Latent Spatial Memory for Video World Models

Weijie Wang^1,*   Haoyu Zhao^1,*   Yifan Yang²   Feng Chen³   Zeyu Zhang¹   Yefei He¹   Zicheng Duan³
Donny Y. Chen⁴   Yuqing Yang²   Bohan Zhuang¹

¹Zhejiang University    ²Microsoft Research    ³Adelaide University    ⁴Monash University

^*Equal contribution

Latent Spatial Memory stores persistent 3D scene content directly as latent tokens.
It avoids repeated RGB rendering and re-encoding from explicit 3D caches, enabling efficient spatial consistency for video world models.

Highlights

Latent Memory Persistent 3D scene context lives directly in latent space.	No RGB Detour Mirage avoids repeated render-and-reencode cache updates.
Memory Lifecycle Initialize, read, denoise, and update across generated chunks.	Efficient Worlds Higher generation efficiency with lower 3D cache memory.

Method

Mirage builds a persistent latent cache from the initial observation. For each generated chunk, it reads target-view memory, uses it during denoising, and writes updated static scene content back to the cache.

Results

10.57x faster generation

55x lower 3D cache memory

70.36 WorldScore average

Citation

If you find this project useful, please cite:

@article{wang2026mirage,
  title   = {Latent Spatial Memory for Video World Models},
  author  = {Wang, Weijie and Zhao, Haoyu and Yang, Yifan and Chen, Feng and Zhang, Zeyu and He, Yefei and Duan, Zicheng and Chen, Donny Y. and Yang, Yuqing and Zhuang, Bohan},
  journal = {arXiv preprint arXiv:2604.24764},
  year    = {2026}
}

Support

See SUPPORT.md for usage support and SECURITY.md for vulnerability reporting.

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution.

For details, visit Contributor License Agreements. When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately. Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information, see the Code of Conduct FAQ or contact opencode@microsoft.com with additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.