Lightweight speech-bubble extraction model for manga/comics; 0.93M parameters, 1.8MB FP16, fast CPU/GPU inference, UNet+MobileNetV3, outputs mask/boundary/confidence, distilled from SAM, fine-tuned on synthetic hard samples.
A lightweight manga, comics speech bubble extraction model.
Highlights:
Click here to check online demo.
git clone https://github.com/letmutex/bubble-pop
cd bubble-pop
uv pip install "torch>=2.6" torchvision numpy opencv-python onnxuv run inference.py web-demo/assets/page-1.jpg --output overlay.jpg# ONNX FP16
uv run export.py model.pt --format onnx-fp16 --output model_fp16.onnx
# ONNX FP16 for Wasm (float32 I/O)
uv run export.py model.pt --format onnx-fp16-wasm --output web-demo/assets/model.onnx
# TFLite FP16
uv pip install "onnx2tf[tensorflow]==2.6.8"
uv run export.py model.pt --format tflite-fp16 --output model_fp16.tflite| Property | Details |
|---|---|
| Architecture | UNet |
| Backbone | MobileNetV3-Small |
| Parameters | 0.93M |
| Input | Grayscale, 768 × 1024 |
| Output | 3-channel logits (mask, boundary, confidence) |
Pre-training: This model is pre-trained on ~12K Web comics/manga pages; distilled directly from SAM 3.1's outputs.
Fine-tuning: The best pre-trained checkpoint is fine-tuned using ~2.4K synthetic pages of hard-samples.
The current version of the model may not perform well in the following cases:
A lightweight manga, comics speech bubble extraction model.
Highlights:
Click here to check online demo.
git clone https://github.com/letmutex/bubble-pop
cd bubble-pop
uv pip install "torch>=2.6" torchvision numpy opencv-python onnxuv run inference.py web-demo/assets/page-1.jpg --output overlay.jpg# ONNX FP16
uv run export.py model.pt --format onnx-fp16 --output model_fp16.onnx
# ONNX FP16 for Wasm (float32 I/O)
uv run export.py model.pt --format onnx-fp16-wasm --output web-demo/assets/model.onnx
# TFLite FP16
uv pip install "onnx2tf[tensorflow]==2.6.8"
uv run export.py model.pt --format tflite-fp16 --output model_fp16.tflite| Property | Details |
|---|---|
| Architecture | UNet |
| Backbone | MobileNetV3-Small |
| Parameters | 0.93M |
| Input | Grayscale, 768 × 1024 |
| Output | 3-channel logits (mask, boundary, confidence) |
Pre-training: This model is pre-trained on ~12K Web comics/manga pages; distilled directly from SAM 3.1's outputs.
Fine-tuning: The best pre-trained checkpoint is fine-tuned using ~2.4K synthetic pages of hard-samples.
The current version of the model may not perform well in the following cases: