CubeComposer pushes native 4K 360 video

This is a specialized research paper, but it tackles a real technical bottleneck in immersive video generation: native resolution, not just flashy demos. The interesting part is less “another video model” and more the systems design that makes 4K 360 generation computationally practical.

• –The core trick is splitting panorama generation across cubemap faces and time windows, which cuts peak memory versus denoising a full 360 scene in one pass
• –The paper’s sparse context attention and continuity-aware blending are aimed directly at the seam and consistency problems that usually make 360 outputs look fake
• –It claims native 4K output where earlier perspective-to-360 methods generally topped out around 1K native generation and leaned on post upscaling
• –The work also introduces a 4K360Vid dataset with 11,832 high-resolution clips, which may matter almost as much as the model for future benchmarking and follow-on research