Lift4D

Lift4D is a research framework for reconstructing the geometry, appearance and deformation of dynamic objects from a single monocular, in-the-wild video. It also completes regions the camera never observed. It is described in the paper "Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild" by Litman, Ma, Shah, Ugrinovic, Kitani, De la Torre and Tulsiani (arXiv 2606.23688, 2026).

The method is test-time optimization. An existing single-view image-to-3D diffusion transformer is adapted with causal latent conditioning, so each frame's 3D latent is initialized from the previous denoised latent. This gives temporally consistent per-frame predictions, which are decoded into Gaussian splats. The splats are consolidated into a deformable 3D Gaussian Splatting representation animated by sparse deformation nodes. The representation is then refined with an occlusion-aware optimization. A view-conditioned diffusion prior fills in occluded and unobserved regions.

The authors report improvements over prior 4D reconstruction baselines on synthetic and in-the-wild footage, especially with severe occlusions and non-rigid motion. The project page offers an interactive viewer of the reconstructions. According to a YouTube creator, code is coming soon. No code, weights or license were listed on the provided pages.