4DAnyone reconstructs 4D humans from one uncalibrated video

4DAnyone claims to reconstruct a moving 4D human from a single uncalibrated video by first hallucinating the missing camera angles, then lifting the generated multiview footage into 4D Gaussian Splatting. The arxiv preprint posted August 20, 2026 by Yudong Jin and collaborators argues that today's camera-controlled video diffusion models cannot scale to the viewpoint counts a 4DGS reconstruction actually needs.

Their diagnosis has a name. "Existing camera-controlled video diffusion models synthesize plausible novel-view videos but fail to maintain consistency when scaled to the tens of target views required for 4DGS reconstruction," the paper says, calling this a "bounded-attention-context problem": once the target views exceed one DiT forward pass, they must be split into groups, and consistency drifts.

The fix has two parts. Reference Context Packing "compresses growing reference views into a fixed-length mixed-resolution context with O(1) reference-context complexity." Target Context Routing "rotates target-view groupings during denoising to share context across groups at high-noise steps and stabilize details at low-noise steps."

Training leans on a proprietary source. The authors built the MVGameHuman dataset "using our in-house game engine" and combined it with light-stage and in-the-wild video. Evaluation is reported on DNA-Rendering and DyMVHumans, where the paper says 4DAnyone "outperforms prior methods in both novel-view video quality and downstream 4DGS reconstruction, with robust in-the-wild generalization."

The abstract publishes no per-benchmark numbers, no runtime, and no code link.