Previously, I took one photo and rearranged its squares to make a version of another photo. How about a sillier version, where you use one image to re-make itself?
The simple answer is to just keep every square in its original place, but that’s no fun at all. Instead, we can add the requirement that no tile can stay put. The simplest way to implement this is to increase the cost of such laziness to ridiculous levels: say a million times greater than the highest cost normally achievable.
At the top of this page are two examples on either side of the original Mona Lisa. The left version just translates tiles, and the right version allows for rotations. These use the optimal assignments.
The original locations of the tiles are easier to understand with the non-rotated tiles, but the image is better with rotations. In the fleshy areas, it shifts around the recognizable bits: eyes in the hands, a chin for decolletage, lips for the wrong side of lips. Gross! Next one!
Oof, also creepy! The more-visible eyes pop out grotesquely. How about a safe one:
Whew. No surprise terrors.