Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild
Summary
Lift4D presents a test-time optimization framework to reconstruct complete 4D geometry, appearance, and deformation from monocular video. It uses causal latent propagation to initialize per-frame 3D latent, a Gaussian Splat representation, and an occlusion-aware, diffusion-prior-based refinement to hallucinate unobserved regions, outperforming prior methods on challenging in-the-wild sequences.