“Current video generation models demonstrate a strong implicit understanding of 3D space, as evidenced by the high accuracy of 3D reconstruction algorithms run on their generated videos.”