Koharu is an AI-powered manga translator written in Rust that automates the translation workflow through machine learning.
The tool addresses the challenge of translating manga at scale by combining object detection, optical character recognition, inpainting, and large language models into a single pipeline. It runs all vision models and language models locally on the user's machine, keeping data private and secure. The workflow is selective, allowing users to apply detection, OCR, translation, and inpainting operations at either individual page or project scope. The tool handles multiple input formats including raster images, archives, and PDFs, and provides a WebGPU-based canvas for manual refinement alongside automated processing.
Koharu suits translators and fan communities working with manga who want to retain control over their data and avoid cloud dependencies. The project is particularly well-suited for batch processing of large manga collections, with support for hardware acceleration through CUDA, ROCm, Metal on Apple silicon, and Vulkan. The tool includes specialized features for manga translation such as speech bubble detection, multilingual text layout with vertical CJK and right-to-left support, and layered PSD export for both flattened delivery and further editing. Users can choose between local GGUF inference for complete privacy or hosted LLM providers for convenience.
The project maintains active community engagement through a Discord server for support and discussion. Development shows consistent attention to expanding hardware acceleration support across multiple platforms and GPU architectures. The codebase demonstrates investment in sophisticated text rendering and layout capabilities tailored to the specific demands of manga translation across different writing systems.
