koharu-rs/koharu

What it solves

Koharu is a local-first manga translation tool that automates the tedious process of translating comics. It eliminates the manual effort of detecting text, removing source text from artwork, and typesetting new translations into speech bubbles.

How it works

Koharu integrates several machine learning models to create a translation pipeline:

• Detection & Segmentation: Identifies text regions and speech bubbles.
• OCR: Extracts text from the identified regions using multimodal OCR models.
• Translation: Translates the extracted text using either local LLMs (via GGUF inference) or hosted cloud providers.
• Inpainting: Uses generative AI to remove the original text and reconstruct the background artwork.
• Typesetting: Automatically fits translated text into the layout, supporting vertical CJK and right-to-left text.

Who it’s for

It is designed for manga translators and scanlation groups who want a private, secure, and automated workflow for translating raster images, archives, and PDFs.

Highlights

• Local-First: Runs vision and LLM models locally to ensure data privacy.
• Hardware Acceleration: Supports CUDA, ROCm/HIP, Metal, and Vulkan for high-performance inference.
• Agent-Based Workflow: Includes an agent for project inspection and pipeline control.
• Professional Export: Supports layered PSD exports for further editing in professional software.
• Flexible Model Support: Compatible with a wide range of local GGUF models (e.g., Gemma 4, Qwen) and cloud APIs.