In the first round, we fed image 1 the style of image 2.
Now we feed this image the style from image 1. You might logically assume it would just turn back to the original image (the original source image fed its own style should just return it), but thats obviously not how a machine “thinks.” (Machine learning introduces enough loss that its impossible to return to the original.)
Instead you get this.
comparing the original source and this first loop, we can see that some elements do return. The flowers re-appear, some of the structure returns to us.
Now we’re just gonna run this a bunch of times more, bouncing back and forth between the two styles.
Suffice to say, you get the most of these feedback loops in the very early stages.