OrthoMotion:Disentangling Camera and Subject Motion via Geometry Semantics Orthogonal Attention

OrthoMotion introduces a novel attention operator designed to disentangle camera and subject motion in controllable video generation, addressing the inherent entanglement in 2D conditioning. By employing a geometric channel for camera motion and a semantic channel for subject motion, OrthoMotion achieves a state-of-the-art reduction in cross-talk by over 2.4 times, as measured by a new Cross-Talk Error (CTE) metric, while maintaining fidelity across various model backbones. This method represents a significant advancement for practitioners, as it guarantees motion disentanglement through architectural design rather than relying on emergent properties.