StoryEngine: A State-Grounded Agentic Framework for Video Storytelling
StoryEngine tries to keep generated video stories coherent by tracking the intended world state between shots.
The paper says current agentic video pipelines can lose causality and visual consistency when they rely on shot text or prior pixels alone. StoryEngine instead maintains structured state for entities, placements, and story-relevant changes, then turns those constraints into render plans. It also uses canonical references for recurring characters and settings, plus a bounded repair loop for local state errors. The authors report better results than state-of-the-art methods on their long-form storytelling benchmark. HF Daily Papers' note
The paper says current agentic video pipelines can lose causality and visual consistency when they rely on shot text or prior pixels alone. StoryEngine instead maintains structured state for entities, placements, and story-relevant changes, then turns those constraints into render plans. It also uses canonical references for recurring characters and settings, plus a bounded repair loop for local state errors. The authors report better results than state-of-the-art methods on their long-form storytelling benchmark. HF Daily Papers' note
score 4