StoryEngine: A State-Grounded Agentic Framework for Video Storytelling

The paper proposes StoryEngine, a state-grounded agentic framework designed to solve visual drift and narrative inconsistency in multi-shot video generation. StoryEngine separates authoritative semantic plans from unreliable visual observations by maintaining a structured representation of entity states and propagating event changes. A bounded evaluation-guided repair loop corrects local inconsistencies, ensuring that causal story progression is maintained across shots without error propagation.

StoryEngine establishes a separation between authoritative semantic plans and unreliable visual observations.

A bounded evaluation-guided repair loop further corrects local state inconsistencies.

StoryEngine consistently outperforms state-of-the-art methods across all evaluation dimensions for video storytelling.