4DAnyone, released by the Ant Research organization on GitHub, tackles a bottleneck in turning ordinary phone footage of a person into a 4D asset: multi-view video generation. Rather than reconstructing directly from a single monocular clip, the model synthesizes dozens of synchronized, view-consistent videos of the same subject and motion from new camera angles, producing the kind of multi-camera capture that downstream Gaussian Splatting pipelines expect but that casual recording never provides. The paper, accepted to SIGGRAPH Asia 2026, uses a two-stage pipeline: a bundled GVHMR submodule recovers pose and skeleton from the source video, and a video diffusion model then denoises new target-view videos conditioned on that skeleton and the original footage. Two variants ship: 4DAnyone-Base runs the standard schedule, while the distilled, four-step 4DAnyone-Turbo (default) reaches near-Base quality per the project's benchmark table, denoising roughly 5.1x-5.9x faster across GPUs from RTX 4090s to H200s. Camera layouts are configurable, from a compact 6-view orbit up to a 48-view, three-pitch-layer rig or a 180-degree frontal arc, set via views-per-layer, pitch, start-yaw, and yaw-span arguments; every view is a 121-frame clip at 704x1280. The project has iterated quickly since its early-August 2026 release: peak GPU memory dropped from just under 32GB to just under 24GB across two rounds, and the 24-view pipeline gained a 1.42x speedup, making single-GPU inference practical on an RTX 4090, where one 121-frame view takes about 27 seconds. A bundled exporter converts one multi-view frame into a Nerfstudio dataset, with a visual-hull point cloud and per-view masks, trainable into a 3D Gaussian Splatting scene via Splatfacto. Full open-source 4D Gaussian Splatting reconstruction is on the roadmap as not yet done, so today's release covers multi-view video synthesis plus a static, per-frame 3DGS export path. Inputs are narrow by design: a single person, full-body or upper-body, minimal camera movement, 121+ frames, 1080p or higher, 9:16 portrait, reflecting a model built for human-subject clips rather than arbitrary scenes. Released under the Apache 2.0 license.