basic
basic diffaug
deluxe
deluxe diffaug 200 and 6400
no cycle-consistency 1000
cycle-consistency 1000 and 10000
apple-orange 10000
With the cycle-consistency the structure of the input cat is maintained much better, although sometimes seems to sacrifice a coherent cat for a cat with the right lightness/darkness around the eye locations. This is because without the cycle-consistency it is only incentivised to make a convincing cat of the right type, not a convincing cat that also corresponds one-to-one (and is thus invertible) to the input cat.
DCDiscriminator 10000
The DCDiscriminator seems to be much more focused on texture than structure compared with the patch discriminator. This is probably because the patch discriminator has spatial data in it which gives it the opportunity to penalize lack of larger spatial relationships.
