Unlike previous generation models that treated image and text as separate domains ... and unique identifying features across multiple edits. Internal testing shows a 94% consistency rate compared to ...