PanoWorld: Real-World Panoramic Generation with Long-Horizon Memory

Forum topic · 小凯 · 2026-07-14

Summary

PanoWorld is a panoramic world model addressing the long-horizon memory challenge in real-world 360-degree scene generation, exploiting the rotation-equivariant nature of omnidirectional representations. The method fixes the camera heading so trajectories reduce to translation, and introduces two key components: Dense Panoramic Ray Conditioning (DPRC) and Geometry-aware Memory Augmentation (GMA). A three-stage training pipeline progressively refines each component. The authors also build World360, a dataset combining panoramic videos captured by real drones with simulated clips generated in AirSim360. Experiments show PanoWorld substantially outperforms alternative approaches. This work is by Haoyuan Li, Dizhe Zhang, and colleagues including Ming-Hsuan Yang and Lu Qi, available on arXiv as 2607.09661.

Overview

Research area: Computer Vision Authors: Haoyuan Li, Dizhe Zhang, Yuemei Zhou, Xiangkai Zhang, Haoran Feng, Xiaofan Lin, Wenjie Jiang, Bo Du, Ming-Hsuan Yang, Lu Qi arXiv: 2607.09661

Key Contributions

Long-horizon memory in panoramic world models: The paper tackles the challenge of maintaining consistent, long-range memory when generating real-world panoramic scenes, leveraging the rotation-equivariant properties of omnidirectional representations.
Fixed-heading trajectory simplification: By fixing the camera heading, the camera trajectory is reduced to pure translation, simplifying modeling of panoramic scene dynamics.
Dense Panoramic Ray Conditioning (DPRC): A conditioning mechanism that provides dense geometric guidance across the full panoramic field of view.
Geometry-aware Memory Augmentation (GMA): A memory module designed to preserve geometric consistency over long horizons.
Three-stage training pipeline: Training progressively refines each component across stages.

Dataset: World360

The authors construct World360, a dataset combining:

Real-world panoramic videos captured by drones
Simulated panoramic clips generated with AirSim360

Results

Experiments demonstrate that PanoWorld significantly outperforms alternative baseline approaches on panoramic generation tasks.

---

*Source: zhichai.net forum post, auto-collected 2026-07-14.*