Harmonizer - Video Harmonization Demo

This is an offline demo of Harmonizer Video Harmonization.

1. Requirements

The basic requirements for this demo are:

• Ubuntu System
• Python 3+

2. Run Demo

We recommend creating a new conda virtual environment to run this demo, as follow:

1. Clone this repository:

   git clone https://github.com/ZHKKKe/Harmonizer.git
   cd Harmonizer
   

2. Download the pre-trained model harmonizer.pth from this link and put it into the folder Harmonizer/pretrained/.

3. Create a conda virtual environment named harmonizer (if it doesn't exist) and activate it. Here we use python=3.8 as an example:

   conda create -n harmonizer python=3.8
   source activate harmonizer
   

4. Install PyTorch and the required python dependencies (please make sure your CUDA version is supported by the PyTorch version installed). In the root path of this repository, run:

   pip install -r src/requirements.txt
   

5. Execute the demo code in the root path of this repository, as:

   python -m demo.video_harmonization.run \
          --example-path ./demo/video_harmonization/example
   

   where ./demo/video_harmonization/example contains an example (the foreground portrait mask is generated by MODNet).
   If you want to test your own samples, please refer to the folder ./demo/video_harmonization/example to prepare:

   • the foreground video and the corresponding foreground mask video.
   • the background video

6. Check the composite video in the folder: ./demo/video_harmonization/example/composite
   Check the video harmonization results in the folder: ./demo/video_harmonization/example/harmonized