Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild

Lift4D is a test-time optimization framework that reconstructs complete dynamic objects from monocular video. It adapts a single-view 3D model for temporally consistent per-frame predictions, then sculpts a deformable 3D Gaussian Splatting representation using occlusion-aware optimization and a diffusion prior. The method outperforms prior approaches on sequences with severe occlusions and non-rigid motion.