Text to Game Characterization: A Starting Point for Generative Adversarial Video Composition

摘要

Recently, image generation using GAN is emerging as a new solution for unsupervised learning. However, a solution of generating data that have complex correlations such as video remain unclear. This paper proposes a generative adversarial network (GAN) for video generation. The proposed model can capture spatio-temporal dependency of a given data distribution by untangling a video using the temporal attention of natural languages and their associated spatial properties. The authors have experimented with GAN learning from the perspective of a GAN model to analyze the associated issues when compared to the baseline, and have attempted to find breakthroughs for providing a bridge for future research.