“I will give a picture, you will create an exact copy, accurate pixel to pixel to the one I give you”.

He got disqualified on purpose in order to make a point

you don't even need to train anything, just run it through on a super low blur and almost nothing will change.

To train a diffusion model that only outputs one image with difference is I think not possible you could do an image to image and then fix the seed so you would get a consistent result and then picking the nearest result that is nearly an identical copy

That probably doesn't count as "AI" do... It's more a very bad form of compression (that may very well make the image file larger)