StoryEngine is a framework for generating long-form video stories with multiple shots. It addresses the problem of visual drift and narrative incoherence by maintaining an explicit representation of the story world state, including entity placements and attributes. This state is updated as events occur, and each shot is planned to depict intended start and end states. The framework uses canonical references for recurring entities and a repair loop to correct local inconsistencies.
Experiments show StoryEngine outperforms existing methods on a new that measures storytelling quality, narrative coherence, and visual consistency. The approach prevents local generation errors from propagating across shots, leading to more coherent and visually consistent stories. This is interesting because it tackles a key limitation in current systems.