PanoWorld: Real-World Panoramic Generation

PanoWorld targets long-range memory in panoramic world models by exploiting the rotation-equivariant nature of omnidirectional representations. It converts camera trajectories into translations under fixed headings and introduces Dense Panoramic Ray-Conditioning (DPRC) and Geometry-aware Memory Augmentation (GMA). The authors also present World360, combining real-world video clips captured by panoramic UAVs with simulated clips generated by AirSim360. The dataset is designed to test physical consistency across large spatial changes and varied illumination, settings that the authors argue are less represented in existing benchmarks. The abstract reports substantial gains over alternative methods, while models, code, and data are planned for public release.