Occlusion-aware rendering
Resolve occlusion ambiguity and visibility before view synthesis.
UNIWORLD-VIEW
UniWorld-View: Large-Baseline View Synthesis via Video Diffusion Models
Peking University · Rabbitpre AI
Benchmark achievement#1WorldScore LeaderboardRanked first, as of July 2026
A unified visual world model
Turn a monocular input into an immersive, controllable world.
UniWorld-View generates photorealistic novel views from a single image or video, even when the target camera moves far beyond the original observation.
It couples occlusion-aware 3D point-cloud guidance with video diffusion to preserve scene appearance, geometry, and temporal consistency across wide-baseline camera trajectories.
Precise camera control3D & 4D view synthesisGeometry-aware video generation
Method
UniWorld-View first estimates a dynamic point cloud from the source video, then renders it from the target camera trajectory. A triple-reprojection process disambiguates occlusions, and normal-based filtering removes invalid back-facing points. The resulting geometry-aware conditions guide video diffusion for consistent novel-view synthesis.
Video diffusion
Use source appearance and rendered geometry as dual conditions.
3D & 4D reconstruction
Generate consistent multi-view video for immersive scenes.
Qualitative Results
UniWorld-View produces high-fidelity novel views across three settings: 3D NVS from a single image, 4D NVS from a source video, and 4D scene generation from a source video.
3D Novel View Synthesis
Input: a single image → Output: a generated video of novel views
Case 1
Case 2
Case 3
Case 4
Case 5
Case 6
Case 7
Case 8
4D Novel View Synthesis
Input: a source video → Output: a generated video from a new camera trajectory
Case 1
Case 2
Case 3
Case 4
Case 5
Case 6
Case 7
Case 8
Case 9
Case 10
Case 11
Case 12
Case 13
Case 14
Case 15
4D Scene Generation
Input: a source video → Output: a reconstructed 4D scene
Case 1
Case 2
Case 3
BibTeX
@misc{zhou2026uniworldviewlargebaselineviewsynthesis,
title={UniWorld-View: Large-Baseline View Synthesis via Video Diffusion Models},
author={Haiyang Zhou and Wangbo Yu and Chaoran Feng and Xunyu Zhou and Yonghong Tian and Li Yuan},
year={2026},
eprint={2608.04701},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2608.04701},
}