How to extend short video clips into minute-long cinematic sequences?

Generating long, coherent videos remains unsolved because current models either preserve a single shot or invent freely without grounding in the original footage. ReCA treats long-video generation as a hierarchical decomposition problem: it recursively breaks the task into smaller, context-bounded chunks, uses frozen short-video generators at each step, and propagates visual state (identity, scene, objects) across time to prevent drift. On a new benchmark for 3–5 minute generation, ReCA outperforms competing methods by 8–16% on overall quality and maintains much better shot consistency.