What they did

The authors designed Co-Director as a multi-agent system with two levels: a global orchestrator using a multi-armed bandit to explore narrative strategies, and local agent modules for video generation with self-refinement to maintain character identity and sequence coherence. They created GenAD-Bench with 400 fictional product advertising scenarios for evaluation. They compared against state-of-the-art video generation pipelines.

Key findings

• Co-Director significantly outperformed baseline methods on GenAD-Bench across semantic coherence and visual consistency metrics.
• The hierarchical parameterization reduced semantic drift and cascading failures common in chained agent pipelines.
• The multi-armed bandit effectively balanced exploration of novel story directions with exploitation of successful configurations.

Why it matters

This work provides a principled optimization framework for agentic video storytelling, addressing the key challenge of maintaining coherence across generated clips. The GenAD-Bench dataset enables standardized evaluation for this emerging task.

Caveats

The evaluation is limited to advertising scenarios; generalization to other narrative forms is suggested but not extensively validated. The framework relies on underlying diffusion models, so improvements in those models could yield different results. The multi-agent architecture may have increased computational cost.