What's this about?
The paper argues that image editing has inherited the training recipe of text-to-image generation, and that this is a mismatch — editing is a different problem, and treating it as generation-with-extra-steps leaves capability on the table. 46 upvotes.
What exactly is the difference between generation and editing?
Generation starts from nothing and only has to satisfy your description. Editing starts from an existing image and has to change one thing while leaving everything else identical. Preserving what you did not ask to change turns out to be the hard part, and it is not something a generation objective ever trains for.
What's the community saying?
The interest is in the diagnosis rather than the numbers — naming a specific structural mismatch between two training paradigms tends to be more useful than another incremental benchmark gain.