Manga OCR

A SafeTensors-packaged Japanese manga OCR model based on kha-white/manga-ocr-base. It uses the Transformers Vision Encoder Decoder architecture with a ViT image encoder and a character-level Japanese BERT decoder.

Capabilities

The model is designed for printed Japanese text, particularly manga-specific layouts:

  • vertical and horizontal text
  • furigana
  • text overlaid on illustrations
  • varied fonts and styles
  • degraded or low-resolution text crops

Loading

from transformers import AutoImageProcessor, AutoTokenizer, VisionEncoderDecoderModel

repo = "mayocream/manga-ocr"
model = VisionEncoderDecoderModel.from_pretrained(repo)
processor = AutoImageProcessor.from_pretrained(repo)
tokenizer = AutoTokenizer.from_pretrained(repo)

Input images should be cropped to the text region before recognition. See the original manga_ocr project for its complete normalization and decoding pipeline.

Model details

  • Task: Japanese image-to-text OCR
  • Encoder: ViT/DeiT-style image encoder
  • Decoder: character-level Japanese BERT with cross-attention
  • Training dataset metadata: Manga109-s
  • Format: Transformers-compatible SafeTensors

Limitations

This model does not detect text regions. Results depend on crop quality and can be unreliable for handwriting, unsupported characters, non-Japanese text, extreme perspective, or heavily obscured glyphs. OCR output should be reviewed before use in archival, legal, or accessibility-critical workflows.

License

Apache-2.0, matching the upstream model repository.