(DIGAN) Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks

Published : 2022/21 Paper : Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks

---

previous works : represents videos as 3D grids of RGB values → neglects continuous dynamics

• INR : encodes a continuous signal into a parameterized neural network

1. INR-based generator
   • modify first layer to handle time coordinate
   • content Generator is same with INR-GAN
   • motion Generator is added
2. Discriminator that identifies unnatural motion without observing entire long frame sequences
   • 2D convolutional network