4DAnyone generates 16 consistent views from one video for 4D human reconstruction
AI video model releases
4DAnyone generates 16 consistent views from one video for 4D human reconstruction
4DAnyone converts a casual monocular video of a person into 16 mutually consistent target viewpoints, then reconstructs a time-varying 4D Gaussian Splatting scene that can be viewed from arbitrary angles.
On DNA-Rendering, 4DAnyone reached 24.15 PSNR versus 20.55 for the nearest compared method; on DyMVHumans it reached 23.28 PSNR versus 19.86.
The framework fine-tunes Wan2.2-TI2V-5B to generate novel human viewpoints from source-video appearance and viewpoint-rendered 3D skeletons estimated with GVHMR.
Reference Context Packing compresses redundant generated views into fixed slots, reducing reference-context complexity from O(N) to O(1) as the number of views grows.
Target Context Routing cyclically reshuffles view groups during high-noise denoising steps to share global structure, then keeps adjacent-view groups fixed during low-noise steps for detail refinement; code and weights are public on GitHub and Hugging Face.