WorldReel: 4D Video Generation with Consistent Geometry and Motion Modeling (CVPR 2026)

Project Page | Paper

Authors: Shaoheng Fang¹, Hanwen Jiang², Yunpeng Bai¹, Niloy J. Mitra^2,3, Qixing Huang¹
Affiliations: ¹The University of Texas at Austin; ²Adobe Research; ³University College London

TL;DR WorldReel jointly generates RGB videos and 4D scene representations, keeping geometry and motion consistent across time, viewpoints, and dynamic scenes.

Overview

WorldReel is a 4D video generator that jointly predicts RGB frames and structured 4D scene representations, including pointmaps, camera trajectory, and dense flow. By modeling appearance, geometry, and motion in one unified framework, WorldReel produces videos that stay consistent across time, viewpoint changes, and dynamic scene content. The model is trained with both synthetic 4D supervision and real videos, enabling strong geometric fidelity together with realistic in-the-wild generation.

Status

• Training code
• Inference code
• Pretrained checkpoints

Code and checkpoints are coming soon.

Citation

If you find this work useful, please consider citing:

@article{fang2025worldreel,
  title={WorldReel: 4D Video Generation with Consistent Geometry and Motion Modeling},
  author={Fang, Shaoheng and Jiang, Hanwen and Bai, Yunpeng and Mitra, Niloy J. and Huang, Qixing},
  journal={arXiv preprint arXiv:2512.07821},
  year={2025}
}