Dynamic 2D Video Using Neural Networks

Dynamic 2D Video Using Neural Networks

Author Name : Yash Worlikar, Omkar Shingan, Sakshi Shinde, Arathi Kamble

ABSTRACT

During the recent years, there has been a rapid boon in the Video rendering, VFX and simulation industry. The current improvements in the hardware and software technology has enabled us to re-create real life phenomenon that were previously impossible to simulate. However, the computation required for such simulations is too large to use it for everyday applications. A single frame may take hours or even days to compute. Not to mention if a minor change is made the entire thing needs to be re-calculated which makes the process of editing much more tedious. The aim of the project is to make it possible to edit such environments without the need of such large computations through the use of Generative adversarial networks (GANs). The project aims to train a world model based on a dataset of a simple 2D sand simulation. We would we training the GAN networks with a large dataset of videos of a 2D sand simulation generated by us. The GAN would then attempt to generate similar videos in an attempt to simulate the dataset it was trained on. Later we would combine the GAN model with a simple video editing tool that would enable the user edit pre-recorded videos in multiple ways. The trained model would attempt to generate an output that would integrate these changes within the video.

Keywords— VFX, video rendering, generative adversarial networks