AutoStruct-Video: LLM-Guided Spatio-Temporal Reasoning for Multi-Event Video Generation

Generating videos that depict multiple distinct events unfolding coherently over time — for example, a sequence of causally or narratively connected actions — remains challenging for current video generation models, which often struggle to maintain long-range narrative and spatio-temporal consistency across multiple depicted events. This project proposes AutoStruct-Video, a framework that uses a large language model to plan the spatio-temporal structure of a multi-event video (event sequencing, spatial layout, temporal transitions) before guiding a video generation model to synthesize content that adheres to this explicit structural plan, improving coherence across complex, multi-event video generation.