Harmonizer: Learning to Perform White-Box Image and Video Harmonization (ECCV 2022)

Paper | Demo | Code | License | Citation | Contact

Harmonizer Result Video | Enhancer Result Video

In our demos, the Harmonizer model is trained on the iHarmony4 dataset, while the Enhancer model is trained on the FiveK + HDRPlus datasets.

Try our online demos for fun without code!

We provide offline demos for image/video harmonization/enhancement.

The training code is released in the folder ./src/train.
Refer to README.md for more details about training.

We provide PyTorch validation code to reproduce the iHarmony4 results reported in our paper, please:

1. Download the Harmonizer model pre-trained on the iHarmony4 dataset from this link and put it in the folder ./pretrained.

2. Download the four subsets of iHarmony4 from this repository and put them in the folder ./dataset/harmonization/iHarmony4.

3. Install python requirements. In the root path of this repository, run:

   pip install -r src/requirements.txt
   

4. For validation, in the root path of this repository, run:

   python -m src.val_harmonizer \
          --pretrained ./pretrained/harmonizer \
          --datasets HCOCO HFlickr HAdobe5k Hday2night \
          --metric-size 256
   

   • You can change --datasets to validate a specific subset.
   • You can remove --metric-size to calculate the metrics without resizing the outputs.
   • The metric values may slightly different from our paper due to the dependency versions.

This project is released under the Creative Commons Attribution NonCommercial ShareAlike 4.0 license.

If this work helps your research, please consider to cite:

@InProceedings{Harmonizer,
  author = {Zhanghan Ke and Chunyi Sun and Lei Zhu and Ke Xu and Rynson W.H. Lau},
  title = {Harmonizer: Learning to Perform White-Box Image and Video Harmonization},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2022},
}

This repository is maintained by Zhanghan Ke (@ZHKKKe).
For questions, please contact kezhanghan@outlook.com.