Flex4DHuman: Flexible Multi-view Video Diffusion for 4D Human Reconstruction

Jen-Hao Cheng^1,2  ·  Yipeng Wang^2,†  ·  Hao Zhang²  ·  Gengshan Yang²  ·  Jenq-Neng Hwang¹

¹ University of Washington    ² World Labs

^† Project lead

---

Overview

Flex4DHuman turns monocular or sparse multi-view videos of dynamic subjects into synchronized dense multi-view videos using only relative camera-pose conditioning — no geometry priors — enabling direct reconstruction of 4D Gaussian splats.

Given one or more reference-view videos, their camera poses, and a set of target camera poses, Flex4DHuman synthesizes consistent novel-view videos across the target views through camera-pose-conditioned video diffusion. The generated synchronized multi-view videos can then be lifted into 4D Gaussian splats, enabling applications in AR/VR, gaming, simulation, and video re-shooting.

Highlights

• Flexible input — works from a monocular or multi-reference views.
• Pose-only conditioning — relies solely on relative camera poses; no SMPL/geometry priors required.
• Synchronized dense multi-view output — temporally and cross-view consistent novel views.
• Direct 4D reconstruction — generated views can be lifted into 4D Gaussian splats.

News

• 2026-06-11 — Project page released. Code release is in preparation — stay tuned.

Data

We release the Multi-view Caption dataset used in this project on Hugging Face:

See andaba/multi-view_caption for contents and licensing.

Citation

If you find our work useful, please consider citing:

@article{cheng2026flex4dhuman,
  title={Flex4DHuman: Flexible Multi-view Video Diffusion for 4D Human Reconstruction},
  author={Cheng, Jen-Hao and Wang, Yipeng and Zhang, Hao and Yang, Gengshan and Hwang, Jenq-Neng},
  journal={arXiv preprint arXiv:2606.13655},
  year={2026}
}

Acknowledgements

I thank Claude Code (Opus) for tireless coding and ideation throughout this project.