Large Video Generative Models: A Systematic Survey

Video generation has wide applications in cinematic production, autonomous systems, interactive entertainment, and domain-specific content creation. In the last couple of years, we have witnessed remarkable progress in video generation technologies, largely driven by large video generative models (LVGMs), which enable the generation of long and high-quality videos. In comparison to traditional video generative models, LVGMs are trained on much more data, have substantially more parameters, and have much stronger zero-shot capabilities on a wide range of video generation tasks. The exciting advances and new challenges in this area call for a systematic analysis on the core technologies and insights. Towards this end, we provide a systematic survey on LVGMs. We first provide a formal definition of LVGMs, and then cover the core technical components including model architectures, data construction, training and evaluation. Finally, we discuss the applications and future directions of LVGMs. This survey serves as a guide for understanding the mechanisms of LVGMs and building them.