xjwxjw / tf_videogan

(Ongoing) A TensorFlow version of "Generating Videos with Scene Dynamics"

tf_videogan

A TensorFlow version of Generating Videos with Scene Dynamics (Keep updating...)

Since the original dataset is 9 TB, which is almost impossible for any individual to use it directly, I here extract the pre-trained model parameter from t7 format files.

For now I have implemented the generator only, which can be used to generate videos. They claimed that the generated videos are not photo-realistic, maybe we can add additional loss functions to improve the model in the future.

Demos

The following videos were generated from sampled noise:

Beach	Golf
Train

Models

You can find the 'beach' model in the repository. The remaining 2 models can be found here

And you can also extract other 2 models with the code load_t7.py. (You will need to install the package torchfile first)

Notes

You might want to see their original torch implementation.