MODNet: Trimap-Free Portrait Matting in Real Time
April 11, 2026 · View on GitHub
Ported weights and clean inference codebase for MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition (AAAI 2022). Provides both PyTorch and ONNX Runtime inference for images, videos, and webcam feeds.
Tip
The models and functionality in this repository are integrated into UniFace — an all-in-one face analysis toolkit.
| Input | Matte | Green Screen | RGBA |
|---|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Installation
pip install -r requirements.txt
Usage
Image Inference
python main.py --source image.jpg --model weights/modnet_photographic.pt --output results/
Image with Custom Background
python main.py --source image.jpg --model weights/modnet_photographic.pt --background bg.jpg --output results/
Video Inference
python main.py --source video.mp4 --model weights/modnet_photographic.pt --background bg.jpg --output out.mp4
Webcam Inference
python main.py --source 0 --model weights/modnet_webcam.pt
ONNX Inference
No PyTorch dependency -- only onnxruntime and opencv-python.
python onnx_inference.py --source image.jpg --model weights/modnet_photographic.onnx --output results/
python onnx_inference.py --source video.mp4 --model weights/modnet_photographic.onnx --background bg.jpg --output out.mp4
python onnx_inference.py --source 0 --model weights/modnet_webcam.onnx
ONNX Export
python -m utils.export_onnx --weights weights/modnet_photographic.pt --output weights/modnet_photographic.onnx
python -m utils.export_onnx --weights weights/modnet_webcam.pt --output weights/modnet_webcam.onnx --simplify
Model Weights
Weights ported from the official repository. Model weights are licensed under Apache 2.0 by the original authors.
| Model | Size | Download |
|---|---|---|
| MODNet Photographic | 25 MB | PyTorch | ONNX |
| MODNet Webcam | 25 MB | PyTorch | ONNX |
Reference
Based on MODNet by Zhanghan Ke