publications

2026

1. arXiv

   

   MUSE: Agentic 3D Scene Authoring via Memory-Grounded Incremental Requirement Satisfaction

   Ruijie Xu, Xinnan Zhu, Jiayu Ying, Daoguo Dong, Yuzhou Ji, and Xin Tan

   arXiv preprint arXiv:2606.14168Ruijie Xu and Xinnan Zhu contributed equally. , 2026

   Abs arXiv Bib Website

   Text-driven 3D scene generation is promising for digital content creation, embodied AI simulation, and interactive design, yet practical workflows require refining existing scenes while preserving non-target content. MUSE formulates controllable 3D scene authoring as incremental requirement satisfaction and introduces a memory-grounded multi-agent framework with an Architect, Sculptor, and Inspector. It also introduces AuthorBench for requirement-level controllability and preservation-aware editing.

   @article{xu2026muse,
     title = {MUSE: Agentic 3D Scene Authoring via Memory-Grounded Incremental Requirement Satisfaction},
     author = {Xu, Ruijie and Zhu, Xinnan and Ying, Jiayu and Dong, Daoguo and Ji, Yuzhou and Tan, Xin},
     journal = {arXiv preprint arXiv:2606.14168},
     year = {2026},
   }

2. arXiv

   

   JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space

   Xinnan Zhu, Ruijie Xu, Jiayu Ying, Daoguo Dong, Jiachen Xu, Yuan Xie, and 1 more author

   arXiv preprint arXiv:2606.13345Xinnan Zhu and Ruijie Xu contributed equally. , 2026

   Abs arXiv Bib Website

   JointEdit3D performs feed-forward 3D scene editing in a unified RGB-geometry latent space. It uses asymmetric latent inpainting and a SceneAnchor Branch to propagate an edit from one reference frame while preserving source-scene structure. The work also introduces the SceneEdit3D-15K paired dataset and the SceneEdit3D-Bench evaluation benchmark.

   @article{zhu2026jointedit3d,
     title = {JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space},
     author = {Zhu, Xinnan and Xu, Ruijie and Ying, Jiayu and Dong, Daoguo and Xu, Jiachen and Xie, Yuan and Tan, Xin},
     journal = {arXiv preprint arXiv:2606.13345},
     year = {2026},
   }

3. arXiv

   

   TSHA: A Benchmark for Visual Language Models in Trustworthy Safety Hazard Assessment Scenarios

   Qiucheng Yu, Ruijie Xu, Mingang Chen, Jianfeng Dong, and Xin Tan

   arXiv preprint arXiv:2603.29759, 2026

   Abs arXiv Bib

   TSHA is a benchmark for trustworthy indoor safety hazard assessment with vision-language models. Its current release contains 66,668 validated question-answer pairs across images, videos, and panoramas, covering 12 parent hazard groups and 64 fine-grained hazard types. Evaluation of 22 popular VLMs shows that current models still lack robust safety hazard assessment capabilities.

   @article{yu2026tsha,
     title = {TSHA: A Benchmark for Visual Language Models in Trustworthy Safety Hazard Assessment Scenarios},
     author = {Yu, Qiucheng and Xu, Ruijie and Chen, Mingang and Dong, Jianfeng and Tan, Xin},
     journal = {arXiv preprint arXiv:2603.29759},
     year = {2026},
   }

2025

1. Vis. Comput.

   

   OSH-Splat: Optimizable Semantic Hyperplanes for Enhanced 3D Language Feature Gaussian Splatting

   Ruijie Xu, Yuzhou Ji, Xin Tan, and Lizhuang Ma

   The Visual Computer, 2025

   Abs DOI Bib

   OSH-Splat constructs a compact 3D language field for accurate and efficient open-vocabulary queries. It extracts hierarchical semantic information at part, subpart, and whole levels, learns compact semantic features on 3D Gaussians, and optimizes a semantic hyperplane for each text query to improve localization accuracy and robustness.

   @article{xu2025oshsplat,
     title = {OSH-Splat: Optimizable Semantic Hyperplanes for Enhanced 3D Language Feature Gaussian Splatting},
     author = {Xu, Ruijie and Ji, Yuzhou and Tan, Xin and Ma, Lizhuang},
     journal = {The Visual Computer},
     volume = {41},
     number = {13},
     pages = {11127--11137},
     year = {2025},
     doi = {10.1007/s00371-025-04091-5},
   }