We include qualitative results demonstrating 3D-consistent video generation in the table below. The input noise is constructed by projecting Gaussian noise from the 3D meshes and adding independent Gaussian noise in the 2D image space, as detailed in Section 4.2 and Section 4.3 of the paper. For improved visualization, the input noise videos are upsampled by a factor of 2.
| Input Noise Video | Generated Video | Ground Truth Video |
|---|---|---|