4DAnyone

4DAnyone is a research model that creates a 4D (animated, free-viewpoint) human from one uncalibrated monocular video. Rather than requiring a calibrated multi-camera rig, it generates the multiview videos such a rig would have recorded, then lifts them into a 4D Gaussian Splatting (4DGS) representation, trained with FreeTimeGS.\n\nThe pipeline extracts a 3D skeleton from the source video and uses it as sparse geometric conditioning for generating target-view videos. Two components address consistency when many views must be generated in separate groups: Reference Context Packing compresses the growing reference context into a fixed budget, and Target Context Routing passes context across groups to limit structural drift. The authors report better novel-view video quality and downstream 4DGS reconstruction than prior state-of-the-art methods. A YouTube reviewer described it as more detailed than ReCamMaster and TrajectoryCrafter and put the model at about 12 GB.\n\nThe project is accompanied by a SIGGRAPH Asia 2026 paper (arXiv 2608.20335) and a project page with demos. The page notes that a community Hugging Face Space and a 4DGS player exist, and that the authors are working on making it faster and easier to run on consumer-grade hardware. Input footage is expected to have mild camera motion. The page does not list code, weights or license details.