Haotian Xue

About & Research Interests

I'm broadly interested in Generative AI across vision and robotics, with a current focus on physical understanding and physical alignment of Vision–Language Models and Video Diffusion Models, through post-training. Earlier in my PhD I worked on adversarial attacks & defenses for GenAI and developed algorithms across diffusion models and diffusion-based policy: Diff-PGD, SDS-Attack, PDM-Pure, and DP-Attacker.

Previously, I completed research internships at Adobe Firefly (2025), NVIDIA DIR (2024), and Microsoft Research Asia (2021). I was also a visiting student at MIT CSAIL (2021). I earned my B.E. in Computer Science (Honors) from Shanghai Jiao Tong University (Zhiyuan College) in 2022.

News

• 2025-04 Joining Adobe Firefly (Summer+Fall) to work on post-training for video diffusion.
• 2024-10 NeurIPS 2024 Scholar Award — see you in Vancouver!
• 2024-09 Three papers accepted to NeurIPS 2024: DP-Attacker, RefDrop, QueST.
• 2024-05 Started summer research intern at NVIDIA DIR Group.
• 2024-04 Released PDM-Pure, a universal purifier against diffusion models.
• 2024-03 ICLR 2024 Travel Award.
• 2024-01 SDS-Attack accepted to ICLR 2024.
• 2023-10 NeurIPS 2023 Scholar Award; invited reviewer for TPAMI.
• 2023-09 Diff-PGD and 3D-IntPhys accepted to NeurIPS 2023.
• 2023-08 Invited reviewer for ICLR 2024.
• 2023-05 Proposed Diff-PGD, a diffusion-based adversarial sample framework.
• 2022-12 Selected as a Top Reviewer of NeurIPS 2022.
• 2022-10 Distance-Transformer accepted to EMNLP 2022 Findings.
• 2022-08 Started PhD at ML@GT.

Selected Publications

• Point-It-Out: Benchmarking Embodied Reasoning for Vision Language Models in Multi-Stage Visual Grounding
  Haotian Xue, Yunhao Ge, Yu Zeng, Max Li, Ming-Yu Liu, Yongxin Chen, Jiaojiao Fan
  In submission
  VLM Robotics
  arXiv · project
• RefDrop: Controllable Consistency in Image or Video Generation via Reference Feature Guidance
  Jiaojiao Fan, Haotian Xue, Qinsheng Zhang, Yongxin Chen
  NeurIPS 2024
  Diffusion Vision
  arXiv · project
• QueST: Self-Supervised Skill Abstractions for Learning Continuous Control
  Arthava Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, Animesh Garg
  NeurIPS 2024
  Robotics Vision
  arXiv · project · code
• Diffusion Policy Attacker: Crafting Adversarial Attacks for Diffusion-based Policies
  Yipu Chen*, Haotian Xue*, Yongxin Chen
  NeurIPS 2024
  Diffusion Safety
  arXiv · project
• Towards More Effective Protection Against Diffusion-Based Mimicry with Score Distillation
  Haotian Xue, Chumeng Liang*, Xiaoyu Wu*, Yongxin Chen
  ICLR 2024
  Diffusion Safety
  arXiv · code · poster
• Pixel is a Barrier: Diffusion Models Are More Adversarially Robust Than We Think
  Haotian Xue, Yongxin Chen
  NeurIPSW: SafeGenAI 2024 · arXiv 2024
  Diffusion Safety
  arXiv · code
• Diffusion-Based Adversarial Sample Generation for Improved Stealthiness and Controllability
  Haotian Xue, Alexandre Araujo, Bin Hu, Yongxin Chen
  NeurIPS 2023
  Diffusion Safety
  arXiv · code · poster
• 3D-IntPhys: Towards More Generalized 3D-grounded Visual Intuitive Physics
  Haotian Xue, Antonio Torralba, Joshua B. Tenenbaum, Daniel L. K. Yamins, Yunzhu Li, Hsiao-Yu Tung
  NeurIPS 2023
  Vision 3D
  arXiv · poster

Reviewer Experience