A Sliced Wasserstein Loss for Neural Texture Synthesis

This is the official implementation of "A Sliced Wasserstein Loss for Neural Texture Synthesis" paper (CVPR 2021).

Input	Output (SCALING_FACTOR=2)

If you use this work, please cite our paper

@InProceedings{Heitz_2021_CVPR,
author = {Heitz, Eric and Vanhoey, Kenneth and Chambon, Thomas and Belcour, Laurent},
title = {A Sliced Wasserstein Loss for Neural Texture Synthesis},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021}
}

This implementation focuses on the key part of the paper: the sliced wasserstein loss for texture synthesis.

Requirements

Librairies

The following libraries are required:

pytorch
numpy
imageio

This code has been tested with Python 3.7.5 on Ubuntu 18.04. We recommend setting up a dedicated Conda environment using Python 3.7.

Pretrained vgg-19 network

A custom vgg network is used, as explained in the supplementals. It has been modified compared to the pytorch standard model:

inputs are preprocessed (including normalization with imagenet stats).
activations are scaled.
max pooling layers are replaced with average pooling layers.
zero padding is remplaced with reflect padding.

Texture generation

To generate a texture use the following command:

python texture_optimization_slicing.py

The parameters inside the file texture_optimization_slicing.py are:

FILE_PATH: Target texture file.
SCALING_FACTOR: Scaling factor of the optimized texture in regards to the example texture (set to 1 to keep the same size, 2 to double the resolution).

Outputs files are:

optimized_texture_N_iterations.png: the intermediate result after N iterations. If there are 20 iterations, there will be 20 output images.

Timing

Timing reference for 10 iterations on 256x256 resolution: