Video Generation Using Large Language Models: Work in Progress

A novel video generation model merges language model capabilities with diffusion-based methods, enhancing performance and enabling zero-shot video generation. It employs a multitask pretraining strategy, unifying various tasks like text-to-video and video-to-video editing. This approach allows for training with unpaired video data, reducing reliance on video-text pairs. The model showcases diverse video generation capabilities and potential for scaling to larger datasets and improved zero- and few-shot task performance.