You would store the diff between the reconstructed version and the original. The diff could be compressed to much less space than the original image, because th

You would store the diff between the reconstructed version and the original. The diff could be compressed to much less space than the original image, because the NN can predict most of the pixels. And get within a few bits of the rest.