Split the image into NxN tiles (e.g. 4x4 or 8x8) and encode each tile either as a direct listing of pixel values, or as a reference to another part of the image. Using 4x4 tiles to compress a 512x512 version the classic Lenna image, I was able to generate something visually quite similar to the original by supplying only 610 tiles of source pixel data, out of 16384. This means that only 3.7% of the tiles in the compressed image come from the source image, the rest are all backreferences to x/y regions of the compressed image.