A GOOD IMAGE GENERATOR IS WHAT YOU NEED FOR HIGH-RESOLUTION VIDEO SYNTHESIS

Supplementary Videos

The webpage includes three parts. The first part shows videos generated by our method. The second part shows videos generated by baseline methods. The third part shows videos used in Amazon Mechanical Turk experiments.


Note: We display videos in a lower resolution to fit into screen.


Each video also contains a button to view in full-screen.


To view videos in full-resolution, please click the .mp4 files.

Video samples of our method


Figure 7: UCF-101 video generation (256x256). Show video


Figure 8: FaceForensics video generation (256x256). Show video


Figure 9 & 10: the 16-frame trained model can synthesis 32- and 64-frame videos. Show 32-frame 64-frame


Figure 11 & 12: each row indicates using the same content code to generate diverse motion, or applying same motion to different content codes. Show same content same motion


Figure 13: Sky Time-lapse video generation (128x128). Show video


Figure 14: (FFHQ, VoxCeleb) cross-domain video generation, BigGAN generator (128x128). Show video


Figure 15: (FFHQ, VoxCeleb) cross-domain video generation (256x256). Show video


Figure 16: (FFHQ, VoxCeleb) cross-domain video generation (1024x1024). Show video


Figure 17 & 18: (AFHQ-Dog, VoxCeleb) cross-domain video generation (512x512). We also interpolate 16-frame video to get 32 frames. Show 16-frame 32-frame


Figure 19: (AnimeFaces, VoxCeleb) cross-domain video generation (512x512). Show video


Figure 20: (LSUN-Church, TLVDB) cross-domain video generation (256x256). Show video

Video samples of other baselines


TGANv2 on FaceForensics (256x256). Show TGAN res


DTVNet on Sky Time-lapse (128x128). Show DTVNet res

Video samples used in Amazon Mechanical Turk experiments


(FFHQ, VoxCeleb) cross-domain video generation, w/o contr loss (256x256). Show video


(FFHQ, VoxCeleb) cross-domain video generation, w/ (left) and w/o (right) mutual info loss. Show w/ mutual loss w/o mutual loss