---
license: gpl-3.0
library_name: candle
pipeline_tag: image-segmentation
base_model: kitsumed/yolov8m_seg-speech-bubble
tags:
  - candle
  - yolo
  - image-segmentation
  - comic
  - manga
  - speech-bubble
  - safetensors
---

# Speech Bubble Segmentation

A Candle-compatible SafeTensors conversion of [`kitsumed/yolov8m_seg-speech-bubble`](https://huggingface.co/kitsumed/yolov8m_seg-speech-bubble) for locating and segmenting speech bubbles in comic and manga pages.

## Model details

- Architecture: YOLOv8m-seg
- Input size: 640 × 640
- Class: `speech bubble`
- Mask prototypes: 192
- Mask coefficients: 32
- Default confidence threshold: 0.25
- Default NMS threshold: 0.45
- Default mask threshold: 0.5

## Files and provenance

- `model.safetensors`: converted floating-point weights with the original Ultralytics tensor names
- `config.json`: loader, class, and postprocessing metadata
- `config.yaml`: original upstream dataset configuration
- Upstream revision: `da4efccf35a15c8a8c2564431a4b7e121d3e0d99`

## Intended use and limitations

Use this model to produce speech-bubble masks for comic layout analysis, text placement, or rendering. It has one class and should not be treated as a general panel, character, or text detector. Validate thresholds and mask quality on the target comic style; training and evaluation artifacts are not included in this conversion.

## License

GPL-3.0, matching this repository and the upstream model card.