✦
Summary
4DAnyone reconstructs a moving human in 4D from a casually captured monocular video. It generates reconstruction-grade, multiview-consistent videos and lifts them into a 4D Gaussian Splatting representation without a calibrated camera array, rig, or known camera poses.
Key Features
Monocular video to 4D human reconstruction.
Multiview-consistent target-view video generation.
4D Gaussian Splatting output.
3D skeleton geometric conditioning.
Reference Context Packing.
Target Context Routing.
Unknown-camera and in-the-wild robustness.
Public paper, code, and model resources.
About 4DAnyone
The method uses a 3D skeleton for sparse but accurate geometric conditioning, Reference Context Packing to keep reference context within a fixed budget, and Target Context Routing to communicate structure across independently generated view groups. The generated views are then used to train a 4DGS model.
4DAnyone is useful for novel-view human video, digital characters, virtual production, telepresence, sports analysis, and 4D scene reconstruction from footage captured in the wild. Its focus on mild camera motion and unknown intrinsics makes it practical for casual video rather than specialized capture setups.