manga-ocr-base

kha-white/manga-ocr-base is an image-to-text model that performs optical character recognition (OCR) for printed Japanese text, with a primary focus on Japanese manga.

Key Strengths

• Robust recognition of both vertical and horizontal text layouts.
• Handles text with furigana (phonetic annotations) accurately.
• Works reliably on text overlaid on images, such as speech bubbles.
• Supports a wide variety of fonts, styles, and low-quality image inputs.

Best for

Developers requiring high-quality Japanese text extraction from manga pages, scanned comics, or any printed Japanese material where conventional OCR may struggle with layout complexity or image degradation.

Architecture

The model uses a Vision Encoder Decoder framework, built on Hugging Face Transformers. Underlying code and training details are available in the official repository.

Hosted by Gigarouter

This model is offered as a managed, OpenAI-compatible API. No installation or local setup is required; simply call the endpoint with an image and receive text output.