World from Motion improves dynamic 3D Gaussian reconstructions by using a video generator as a controllable prior. We condition generation on a persistent 4D representation, sample new dynamic viewpoints, and distill the generated observations back into the reconstruction.
Browser-based 4D Gaussian previews with scene switching and camera controls.
4D Reconstruction Benchmark on DyCheck
| Method | Covisible mPSNR ↑ | Covisible mSSIM ↑ | Covisible mLPIPS ↓ |
|---|---|---|---|
| Shape-of-Motion | 17.32 | 0.598 | 0.296 |
| MoSca | 19.32 | 0.706 | 0.264 |
| WorldTree | 19.75 | 0.728 | 0.240 |
| ViDAR | 19.69 | 0.713 | 0.223 |
| World-from-Motion | 20.26 | 0.732 | 0.215 |
4D Novel-View Synthesis Benchmark on DyCheck
| Method | mPSNR ↑ | mSSIM ↑ | mLPIPS ↓ |
|---|---|---|---|
| ReCamMaster | 10.96 | 0.262 | 0.755 |
| GEN3C | 12.06 | 0.260 | 0.679 |
| TrajectoryCrafter | 13.06 | 0.320 | 0.656 |
| Vista4D | 14.14 | 0.310 | 0.514 |
| World-from-Motion | 18.45 | 0.635 | 0.362 |
Virtual-camera ablation with mPSNR, mSSIM, and mLPIPS from the paper table.
PCK@0.05 compares track quality on DyCheck.
Inference-time guidance on dense 4D scaffold rendering improves both fidelity and accuracy.
We thank Yang Zheng, Zhengfei Kuang, Lior Yariv, and Jianhao Zheng for fruitful discussions. We also thank Yijia Weng and Jiahui Lei for providing evaluation details for MoSca, Kuan Heng Lin for providing Vista4D evaluation details, and Michal Nazarczuk and Eduardo Pérez-Pellitero for providing evaluation details for ViDAR. This website builds on the templates from RealmDreamer and CAT4D.
@misc{zhu2026worldfrommotion,
title = {World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video},
author = {Liyuan Zhu and Shengyu Huang and Amrita Mazumdar and Tianye Li and Zan Gojcic and Gordon Wetzstein and Iro Armeni and Shalini De Mello and Alex Trevithick},
year = {2026}
}