> _

Hengyuan Xu

AI Researcher · Generative Models · Transformers · ID-Consistency

Master's Student @ Fudan University · FVL Lab · Intern @ Tencent Hunyuan

🚧 🚧 This site is under permanent construction — just like the person behind it.

01 About

Hengyuan Xu

I am a Master's student in FVL Lab at the School of Computer Science, Fudan University, advised by Xingjun Ma and Yu-gang Jiang. I received my B.S. from Shanghai Jiao Tong University. I am currently a research intern at Tencent Hunyuan, working on improving unified model consistency across a large number of reference subjects.

My research spans generative AI, mathematical properties of Transformers, privacy-preserving ML, and model IP protection. I am open to PhD opportunities and research collaborations.

Unified Models Diffusion Models Text-to-Image ID-Consistency Transformers Model Watermarking AI Safety
Open to PhD & research internships

Academic Service

🎤
ICLR 2026 — Oral Presentation
📋
CVPR 2024 — Poster Presentation (Permutation Equivariance)
🔍
Reviewer for TVCG, TIP, and multiple AI conferences.
8+Publications
5+First / Co-first
260+Citations
1K+GitHub Stars

02 Selected Publications

WithAnyone
ICLR 2026 WithAnyone: Towards Controllable and ID Consistent Image Generation Hengyuan Xu, Wei Cheng, Peng Xing, Yixiao Fang, Shuhan Wu, Rui Wang, Xianfang Zeng, Daxin Jiang, Gang Yu, Xingjun Ma, Yu-Gang Jiang 6B FLUX-based IPA model. Million-level paired ID-centric dataset. arXiv Code Demo
PixelSmile
CVPR 2026 PixelSmile: Toward Fine-Grained Facial Expression Editing Jiabin Hua*, Hengyuan Xu*, Aojie Li, Wei Cheng, Gang Yu, Xingjun Ma, Yu-Gang Jiang arXiv Project
Permutation Equivariance
CVPR 2024 Permutation Equivariance of Transformers and Its Applications Hengyuan Xu, Liyao Xiang, Hangyu Ye, Dixi Yao, Pengzhi Chu, Baochun Li Mathematically formulated, extended and proved permutation equivariance in row-wise and column-wise directions. arXiv Code Video
Hufu
Preprint Hufu: A Modality-Agnostic Watermarking System for Pre-Trained Transformers Hengyuan Xu, Liyao Xiang, Xingjun Ma, Borui Yang, Baochun Li arXiv
Preprint A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, and More Xingjun Ma, Yixu Wang, Hengyuan Xu, et al. arXiv 2026. arXiv
Preprint Safety at Scale: A Comprehensive Survey of Large Model Safety Xingjun Ma, et al. (incl. Hengyuan Xu). Trends in Privacy and Security, 2025. arXiv
CVPR 2026 VistoryBench: Comprehensive Benchmark Suite for Story Visualization Cailin Zhuang, et al. (incl. Hengyuan Xu). CVPR 2026. arXiv
ICDM 2022 Privacy-Preserving Split Learning via Patch Shuffling over Transformers Dixi Yao, Liyao Xiang, Hengyuan Xu, Hangyu Ye, Yingqi Chen. ICDM 2022.

04 Experience & Education

2025 — Present Shanghai

Research Intern @ Tencent Hunyuan

Working on improving unified model consistency across a large number of reference subjects.

2024.11 — 2025 Shanghai

Research Intern @ StepFun

Algorithm development for Lipu APP. Open-sourced WithAnyone (ICLR 2026) — collected and processed million-level paired ID-centric datasets, trained 6B FLUX IPA model from scratch. Explored MLLM reflection for text-to-image editing and character reference models.

2022.11 — 2024.06 Shanghai

Research Intern @ John Hopcroft Center, SJTU

Advised by Liyao Xiang. Worked on privacy-preserving deep learning and model IP protection. Explored mathematical properties of Transformers, leading to CVPR 2024 paper on permutation equivariance and its applications in watermarking and split learning.

2024.09 — Present Shanghai

M.S. in Computer Science @ Fudan University

2020.09 — 2024.06 Shanghai

B.S. in SEIEE @ Shanghai Jiao Tong University