4Director: Controlling Video World Models with Rigid 3D Geometry
4Director steers generated video with explicit 3D object meshes and per-frame rigid motion, instead of flatter image-plane controls.
The paper says each object is reconstructed once from the input image as a canonical mesh, then moved through prescribed rigid transformations frame by frame. That geometry is rendered as a depth video, which a Motion Adapter turns into video while keeping appearance, lighting, and non-rigid details consistent. The authors also introduce RealCOD-Rigid, a 20,774-clip training dataset, and IG-IoU, a metric for motion adherence and identity preservation. They report that 4Director beats prior methods on visual quality and camera/object control. HF Daily Papers' note
The paper says each object is reconstructed once from the input image as a canonical mesh, then moved through prescribed rigid transformations frame by frame. That geometry is rendered as a depth video, which a Motion Adapter turns into video while keeping appearance, lighting, and non-rigid details consistent. The authors also introduce RealCOD-Rigid, a 20,774-clip training dataset, and IG-IoU, a metric for motion adherence and identity preservation. They report that 4Director beats prior methods on visual quality and camera/object control. HF Daily Papers' note
score 5