UniWorld-View logoUNIWORLD-VIEW

UniWorld-View: Large-Baseline View Synthesis via Video Diffusion Models

Haiyang Zhou, Wangbo Yu, Chaoran Feng, Xunyu Zhou, Yonghong Tian, and Li Yuan

Peking University  ·  Rabbitpre AI

Overview

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

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

Single-stage generation13 cases

A single generation pass from the input image.

Case 13 · Warehouse robot
Warehouse robot input image
Input image
Generated video · Front-side sweep + rising loop
Case 14 · Knight street
Knight street input image
Input image
Generated video · Curved street glide + crane
Case 1
3D NVS case 1 input
Input image
Generated video
Case 2
3D NVS case 2 input
Input image
Generated video
Case 3
3D NVS case 3 input
Input image
Generated video
Case 4
3D NVS case 4 input
Input image
Generated video · Fixed-position pan, 45°
Case 5
3D NVS case 5 input
Input image
Generated video · 90° sweep + crane up
Case 6
3D NVS case 6 input
Input image
Generated video · S-curve, 110° turn
Case 7
3D NVS case 7 input
Input image
Generated video
Case 8
3D NVS case 8 input
Input image
Generated video · 150° sweep + pullback
Case 9 · Pool atrium
Single-image pool atrium input
Input image
Generated video · Curved dolly
Case 12 · Painted café terrace
Painted café terrace input image
Input image
Generated video · Sideward dolly · Painted scene
Case 18 · Knitted harbour
Knitted harbour input image
Input image
Generated video · Lateral flight

Two-stage generation1 case

A forward-progressing trajectory in two stages. The final frame of the first stage conditions the second stage. One external input image; the video is presented in 10 seconds.

Case 10 · Medieval street
Single-image medieval street input
Input image
Generated video · Rooftop flyover · 2 stages

4D Novel View Synthesis

Input: a source video  →  Output: a generated video from a new camera trajectory

Case 1
Input video
Generated video
Case 2
Input video
Generated video
Case 3
Input video
Generated video
Case 4
Input video
Generated video
Case 5
Input video
Generated video
Case 6
Input video
Generated video
Case 7
Input video
Generated video
Case 8
Input video
Generated video
Case 9
Input video
Generated video
Case 10
Input video
Generated video
Case 11
Input video
Generated video
Case 12
Input video
Generated video
Case 13
Input video
Generated video
Case 14
Input video
Generated video
Case 15
Input video
Generated video

4D Scene Generation

Input: a source video  →  Output: a reconstructed 4D scene

Case 1
Input video
Reconstructed 4D scene
Case 2
Input video
Reconstructed 4D scene
Case 3
Input video
Reconstructed 4D scene

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}, 
}