You seem to be assuming that the pixel data only needs to be "visually" correct, but that isn't always the case. It's entirely possible that the data will be subject to additional processing which can magnify any errors. If you're going to add lossy compression (even if it's only "slightly" lossy), it needs to be optional.