LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation

Recent advances in diffusion models have significantly improved text-to-video generation, enabling personalized content creation with fine-grained control over both foreground and background elements. However, precise face-attribute alignment across subjects remains challenging, as existing methods lack explicit mechanisms to ensure intra-group consistency. We propose LumosX, a framework that advances both data and model design to achieve state-of-the-art performance in fine-grained, identity-consistent, and semantically aligned personalized multi-subject video generation.

---

💻 Authors

Jiazheng Xing^1,4,2,*, Fei Du^2,3,*, Hangjie Yuan^2,3,1,*, Pengwei Liu^1,2, Hongbin Xu⁴, Hai Ci⁴, Ruigang Niu^2,3, Weihua Chen^2,3†, Fan Wang², Yong Liu^1†

¹Zhejiang University, ²DAMO Academy, Alibaba Group, ³Hupan Lab, ⁴National University of Singapore

^*Equal contributions  ·  ^†Corresponding authors

Contact: jiazhengxing@zju.edu.cn, kugang.cwh@alibaba-inc.com, yongliu@iipc.zju.edu.cn

📘 Click to view Abstract

Recent advances in diffusion models have significantly improved text-to-video generation, enabling personalized content creation with fine-grained control over both foreground and background elements. However, precise face-attribute alignment across subjects remains challenging, as existing methods lack explicit mechanisms to ensure intra-group consistency. Addressing this gap requires both explicit modeling strategies and face-attribute-aware data resources. We therefore propose LumosX, a framework that advances both data and model design.

On the data side, a tailored collection pipeline orchestrates captions and visual cues from independent videos, while multimodal large language models (MLLMs) infer and assign subject-specific dependencies. These extracted relational priors impose a finer-grained structure that amplifies the expressive control of personalized video generation and enables the construction of a comprehensive benchmark.

On the modeling side, Relational Self-Attention and Relational Cross-Attention intertwine position-aware embeddings with refined attention dynamics to inscribe explicit subject-attribute dependencies, enforcing disciplined intra-group cohesion and amplifying the separation between distinct subject clusters. Comprehensive evaluations on our benchmark demonstrate that LumosX achieves state-of-the-art performance in fine-grained, identity-consistent, and semantically aligned personalized multi-subject video generation.

📜 News

[2026/1/26] Accepted by ICLR 2026 !

[2026/3/21] Code is available in Lumos-Custom / LumosX !

📎 Citation

If you find this work useful, please cite:

@inproceedings{xinglumosx,
  title={LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation},
  author={Xing, Jiazheng and Du, Fei and Yuan, Hangjie and Liu, Pengwei and Xu, Hongbin and Ci, Hai and Niu, Ruigang and Chen, Weihua and Wang, Fan and Liu, Yong},
  booktitle={The Fourteenth International Conference on Learning Representations}
}

📣 Disclaimer

This is the official release channel for LumosX weights.