A Survey of Generative Video Models as Visual Reasoners

The rapid advancements in generative artificial intelligence have ushered in a new era for video content creation, enabling the synthesis of increasingly realistic and complex visual sequences. This survey delves beyond the conventional role of generative video models as mere content creators, exploring their burgeoning capabilities as sophisticated visual reasoners. We provide a comprehensive overview of the foundational architectures driving modern video generation, including Generative Adversarial Networks, Variational Autoencoders, and Diffusion Models, highlighting their unique challenges in ensuring temporal consistency and high fidelity. A significant portion of this work is dedicated to defining and categorizing visual reasoning tasks, from object interaction and temporal dynamics to causal inference, setting a framework for evaluating the reasoning prowess of these models. We then meticulously examine how generative video models either explicitly or implicitly demonstrate these reasoning capabilities through tasks like future frame prediction, controllable video synthesis, and multimodal question answering. Furthermore, we address the critical challenges and limitations, such as data scarcity, computational demands, the prevalence of hallucinations, and the imperative for explainability and trustworthiness, alongside ethical considerations like bias and potential misuse. Concluding with a synthesis of key findings, this survey identifies promising future research directions, emphasizing the development of robust evaluation metrics, the exploration of multimodal reasoning, and the design of architectures that inherently integrate reasoning, offering a forward-looking perspective on their transformative impact across various domains.