See the original code and links to paper at https://github.com/nkolkin13/STROTSS
See the forked version we've based ourselves on at https://github.com/futscdav/strotss
We've created altered versions of the stross.py file with different distances for the content loss.
-
In strossEuclidianDistance.py we've deleted the original cosine distance and used only the euclidian distance already defined, thus it doesn't the scale invariance needed for self similarity and the results are unsatisfactory.
-
In stross.py we've used the cosine distance in the content loss, which is the original implementation. This version clearly mantains the scale invariance.
-
In strossPearsonDistance.py we've used the pearson distance in the content loss. This version mantains the scale invariance and wields very similar results to the original implementation.
- In strossAngularPearsonDistance.py we've used the analogous angular formulation of the pearson distance in the content loss. This version still mantains the scale invariance and yields unconsistent results depending on the images used.
Usage:
python strotss{specify which}.py <content> <style> [--weight 1.0] [--output strotss.png] [--device "cuda:0"]
or if you want to use the scripts for batch processing:
python script{specify which}.py
the naming covention for the input images is:
content{3 digit number}.jpg
style{3 digit number}.png
