A Sliced Wasserstein Loss for Neural Texture Synthesis - PyTorch version

This is an unofficial, refactored PyTorch implementation of "A Sliced Wasserstein Loss for Neural Texture Synthesis" paper [CVPR 2021].

Notice:

• The customized VGG-19 architecture might be different from the original Tensorflow implementation. Thus, some results might be inconsistent to the paper. Feel free to give advice.
• The spatial tag part is not included in this implementation.

Prerequisites

• Python 3.7.10
• PyTorch 1.9.0

Data

I have collected data in the data folder from the official repository and from "Deep Correlations for Texture Synthesis " [Siggraph 2017].

Run

First cd pytorch and then run some random examples:

python texturegen.py --data_folder=SlicedW --img_name=input.jpg
python texturegen.py --data_folder=SlicedW --img_name=2.png
python texturegen.py --data_folder=SlicedW --img_name=berry.png
python texturegen.py --data_folder=SlicedW --img_name=64.png
python texturegen.py --data_folder=DCor --img_name=Texture13.png
python texturegen.py --data_folder=DCor --img_name=Texture32.jpg
python texturegen.py --data_folder=DCor --img_name=Texture19.png
python texturegen.py --data_folder=DCor --img_name=Texture22.png

After slightly more than 1 minute for each scene, you can find intermediate outputs in outputs folder, and final results in results folder.

Sample Results

Input	Synthesized

References

• Official Tensorflow version