Your statement implies the source dataset's file sizes represent the amount of visual information in them. This is almost certainly not the case.