Tag
This paper introduces GEAR, a Geometry-Enabled Attention Routing framework that uses geometry as explicit token-level addresses for visual memory to enhance long-horizon camera-controlled video generation, achieving state-of-the-art results.