Light-X

Generative 4D video rendering with joint camera and illumination control (ICLR 2026): re-renders monocular video along novel camera trajectories while relighting the scene, unifying trajectory control and illumination editing in one video-generation model — a step toward controllable neural rendering of dynamic real-world scenes.

Joint HUST / NTU (S-Lab, Ziwei Liu) / Tsinghua-AIR / BAAI / CUHK-Shenzhen work with NTU as co-lead — representative of S-Lab's generation-and-restoration line alongside the ByteDance-branded SeedVR/SeedVR2 (filed with ByteDance Seed per attribution rules).