MUGEN: Interactive Panoramic World Exploration via Camera Control

SIGGRAPH Asia 2026

MUGEN brings together a large-scale real-world panoramic video dataset and Wan360, a model for generating immersive 360° videos with camera control from a single panoramic image. The dataset provides semantic and geometric annotations, while Wan360 adapts a perspective video backbone to panoramic geometry and user-specified camera trajectories.

MUGEN dataset

The project reports 1,318 hours of curated panoramic footage with at least 4K source resolution, organized into 60-second clips, and a 300-hour MUGEN-HQ training subset. Annotations cover captions, scene attributes, locations, camera trajectories, instance masks and depth.

The current CSV release contains 77,023 clips (approximately 1,283.7 hours), including 17,655 HQ clips. Captions and category annotations are included in the CSVs; camera trajectories are available on Hugging Face. Depth is pending release. See the dataset guide for the current release contents and reconstruction workflow.

Wan360 model

Wan360 uses three parameter-free components—periodic longitude RoPE, ERP-aware padding and random roll yaw—to handle cyclic seams and yaw consistency in equirectangular projection (ERP). A panoramic Plücker embedding encodes camera motion using spherical ERP rays.

The project page demonstrates forward, backward, left and right camera translations from the same input panorama. Follow the model guide to run inference.

Getting started

Component	Guide	What you can do
Wan360 model	Model guide	Set up inference and generate forward, backward, left and right camera translations.
MUGEN dataset	Dataset guide	Access annotations, download source videos and reconstruct clips following the Sekai workflow.

MUGEN/
├── model/       # Wan360 inference code and example image
├── dataset/     # Dataset download, reconstruction and annotation tools
└── README.md    # Project overview

Each guide specifies its own environment and working directory.

Resources

Resource	Link / status
Project page and demos	Interactive Panoramic World Exploration via Camera Control
Paper	arXiv
Dataset and annotations	AlayaLab/MUGEN on Hugging Face
Wan360 weights	AlayaLab/Wan360 on Hugging Face

Acknowledgements and licenses

Dataset tools are adapted from Sekai. Model code is based on DiffSynth-Studio.

• Dataset code: dataset/LICENSE, including its non-commercial research restriction.
• Model code: model/LICENSE, Apache-2.0.
• Example photograph: Captured by the MUGEN team and released under CC BY 4.0.
• Wan360 weights: model/LICENSE-WEIGHTS, Apache-2.0.
• MUGEN annotations: dataset/LICENSE-ANNOTATIONS, CC BY 4.0, including captions, category annotations, camera trajectories and depth when released.

The annotation license covers only rights held by the MUGEN contributors in the annotations and database; it does not license source videos or reconstructed video clips. Third-party assets retain their original licenses and notices. See model/THIRD_PARTY_NOTICES.md for the bundled model components.

Citation

@misc{tan2026mugeninteractivepanoramicworld,
      title={MUGEN: Interactive Panoramic World Exploration via Camera Control},
      author={Jiaming Tan and Zhen Li and Shuwei Shi and Minggui Teng and Siqi Yang and Yuwei Wu and Bo Zheng and Chuanhao Li and Kaipeng Zhang},
      year={2026},
      eprint={2609.38077},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.38077},
}