Large Scale Video Generation

This repo is an implentation of adversarial video generation based on the model proposed in DVD-GAN. The implementation extends the BigGAN framework to a video setting by adding a video discriminator and a convolutional GRU to generate frame sequences.

Code is based on the BigGAN Github repo. The video generation code in this repo is mainly developed by Xiaodan Du and Jugat Lamba.

Using the Code

We trained a model on Kinetics-400 to generate 12 frame 64x64 resolution videos. In similar fashion to the BigGAN repo, the scripts/launch_DVDGAN_64.sh script launches model training for a Kinetics-400 video GAN at 64x64 resolution with the hyperparameters we used.

We trained our model for around 2-3 days on 8 V100s. Example generated videos are shown below:

Acknowledgments

This code is built on the PyTorch BigGAN codebase.