GraphVid: Interactive Graph-Controllable Video Generation

Researchers introduced GraphVid, an image-to-video model that uses structured interaction graphs to control multi-object behavior more precisely than text prompts or trajectory drawing.

They also released GraphVid-Bench, a relationally annotated dataset designed to train interaction-aware video models.

GraphVid reportedly outperforms prior motion-control methods such as Motion-I2V on quality and controllability, while using less data and fewer trainable parameters.

The approach offers a more scalable and less ambiguous way to direct complex video scenes, improving both generation quality and user control.