Regarding the first trick: the value really isn't very much. Storing 2x2 downsampled versions of an image at all levels is only a 33% increase in size. This is the mipmap trick, and can be done right now with nothing but a size check on the client. I don't think 33% is worth mucking with file formats over.