What if it just gets scaled down with nearest neighbour? Then the pixels would be forced on to a grid of a particular size. Maybe some kind of algorithm that compares the % difference to the full scale image to see what scale makes the most sense for it?