-source tool built in Rust that fully automates the complex process of translating manga. The application is a multi-stage pipeline: it first uses a YOLO model to detect text bubbles in an image, then employs a custom OCR (optical character recognition) model to extract the Japanese text. Following extraction, the tool leverages a LaMa model—a specialized AI for image inpainting—to cleanly remove the original text from the bubble, preparing the canvas for the new translation.
The core translation is handled by configurable large language models (LLMs), allowing users to choose models for quality or speed. Finally, a custom text rendering engine blends the newly translated text back into the image, matching the style and flow of the original comic. Notably, Koharu is distributed as a standalone application with CUDA libraries bundled, meaning users with compatible NVIDIA GPUs can run it immediately without any complex environment setup or dependency installation, lowering the barrier to high-speed, AI-assisted translation.
- Uses a multi-model AI pipeline (YOLO, custom OCR, LaMa inpainting, LLMs) for end-to-end manga page translation.
- Distributed as a zero-setup, standalone Rust application with CUDA support bundled for immediate GPU acceleration.
- Open-source project allows customization of translation LLMs and the text rendering engine for different styles.
Why It Matters
Democratizes high-quality manga translation for fans and small publishers, automating a labor-intensive process that requires multiple skilled tasks.