Nvidia said that previous models that transform 2D images into 3D models, such as ShapeNet, need to be trained using datasets of 3D shapes in order to achieve the same results.