Wonder: Video World Model Done Better
Adobe Research, Johns Hopkins University
* Equal Contribution † Project Lead
Overview
Abstract
Given an image or video, Wonder builds a persistent interactive world that users can navigate by moving the camera, revealing unseen regions, and returning to previously observed areas with constant latency for up to one minute. This requires system-level co-design across control representation, memory mechanism, and training strategy. Wonder turns camera motion into dense visual evidence, retrieves relevant full-fidelity history through sparse attention, and uses a rectified distillation pipeline to preserve control and long-term consistency. Together, these components enable diverse 16 FPS rollouts with coherent geometry, appearance, dynamics, and interactive control across image-to-video and video-conditioned generation.
Method
System-Level Co-Design
Control Signal
Rendering a synthetic camera space with a 3D scaffold and environment map, turning translation and rotation into dense, frame-aligned visual evidence.
Memory Mechanism
Keeping full-fidelity history KV caches and adaptively selecting relevant entries via sparse attention to preserve long-horizon memory at constant latency.
Training Strategy
Improving student capacity with a Mixture-of-Students design and resolving camera drift during distillation via GAN Control Regularization.
Demos
Image-to-Video Game Worlds
Demos
Image-to-Video Cartoon Worlds
Demos
Image-to-Video Real Worlds
Demos
Complex Camera Movement
Demos
Video-to-Video Real Worlds
Demos
Video-to-Video Cartoon & Game Worlds
Comparison
Compare with SOTA Models
Citation
@article{xu2026wonder,
title={Wonder: Video World Model Done Better},
author={Xu, Jiacong and Jiang, Hanwen and Shu, Zhixin and Sunkavalli, Kalyan and Patel, Vishal M. and Mei, Yiqun},
journal={arXiv preprint arXiv:2607.26037},
year={2026},
url={https://arxiv.org/abs/2607.26037}
}Notice: All images and videos on this page are used solely for research and demonstration purposes. Copyrights remain with their respective owners. If you believe any content infringes your rights, please contact us, and we will promptly remove the corresponding example.