CMU Lift4D: Single-view 4D reconstruction in-the-wild

rsasaki0109 · x · 2026-08-22

Carnegie Mellon University released Lift4D, a test-time optimization framework for reconstructing complete dynamic objects from monocular in-the-wild videos. Lift4D adapts a single-view 3D reconstruction model for temporally consistent predictions to initialize a deformable 3D Gaussian Splatting representation. It then refines this representation via occlusion-aware optimization to match the video, recovering visible details and completing unobserved regions using a view-conditioned diffusion prior.

Original post →

More from Multimodal

Multimodal channel →