Current video diffusion models generate high-fidelity clips but fail at coherent storytelling. Agentic pipelines that chain modules suffer from semantic drift and cascading failures. Co-Director addresses this with a hierarchical multi-agent framework that treats storytelling as a global optimization problem.
The framework uses a multi-armed bandit to globally explore and exploit creative directions, while a local multimodal loop ensures identity and sequence consistency. It is evaluated on GenAd-Bench, a new 400-scenario dataset of fictional products for personalized advertising. Co-Director outperforms state-of-the-art baselines on both automated metrics and human evaluation.
This is interesting because it provides a principled way to balance exploration of novel narrative strategies with exploitation of effective configurations, reducing the brittle handcrafted prompting of prior work. The approach is claimed to generalize to broader cinematic narratives beyond advertising.