What it is
Reconstructs 4D humans from a casual monocular video by generating multiview-consistent videos and lifting them into 4D Gaussian Splatting. Instead of denoising many target views in isolated batches, it keeps all prior views as fixed-cost packed context and rotates target-view groups during diffusion to share structure across groups. It uses sparse 3D skeleton conditioning rather than dense depth or known cameras.
Key results
- Beats prior methods on DNA-Rendering novel-view video quality
- Beats prior methods on DyMVHumans downstream 4DGS reconstruction
- Generalizes robustly to in-the-wild monocular videos
Shows that reconstruction-scale multiview diffusion needs both scalable reference conditioning and cross-group target communication, not just camera control.