What is the significance of the blur in the skeleton on Example 2?

[schollz]

I refer to this example:

In this example, image A is a slightly blurred version of image A'. What is the significance of this for the NN? Why not just use the non-blurred version of the skeleton face for both A and A'? Does the amount of blur make a difference?

My own examples

Test 1

In my own results I tested going from photo (A) -> watercolor of the photo (A'). Then I use another similar photo (B) to generate a watercolor for that photo (B'). This way produces pretty unsatisfactory results.

A

A'

B

B'

Test 2

Following the skeleton method, I use a blurred image of watercolor of a photo (A) -> watercolor of the photo (A'). Then I use a similar photo (B) to generate a watercolor for that photo (B') which produces much better results. I don't know why this method works better than the former, it is somewhat non-intuitive for me.

A

A'

B

B'