I haven't seen even visual diffing, let alone merge between formats
(binary or textual, doesn't matter) - I'm not talking about images, but
(usually stored in proprietary and/or ad hoc formats because of efficiency
requirements) animations, geometry etc
What you are proposing is a nice idea, but it would took an enormous
amount of work to be generally applicable (merge of two skinned characters
with different number of bones, anyone?) and still will be error-prone.
Moreover, merging between two different data sets is not solely a
technical problem, it requires artistic eye because even correctly merged
result of two great-looking changes may still look like shit.
It's so much easier to just lock the files, really!