The paper states that it shows swapping of texture between apples by using a manual segmentation input. A manual mask is also used to transfer the fireball texture to the perfume bottle.

Presumably it is a demonstration of the network extracting the texture of an image, ignoring "spatiality", such that when a style/texture transfer is applied on one and the same image, the texture may very well order up differently than on the original image.

More specifically, ignoring spatiality is the effect of matching the response of the covariance matrix of the feature vectors, which was the key insight in the original style transfer paper by Gatys.

All 3 looked the same to me, but I guessed it was because I'm colour blind. It's not just that then?

I think the input and the style are the same photo, which is odd. But if you look at the output; you can see that even though the colors have switched, the patterns and shapes have stayed the same. The middle apple has dots on it, but it is now green. The right apple is kinda spotty, but it is now red.

If you look at the URLs, most of them have the input in the 'examples/input/' dir, and the reference in the 'examples/style/' dir. The apple reference is in the 'examples/input/' dir. Perhaps that really does mean that the input and style were identical, and they deliberately re-used the same image URL, but to me it feels like a copy-and-paste error :)