Can anyone tell me exactly how the 'image painter' network works?

[threeMetreJim]

Been playing with a version of the 'image painter' for some time.  I have discovered a way to get it to store multiple images but was wondering exactly how that is possible.  You can tell it which picture you'd want to paint with suitable modification.  Not sure if anyone else has tried this.  Also could it be used as a 'time averaged' classifier? as it reads the pixels randomly, this would make it position and rotation invariant if it did indeed turn out to work.  Random pixels from image 1---> output class 1 (on average), random pixels from image 2 ---> output class 2 (on average).

Could be interesting to try. There's no real reason to expect it to work, but they tend to throw up surprising results.