Code inconsistency from the Paper

[harveyslash]

I was unable to find any code relating to the sigmoid blending.
What are you using instead ?

[harveyslash]

also,
in the distance calculation in patch_match , there seems to be an element wise multiplication instead of euclidian distance calculation.

[rozentill]

@harveyslash
Hi, thanks for your interest on our project.

1. The step function with a threshold 0.05 is the fast approximation to the sigmoid function mentioned in our paper. It is just used for acceleration purpose.

2. The negative cosine similarity of normalized vectors is equivalent to the euclidean distance.

[harveyslash]

Thank you for your reply

I am maintaining the Python version of your paper , https://github.com/harveyslash/Deep-Image-Analogy-PyTorch
I even wrote a pycuda implementation of patchmatch for acceleration.
However, the quality is substantially poorer than yours.

[rozentill]

Thanks for your effort on reproduction to python. I will take a look to see if I can help you.

[harveyslash]

If you would like, i can set up my server for you so that you can run everything in jupyter without having to set up anything. Let me know if thats suitable for you.