Comic Text Detector

A split SafeTensors conversion of the comic text detector from dmMaze/BallonsTranslator. It combines a YOLOv5 feature extractor, a UNet text mask head, and a DBNet-style line detection head.

Model details

• Task: comic text-region and text-line detection
• YOLO component: FP16 weights
• UNet and DBNet components: FP32 weights
• Outputs: text mask, line probability maps, and structured text blocks after postprocessing
• Format: SafeTensors state dictionaries with source tensor names preserved

Files

• yolo-v5.safetensors: YOLOv5 backbone, neck, and detection-head state
• unet.safetensors: text segmentation head
• dbnet.safetensors: differentiable-binarization line head
• config.json: component and runtime metadata

The three files exactly preserve the key sets, shapes, dtypes, and tensor values from their corresponding sections in the original comictextdetector.pt checkpoint.

Intended use and limitations

Use this model to identify text regions and individual lines before comic OCR. It is specialized for illustrated pages and may perform poorly on photographs, handwriting, unusual layouts, or very small text. Detection thresholds and postprocessing are part of the complete BallonsTranslator/Koharu pipeline; loading only the tensor files is not sufficient to reproduce final text blocks. No evaluation metrics or training dataset artifacts are included here.

License

GPL-3.0, following the BallonsTranslator implementation.