MUGEN / Wan360

MUGEN

Interactive Panoramic World Exploration via Camera Control

SIGGRAPH Asia 2026

Jiaming Tan1,2,*Zhen Li1Shuwei Shi1Minggui Teng1Siqi Yang1
Yuwei Wu2,†Bo Zheng1Kaipeng Zhang1,3,†Chuanhao Li1,†

1 Alaya Lab2 Beijing Institute of Technology3 Shanghai Innovation Institute

* Work done during an internship in Alaya Lab.   † Corresponding authors.

Paper Code & Data
Project introduction
MUGEN combines high-resolution panoramic videos with camera trajectories, instance masks, depth maps, and semantic annotations.
A data foundation and a camera-controllable model for panoramic world exploration.

Overview

Explore beyond the frame.

Interactive panoramic video generation aims to synthesize immersive 360° videos that remain visually coherent while following user-specified camera trajectories during exploration. However, progress is limited by a coupled data-and-model gap: existing panoramic video datasets are often short, weakly annotated, or lack camera trajectories, while existing camera-controlled video generation models are designed for perspective videos and do not directly support panoramic geometry.

We introduce MUGEN, a large-scale real-world panoramic video dataset with over 1,300 hours of at least 4K videos and rich semantic and geometric annotations, and Wan360, a camera-controllable panoramic video generation model. Wan360 adapts a perspective video backbone with three parameter-free ERP-aware components and a panoramic Plücker embedding, enabling high-quality, temporally coherent 360° video generation.

01 / Generation

One panorama. Four directions.

The same input panorama, with four user-specified camera trajectories: forward, backward, left, and right.

ForwardSame input panorama
BackwardSame input panorama
LeftSame input panorama
RightSame input panorama

Four camera-control results from our fast-forward video, shown in equirectangular projection (ERP).

02 / Dataset

MUGEN

Real-world panoramic videos with captions, scene attributes, locations, camera trajectories, instance masks, and depth.

1,318 hours
Curated panoramic footage
≥ 4K
Source video resolution
60 seconds
Continuous clips
300 hours
MUGEN-HQ training subset
MUGEN curation pipeline: collection, preprocessing, semantic and geometric annotation, quality filtering, and high-quality subset sampling.
From 2,893 hours of source footage to MUGEN: systematic annotation and quality filtering, followed by stratified sampling for MUGEN-HQ.

03 / Method

Wan360

Adapting perspective video generation to the geometry of a 360° world.

Wan360 architecture with random roll yaw, ERP-aware padding, periodic longitude RoPE, and panoramic Plücker embeddings.

Panorama-aware generation

Periodic longitude RoPE, ERP-aware padding, and random roll yaw address cyclic seams and yaw consistency without adding trainable parameters.

Camera trajectory control

A panoramic Plücker embedding represents camera motion using spherical ERP rays in place of perspective pinhole rays.

Reference

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}, 
}