✦
Summary
Lift4D is a test-time optimization framework for reconstructing complete dynamic 4D objects from single monocular in-the-wild videos. It aims to recover geometry, appearance, and deformation, including regions that the camera never directly observes.
Key Features
Reconstructs complete 4D dynamic objects from monocular in-the-wild video.
Uses causal latent conditioning for temporally consistent per-frame 3D predictions.
Initializes a deformable 3D Gaussian Splatting representation.
Uses canonical Gaussians animated by sparse deformation nodes.
Refines appearance with occlusion-aware losses and inpainted frames.
Uses a novel-view diffusion prior for unobserved and occluded regions.
Targets severe occlusions and non-rigid motion.
Provides arXiv, public code link, and direct video demos.
About Lift4D
The method adapts a single-view 3D reconstruction model to generate temporally consistent per-frame 3D predictions through causal latent conditioning. These predictions initialize a deformable 3D Gaussian Splatting representation, which is then refined with occlusion-aware appearance optimization and a view-conditioned diffusion prior.
Lift4D is useful for 4D reconstruction research, dynamic object capture, and monocular video-to-asset workflows. It improves over prior baselines on challenging sequences with occlusion and non-rigid motion by combining observed details with learned completion priors.