3D-Aware Video Generation

Results on FaceForensics Dataset
Ours with Forward-Facing Camera
Ours with Forward-Facing Camera and Zoom Effect
Ours with Rotating Camera
Ours with Motion and Content Decomposition
(random motions applied to each identity)
Ours pre-trained on FFHQ with Forward-Facing Camera
Ours pre-trained on FFHQ with Rotating Camera along Two Axes
Ours simultaneously trained on FFHQ and FaceForensics with Rotating Camera along Two Axes
Ours pre-trained on FFHQ with Motion and Content Decomposition
(two same motions applied to 4 identities)
Ours with Content Interpolation
Ours with Motion Interpolation
Ours with Random Motions
Samples Generated with MoCoGAN-HD
Samples Generated with DIGAN
Samples Generated with StyleGAN-V
Results on MEAD Dataset
Ours with Forward-Facing Camera
Ours with Forward-Facing Camera and Zoom Effect
Ours with Different Camera Positions
Samples Generated with StyleGAN-V
Results on TaiChi Dataset
Ours with Rotating Camera and Static Motion
Ours with Rotating Camera and Dynamic Motion
Ours with Forward-Facing Camera
Ours with Forward-Facing Camera and Zoom Effect
Samples Generated with DIGAN
Results on SkyTimelapse Dataset
Ours with Forward-Facing Camera
Ours with Rotating Camera along First Axis and Static Motion
Ours with Rotating Camera along First Axis and Dynamic Motion
Ours with Rotating Camera along Second Axis and Static Motion
Ours with Rotating Camera along Second Axis and Dynamic Motion
Samples Generated with DIGAN
Samples Generated with StyleGAN-V
Notes

For FaceForensics and SkyTimelapse comparisons to MoCoGAN-HD, DIGAN, and StyleGAN-V we use generated results provided on the StyleGAN-V website.