But that’s not how models handle images either. They spatially segment and reason about title bars, placement, etc.