CVPR2023

Persistent Nature: A Generative Model of Unbounded 3D Worlds

Lucy Chai, Richard Tucker, Zhengqi Li, Phillip Isola, Noah Snavely

Abstract

us to generate arbitrary camera trajectories, such as the six numbered views shown along a cyclic camera path (center). The persistence inherent to our representation stands in contrast to prior auto-regressive methods [42] that do not preserve consistency under circular camera trajectories (right); while the two images shown on the right are at the start and end of a cyclic path, the terrain depicted is completely different. Our method is trained solely from unposed, single-view landscape photos.