Seems like it's reversed along the Y axis as well? I'm curious what led to that. The nefarious side of my brain say it was a very basic attempt at making the source training data less immediately recognizable in any generated output, but I do wonder if there's a more innocent explanation.
A "more innocent explanation" could simply be data augmentation. It seems pretty clear they don't care that it's obviously using watermarked Shutterstock videos.
