Generative Cinematographer: Composing Camera and Object Motion in 3D

Generative Cinematographer (GenCine) enables joint control of camera and object motion for video generation through an editable 3D scene scaffold reconstructed from a single image. Users specify camera trajectories and manipulate foreground regions with local 3D motion handles. The system projects these controls into guidance maps that encode region locations, handle identities, and 3D positions in a shared world coordinate system. A lightweight guidance branch and LoRA adapters condition a pretrained Wan video model on these maps. Experiments demonstrate controllable motion and improved geometric consistency as viewpoints change.

Under review at ICLR 2027