Smooth and long video generation

Abstract

Diffusion models have gained significant traction in recent years for image generation tasks due to their ability to produce high-quality outputs while avoiding common pitfalls of adversarial approaches like Generative Adversarial Networks. Their probabilistic framework, which progressively denoises random noise, ensures stability during training and allows precise control over the generative process. More recently, researchers have extended diffusion models to video generation, with promising results from methods such as AnimateDiff and LaVie. However, generating long videos remains an open challenge, as issues such as temporal coherence deterioration and high computational costs limit scalability. This final year project explores training-free approaches to address these challenges, leveraging pre-trained video generation models to improve the length and temporal consistency of generated videos. Building on the capabilities of AnimateDiff, the project aims for computational efficiency, avoiding resource-intensive training while delivering advancements in video length and temporal coherence. These contributions pave the way for practical applications of video diffusion models in fields such as film, virtual reality, education, gaming, and the advertising industry.