GraphVid: Interactive Graph-Controllable Video Generation

TL;DR

GraphVid introduces a graph-based interface for controlling video generation, allowing users to specify complex multi-object interactions through structured interaction graphs rather than manual trajectory drawing. It achieves high-quality video synthesis with improved metrics over prior motion-control methods, despite using less training data and fewer parameters.

Why It Matters

This approach simplifies the control of complex video scenes by replacing error-prone manual tracking with intuitive, semantically meaningful interaction graphs, making controllable video generation more accessible and scalable for real-world applications.

Key Findings

• GraphVid uses interaction graphs to enable precise, flexible control over multiple objects in video generation, reducing reliance on pixel-level motion inputs.
• The model outperforms Motion-I2V in key quality metrics, achieving lower FID and FVD scores and higher PSNR and SSIM values.
• A new benchmark dataset, GraphVid-Bench, provides structured relational annotations to support training of interaction-aware video models.