EduStory: A Unified Framework for Pedagogically-Consistent Multi-Shot STEM Instructional Video Generation

Researchers have introduced EduStory, a novel unified framework designed to enhance the generation of long-horizon instructional videos, particularly within STEM domains. While current video generation models have improved in visual quality, they often fail to maintain knowledge consistency and coherent pedagogical narratives across multiple shots. EduStory addresses these limitations by integrating pedagogical state modeling to track persistent knowledge states, employing script-guided structured control for narrative organization, and utilizing learning-oriented evaluation metrics to assess knowledge fidelity. To facilitate rigorous testing, the team also launched EduVideoBench, a diagnostic benchmark featuring multi-granularity annotations such as pedagogical storyboards and knowledge state transitions. Extensive experiments indicate that this domain-aware approach significantly reduces narrative breakdowns and improves alignment with instructional intent. The study underscores the critical role of domain-specific structural constraints and tailored benchmarks in advancing reliable, controllable, and trustworthy AI-driven educational content generation, marking a significant step forward in computer vision and artificial intelligence applications for education.