Co-Director: Agentic Generative Video Storytelling

Google researchers have introduced Co-Director, a hierarchical multi-agent framework that addresses a critical flaw in generative video storytelling: maintaining narrative coherence across scenes. While diffusion models can produce stunning individual clips, chaining them into a story often leads to semantic drift—characters change appearance, plots lose direction, or inconsistencies pile up. Co-Director formalizes video storytelling as a global optimization problem, balancing the exploration of creative narrative strategies with the exploitation of effective configurations.

At the core of Co-Director is a multi-armed bandit algorithm that globally identifies promising creative directions, while a local multimodal self-refinement loop keeps character identity and sequence-level consistency intact. The team also built GenAD-Bench, a 400-scenario dataset of fictional products for personalized advertising, to evaluate the system. Results show Co-Director significantly outperforms state-of-the-art baselines, offering a principled approach that generalizes to broader cinematic narratives. This work, led by Yale Song and 15 other authors, represents a step toward turning AI video generation into a reliable storytelling engine.