MIT 48px OCR

A SafeTensors conversion of the 48-pixel OCR model used by zyddnys/manga-image-translator and BallonsTranslator. The model recognizes cropped comic text lines and also predicts foreground and background colors.

Model details

• Input text height: 48 pixels
• Maximum configured width: 8100 pixels
• Visual backbone: ConvNeXt feature extractor
• Sequence model: four-layer Transformer encoder and five-layer Transformer decoder
• Embedding dimension: 320
• Attention heads: 4
• Default beam size: 5
• Default maximum sequence length: 255

Files

• model.safetensors: converted model weights
• config.json: architecture and decoding configuration
• alphabet-all-v7.txt: tokenizer alphabet and special-token vocabulary

Token IDs are 0 for padding, 1 for beginning-of-sequence, and 2 for end-of-sequence. <SP> represents a space.

Intended use and limitations

The model expects already detected, cropped, and normalized comic text regions. It does not locate text on a page. Recognition quality depends strongly on crop quality, text scale, language coverage in the supplied alphabet, and image degradation. Training data details and evaluation metrics are not included with this conversion.

License

GPL-3.0, following the upstream manga-image-translator and BallonsTranslator implementations.