A computer-vision preprint reports a benchmark lead for 4DAnyone, a system designed to turn an uncalibrated video into views from other angles and a downstream 4D reconstruction. On the DNA-Rendering and DyMVHumans benchmarks, the authors say it outperformed all listed baselines in generated-video consistency, 4DGS reconstruction and generated-video reconstruction.
The study asks whether a 4D human can be reconstructed from an uncalibrated monocular video. The task is to generate consistent novel views from a single source and then use those views to build a 4DGS model.
From one source to a set of views
4DAnyone starts with the source video and uses an HMR-derived, depth-buffered 3D skeleton as conditioning. It then generates multiview-consistent human videos and uses those generated views to reconstruct a 4D Gaussian Splatting, or 4DGS, model.
One of the proposed components, Reference Context Packing, compresses the growing set of generated-view references into a fixed-length context. The paper describes this as keeping the cost of reference conditioning constant as the reference material grows.
The second component, Target Context Routing, changes how target views are grouped during denoising. In the high-noise phase, it cyclically shifts and repartitions target views into groups of four; in the low-noise phase, it fixes adjacent four-view groups.
In the proposed pipeline, the skeleton provides structural guidance while the two context mechanisms organize the visual information used to generate target views. The authors present the resulting videos as material suitable for downstream 4D reconstruction.
A broad training mix, followed by a limited test
The training mix included MVGameHuman, with 38,000 videos, 24 cameras and 318 actors; SynCamVideo, with 34,000 videos, 10 cameras and 66 actors; DNA-Rendering, with 51,000 videos, 48 cameras and 548 actors; TedTalk, with 42,000 videos, one camera and 413 actors; and Pexels, with 20,000 videos, one camera and 1,411 actors.
These figures describe video and scene collections rather than an enrolled human-participant sample. The benchmark results are therefore evidence about a computer-vision system evaluated on datasets, not direct evidence about people, health, behavior or safety.
The implementation was built on Wan2.2-TI2V-5B and trained at 704×1280 with a learning rate of 1 × 10−5 on 128 H20-3E GPUs. At inference, it used 20 denoising steps and a TCR switching ratio of 0.2.
For quantitative evaluation, the researchers used 10 DNA-Rendering test scenes and three DyMVHumans test scenes. Each scene contained 16 approximately uniformly distributed cameras and 98 frames. Starting from a single front-view source video, 4DAnyone generated 16 uniformly spaced target-view videos.
The comparison included MV-Performer, TrajectoryCrafter and a fine-tuned ReCamMaster. The baseline conditions were not identical: ReCamMaster was fine-tuned, while MV-Performer and TrajectoryCrafter were evaluated with released weights in zero-shot use.
The reported scores favored the new system
On DNA-Rendering, the system's reported scores for generated-video consistency were 24.33 PSNR, 0.862 SSIM and 0.163 LPIPS. For 4DGS reconstruction, the figures were 24.15, 0.863 and 0.159. For generated-video reconstruction, they were 23.69, 0.850 and 0.165.
On DyMVHumans, the reported generated-video consistency scores were 24.48 PSNR, 0.862 SSIM and 0.109 LPIPS. The 4DGS reconstruction results were 23.28, 0.846 and 0.117, while generated-video reconstruction produced 21.03, 0.808 and 0.143.
Across both tables, the system's reported row had higher PSNR and SSIM and lower LPIPS than the listed baselines. That metric pattern is the basis for the authors' claim that 4DAnyone led the comparisons across all three evaluation dimensions.
The three measures describe different parts of the reported comparison: whether the generated views remain consistent, how well the generated views support a 4DGS reconstruction, and how closely the generated videos match the evaluation target. The paper reports these as descriptive metric comparisons rather than as results accompanied by inferential tests.
The ablation table compared different configurations
An ablation study removed or changed parts of the system and measured generated-video consistency. Without both TCR and RCP, the reported PSNR/SSIM/LPIPS triplet was 21.09/0.766/0.216. Removing RCP produced 22.03/0.780/0.203, while removing TCR produced 22.21/0.788/0.196. The Full (Random) configuration produced 22.20/0.788/0.197.
Among the listed configurations, Full Sliding had the highest reported PSNR and SSIM and the lowest reported LPIPS. The table therefore places the strongest metric combination with the full system and its sliding routing arrangement, while the versions missing RCP or TCR scored lower.
The difficult cases are still visible
Beyond the controlled benchmark tables, the authors report qualitative generalization across diverse human-centric videos and challenging inputs. Their figures show consistent target-view videos and high-quality 4DGS renderings in those examples.
The reported comparisons include no confidence intervals, significance tests or other uncertainty estimates. The tables show the direction and size of the metric differences, but they do not provide a statistical estimate of how uncertain those gaps are.
The paper also reports clear failure cases. Flowing fabric was inconsistent across views, and in one example every generated view inherited an estimated flat-foot pose from an inaccurate HMR result.
Those examples expose two boundaries of the pipeline: clothing that moves independently of the body skeleton may not stay consistent across views, and errors in the estimated pose can be carried into the generated views. The reported system therefore does not establish reliable reconstruction for every clothing type, motion pattern or monocular video.
The quantitative evidence is also drawn from a limited set of reported benchmark scenes, while the ablation uses selected challenging DNA-Rendering sequences. Results from those settings cannot by themselves answer how the method will perform across a larger and more varied collection of real-world videos.
The analysis leaves open whether the reported ranking would remain the same if every baseline received the same training or fine-tuning treatment. It also identifies broader tests of loose garments, strong non-body motion, HMR errors, motion blur, occlusion and camera movement as unresolved questions.
A technical result, still awaiting broader testing
The work is a preprint. Its front matter identifies it as arXiv:2608.20335v1, dated 20 August 2026, and directs readers to a project page for video results and source code.
The paper says the work was partially supported by the National Key R&D Program of China, NSFC, the Zhejiang Provincial Natural Science Foundation of China, Ant Group, and the Information Technology Center and State Key Lab of CAD&CG at Zhejiang University.
Taken on its own terms, the study presents a reported computational advantage for turning one uncalibrated source video into multiview material for 4DGS reconstruction. Its main caveat is equally specific: the approach remains vulnerable when clothing motion departs from the body skeleton or when the initial pose estimate is wrong.
Paper data and sources
Original title: 4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Authors: Yudong Jin, Tao Xie, Qihang Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text