Lift4D reconstructs non-rigid 4D objects from single-view video
Lift4D introduces a framework for reconstructing dynamic, non-rigid objects from monocular in-the-wild video. By adapting single-view 3D models with causal latent conditioning and utilizing deformable 3D Gaussian Splatting, it generates temporally consistent 4D representations even under severe occlusions.
Lift4D tackles one of the hardest problems in generative vision by achieving coherent 4D object reconstruction from a single moving camera without massive 4D training datasets.
- –Causal latent conditioning adapts existing single-view 3D models to produce temporally consistent per-frame predictions
- –Deformable 3D Gaussian Splatting efficiently models complex non-rigid motion and appearance over time
- –An occlusion-aware optimization paired with a view-conditioned diffusion prior intelligently hallucinates unseen regions
DISCOVERED
1h ago
2026-06-25
PUBLISHED
1d ago
2026-06-23
RELEVANCE
AUTHOR
_akhaliq