Code and data for paper "Deep Painterly Harmonization"

This software is published for academic and non-commercial use only.

This code is based on torch. It has been tested on Ubuntu 16.04 LTS.

Dependencies:

• Torch (with loadcaffe)
• Matlab or Octave

CUDA backend:

• CUDA
• cudnn

Download VGG-19:

sh models/download_models.sh

Compile cuda_utils.cu (Adjust PREFIX and NVCC_PREFIX in makefile for your machine):

make clean && make

To generate all results (in data/) using the provided scripts, simply run

python gen_all.py

in Python and then

run('filt_cnn_artifact.m')

in Matlab or Octave. The final output will be in results/.

Note that in the paper we trained a CNN on a dataset of 80,000 paintings collected from wikiart.org, which estimates the stylization level of a given painting and adjust weights accordingly. We will release the pre-trained model in the next update. Users will need to set those weights manually if running on their new paintings for now.

Update: Removed a few examples due to image copyright issue. You can still find them in data/ folder for testing use only.

Here are some results from our algorithm (from left to right are original painting, naive composite and our output):

• Our torch implementation is based on Justin Johnson's code;
• Histogram loss is inspired by Risser et al.

If you find this work useful for your research, please cite:

@article{luan2018deep,
  title={Deep Painterly Harmonization},
  author={Luan, Fujun and Paris, Sylvain and Shechtman, Eli and Bala, Kavita},
  journal={arXiv preprint arXiv:1804.03189},
  year={2018}
}

Feel free to contact me if there is any question (Fujun Luan [email protected]).