Exploring maximum-mean-discrepancy-based descriptive neural image style transfer using PyTorch.
Many years have passed since this project was done in 2018. This project is now lack of maintenance. Links to the model and some of the images are broken. One in need of the model may find this issue useful.
./download_models.sh provided by [9].
./download_images.sh
python3 neural_image_style_transfer.py built upon [9].
This project studies the basic form of the neural image style transfer technique proposed by Gatys et al. ([1.1]). First, we reproduce the result of [1.1], which uses Gramian matrix to capture the concept of style. Next, we substitute mean vector, variance vector, and covariance matrix for Gramian matrix, and find that these new components capture different aspects of style and some of them even improve the quality of style transfer. Finally, we implement the idea "activation shift" in [3], which is another technique to improve the quality of style transfer.
For details, refer to the report.
[1.1] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. 2016.
[3] R. Novak and Y. Nikulin. Improving the neural algorithm of artistic style. 2016.
[9] PyTorch implementation of [1.1] by Gatys. https://github.com/leongatys/PytorchNeuralStyleTransfer