That's pretty much what I had in mind: diff at language level not at text level.

Actually, ideally you would have content-aware DIFFs. So source code has one visualiser, images another, each binary format another.

That would make the experience much better.