Co-Director: Hierarchical Multi-Agent Framework for Video Storytelling
A new study introduces Co-Director, a structured approach to help multiple agents work together on video storytelling by treating it as an overall optimization problem. It uses a multi-armed bandit strategy to explore creative possibilities, while also implementing a local self-refinement loop that helps maintain consistency in character identity throughout the story. This technique strikes a good balance between trying out new narrative styles and using effective methods. The researchers also present GenAD-Bench, a dataset with 400 scenarios that feature fictional products for personalized ads. Results show that Co-Director improves semantic coherence compared to current agent-based systems, which often struggle with semantic drift. You can find the paper on arXiv under the ID 2604.24842.
Key facts
- Co-Director is a hierarchical multi-agent framework for video storytelling
- It formalizes video storytelling as a global optimization problem
- Uses a multi-armed bandit for global creative direction
- Local multimodal self-refinement loop mitigates identity drift
- GenAD-Bench dataset contains 400 scenarios of fictional products
- Dataset designed for personalized advertising evaluation
- Addresses semantic drift and cascading failures in current pipelines
- Published on arXiv with ID 2604.24842
Entities
Institutions
- arXiv