OmniRoam: World Wandering via Long-Horizon Panoramic Video Generation

Yuheng Liu; Xin Lin; Xinke Li; Baihan Yang; Chen Wang; Kalyan Sunkavalli; Yannick Hold-Geoffroy; Hao Tan

cs.CVMar 31, 2026

OmniRoam: World Wandering via Long-Horizon Panoramic Video Generation

Authors: Yuheng Liu, Xin Lin, Xinke Li, Baihan Yang, Chen Wang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Hao Tan,

Organizations: University of California, Irvine, USA · University of California, San Diego, USA · City University of Hong Kong, Hong Kong, China · University of Pennsylvania, USA ·

Abstract

Modeling scenes using video generation models has garnered growing research interest in recent years. However, most existing approaches rely on perspective video models that synthesize only limited observations of a scene, leading to issues of completeness and global consistency. We propose OmniRoam, a controllable panoramic video generation framework that exploits the rich per-frame scene coverage and inherent long-term spatial and temporal consistency of panoramic representation, enabling long-horizon scene wandering. Our framework begins with a preview stage, where a trajectory-controlled video generation model creates a quick overview of the scene from a given input image or video. Then, in the refine stage, this video is temporally extended and spatially upsampled to produce long-range, high-resolution videos, thus enabling high-fidelity world wandering. To train our model, we introduce two panoramic video datasets that incorporate both synthetic and real-world captured videos. Experiments show that our framework consistently outperforms state-of-the-art methods in terms of visual quality, controllability, and long-term scene consistency, both qualitatively and quantitatively. We further showcase several extensions of this framework, including real-time video generation and 3D reconstruction. Code is available at https://github.com/yuhengliu02/OmniRoam.

Figures & tables

Figure 1. Two-stage Long-horizon Panoramic Scene Wandering . Starting from an input frame/video, OmniRoam first produces a coarse 480 $\times$ 960 preview with accelerated playback speed to quickly traverse the scene along the user-specified camera trajectory (left). It then upscales this video to a high-quality 720 $\times$ 1440 resolution (middle), jointly improving spatial details and temporal coherence, creating an immersive exploration experience. The resulting wandering path is visually represented as a series of viewpoints within the 3D scene (right).

Figure 2. OmniRoam Pipeline consists of (1) a preview stage that takes an input image/video and a user-specified camera trajectory, and generates an accelerated, mid-resolution panoramic video via decomposed trajectory conditioning (flow + scale), and (2) a refine stage that performs scale alignment and segment-wise diffusion to produce the final high-resolution long-horizon panoramic video. The framework further supports real-time autoregressive preview and downstream 3D reconstruction. Both models in preview and refine stage are finetuned from a pretrained Diffusion Transformers.

Figure 3. Dataset Visualization. We curate a hybrid training set containing real-world handheld panoramic videos captured across diverse scenes (top) and synthetic panoramic sequences rendered from 3DGS scenes with synthesized wide-ranging, physically feasible trajectories (bottom).

Method	Resolution	No. of Frames	Visual Quality			Trajectory Controllability			Loop Consistency
Method	Resolution	No. of Frames	FAED $\downarrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR’25 $\uparrow$	PSNR’55 $\uparrow$	PSNR’75 $\uparrow$	Loop Consistency
Matrix-3D	480p	81	8.64	0.63	0.37	19.05	17.35	17.04	1.38
Imagine360	480p	81	23.35	0.33	0.61	–	–	–	–
Ours	480p	81	5.27	0.70	0.18	23.06	20.69	19.87	2.34
Matrix-3D	720p	81	9.76	0.66	0.36	19.06	17.63	17.12	1.41
Ours	720p	641	5.07	0.66	0.33	19.75	18.59	18.24	1.96

Table 1. Quantitative Comparison. Our method consistently outperforms prior approaches ( Yang et al., 2025 ; Tan et al., 2024 ) across all evaluated metrics, achieving superior visual quality, stronger trajectory controllability, and higher loop consistency. FAED, SSIM, and LPIPS are evaluated on 81 frames, while loop consistency is reported on the full sequences.

Figure 4. Qualitative Comparisons. Our method generates panoramic videos that closely follow the input trajectory and preserve sharper details from the input images, while prior methods often produce blurry artifacts or semantically inconsistent content. Zoomed-in perspective crops are shown in the red inset.

Method	No. of Frames	Visual Quality			Trajectory Controllability				Loop Consistency
Method	No. of Frames	FAED $\downarrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR’25 $\uparrow$	PSNR’55 $\uparrow$	PSNR’615 $\uparrow$	PSNR’635 $\uparrow$	Loop Consistency
Perspective (Preview)	81	16.90	0.62	0.44	17.72	16.06	–	–	1.70
Perspective (Refine)	641	15.48	0.63	0.57	16.60	14.94	13.76	14.07	1.42
Autoregressive	641	16.04	0.33	0.64	15.44	14.84	10.14	10.11	0.89
Ours	641	7.70	0.58	0.44	19.75	18.59	15.55	15.63	1.96

Table 2. Design Analysis. We analyze key design choices, including video representation (ours: panoramic vs. perspective) and generation strategy (ours: global-to-local vs. direct autoregressive). We report FAED, SSIM, LPIPS, and loop consistency over the full video sequences. For long videos (641 frames), we additionally report the average PSNR over extended temporal windows, specifically frames 610–615 (PSNR’615) and 630–635 (PSNR’635).

Figure 5. CLIP Similarity over Loop Trajectories. We show temporal CLIP similarity to the first frame under loop trajectories for long-horizon generation (641 frames). Our method shows the expected trend: similarity decreases as the camera moves away and recovers as the trajectory closes the loop; whereas (a) the autoregressive variant exhibits a largely monotonic decline (drift from the initial view) and (b) the perspective-video variant shows weaker recovery with noticeable structural degradation. Representative frames are shown alongside the curves.

Figure 6. Additional Qualitative Comparisons. Our method maintains better geometric stability and finer textures during scene exploration.

Figure 7. Results from Real-Time Previewer-based Two-Stage Generation. Videos are generated in real time at $480\text{p}$ and subsequently refined to $720\text{p}$ .

Figure 8. Long Video Generation with Loop Trajectory. We present 641-frame generated video sequences that traverse the scene in a closed loop, where the last frames closely resemble the initial frame, demonstrating strong long-term consistency.

Panel of Figure 9 — Figure 9. Visualization of Real-time Previewer and 3DGS. (a) demonstrates trajectory-following panoramic wandering with our real-time previewer. (b) shows the reconstructed 3D Gaussian splatting results and corresponding panoramic samples.

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 364 papers

PanoWorld: Geometry-Consistent Panoramic Video World Modeling
May 14, 2026Le Jiang, Xiangyu Bai, Bishoy Galoaa +7Physical Consistency in Video Generation Video Generation
Northeastern University
UCM: Unified Modeling of Camera Control and Memory with Time-aware Positional Encoding Warping for World Models
Feb 26, 2026Tianxing Xu, Zixuan Wang, Guangyuan Wang +5Camera-Controlled Video Generation Video Diffusion Models
Tsinghua University China · Tongyi Lab, Alibaba China · University of Science and Technology of China China
MUGEN: Interactive Panoramic World Exploration via Camera Control
Sep 29, 2026Jiaming Tan, Zhen Li, Shuwei Shi +6Camera-Controlled Video Generation
Alaya Lab · Beijing Institute of Technology · Shanghai Innovation Institute

May 14, 2026J/K move · Enter open · S save

PanoWorld: Geometry-Consistent Panoramic Video World Modeling

Le Jiang, Xiangyu Bai, Bishoy Galoaa, Shayda Moezzi+6

Northeastern University

Physical Consistency in Video Generation Video Generation Temporal Consistency in Video Generation Video World Models

Open paper AskPDF arXiv

We present PanoWorld, a panoramic video world model that generates geometry-consistent 360

\degree

video from a single image and a caption. Existing panoramic video methods optimize primarily for visual realism and do not explicitly constrain the underlying 3D scene state, producing outputs that appear plausible yet exhibit inconsistent depth, broken correspondences, and implausible motion across the spherical surface. We address this gap by framing panoramic video generation as a geometry- and dynamics-consistent latent state modeling problem rather than pure visual synthesis. Building on a pre-trained perspective video world model, we introduce two lightweight regularizers: a depth consistency loss against pseudo ground-truth panoramic depth, and a trajectory consistency loss that supervises the 3D world-frame positions of tracked points across time. We further apply spherical-geometry-aware adaptation to the conditioning and positional encoding. We additionally introduce PanoGeo, a unified geometry-aware panoramic video dataset with consistent depth, trajectory, and prompt annotations across diverse real and synthetic sources, used for both training and stratified evaluation. Experiments show that PanoWorld improves geometric consistency over prior panoramic generation methods while maintaining competitive visual realism, establishing that panoramic video generation must be treated as a geometric modeling problem to support the holistic spatial understanding requirements of embodied AI applications. Code is available at https://github.com/ostadabbas/PanoWorld.

Abstract

Video Generation

Explore similar work

PanoWorld: Geometry-Consistent Panoramic Video World Modeling

UCM: Unified Modeling of Camera Control and Memory with Time-aware Positional Encoding Warping for World Models

MUGEN: Interactive Panoramic World Exploration via Camera Control