This paper proposes a novel neural rendering approach that represents a scene as a compact set of "compressed light-field tokens (CLiFTs)", retaining rich appearance and geometric information of a scene.