GroundShot: Visually Consistent Multi-Shot Long Video Generation via Entity-Grounded Shot Scheduling

GroundShot is a novel framework for generating visually consistent multi-shot videos, addressing the challenge of entity drift across shots without requiring additional training or model modifications. It employs an entity-level visual memory system to schedule shot generation based on the reliability of entities, thereby enhancing consistency in their appearances. Additionally, the introduction of GroundBench provides a diagnostic benchmark to evaluate consistency at the entity level, demonstrating that GroundShot significantly outperforms existing methods in maintaining visual coherence across multi-shot sequences.