“Gini3 achieves world consistency and spatial memory by generating frame-by-frame, deliberately avoiding explicit 3D representations like NeRFs or Gaussian splatting.”