would you ever need to version individual images? at a high level, you could version the collection as a whole by filtering by timestamp, and moving deleted files to a "deleted" directory, or maintaining filename and timestamps in a database. I'm sure there are lots of corner cases that would come up when you actually tried to build such a system, but I don't think the overall scheme needs to be as conceptually complex as source code version control
Would this be a good fit for large machine learning data sets?
Every time a new model comes out that touts having been trained on something like a quarter of a billion images, I ask myself "how the heck does someone manage and version a pile like that?"