Xiangpeng Yang

Hi, I'm Xiangpeng. I am currently a Ph.D. student at the University of Technology Sydney (UTS). I am also a research intern with the Tongyi MAI Z-Image Team, advised by Dr. Peng Gao and Prof. Steven Hoi. I am an incoming Student Researcher at Google, working with Dr. Yale Song on video editing improvements for Veo 3.1.

My research interests involve Generative AI, Video Generation, and Multi-modal Learning. Specifically, I focus on video world models, video generation, and multi-modal foundation models.

Looking ahead, I am deeply motivated to build unified video models capable of jointly understanding dynamic visual environments and generating coherent future content within a single framework. I believe this direction is a crucial step toward world models, where systems can reason about and interact with the physical world through continuous video understanding and prediction.

I am seeking full-time opportunities starting in Fall 2026 in image/video generation, world models, embodied AI, and unified models. Please feel free to reach out!

• [05/2026] Gave a talk on VideoCoF at Sydney AI Meetup.
• [03/2026] I received an offer to join Google as a Student Researcher, working with Dr. Yale Song on video editing improvements for Veo 3.1.
• [02/2026] VideoCoF is accepted to CVPR 2026 as a Highlight paper (ranked Top 0.65%)!
• [12/2025] We Release the Code and Model of VideoCoF.
• [07/2025] Gave an invited talk at Sydney AI Meet-Up.
• [01/2025] VideoGrain is accepted to ICLR 2025.
• [12/2023] DGL is accepted to AAAI 2024.

• Z-Multi-Shot Video

  Sole Contributor: Xiangpeng Yang · Tongyi MAI, Z-Image Team

  Independently curated billion-scale data, conducted large-scale training, and proposed a novel scalable learning paradigm.

  Research Project, 2026

  TL;DR: The industry's first model to directly output 60-second multi-shot audio-video, surpassing Seedance 2.5's 30 seconds and Wan 3.0's 30 seconds.

  Demo Z-Image Team
• VideoCoF: Unified Video Editing with Temporal Reasoner

  Xiangpeng Yang, Ji Xie, Yiyuan Yang, Yue Ma, Yan Huang, Min Xu, Qiang Wu

  Conference on Computer Vision and Pattern Recognition (CVPR), 2026 Highlight · Top 0.65%

  TL;DR: A Chain-of-Frames reasoning framework for unified video editing with 16× length extrapolation.

  ArXiv Project Code HF Daily Papers · Top 3 Video Dataset media[机器之心]
• ActAlign: Action-Aware Representation Alignment for Generalizable Vision-Language-Action Models

  Yiyuan Yang, Guodong Long, Xiangpeng Yang, Dikai Liu, Qinghua Lu, Jing Jiang

  AAAI Conference on Artificial Intelligence (AAAI), 2027 (In Submission)

  TL;DR: Action-aware representation alignment for generalizable embodied VLA models.

  Paper Demo
• VideoGrain: Modulating Space-Time Attention for Multi-Grained Video Editing

  Xiangpeng Yang, Linchao Zhu, Hehe Fan, Yi Yang

  The International Conference on Learning Representations (ICLR), 2025

  TL;DR: The first work propose Multi-Grained Video Editing, including Class, Instance and Part-level

  Paper Project Code HF Daily Papers · Top 1 media(ak) media(gradio) media(量子位)
• 
  D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

  Dengyang Jiang, Xin Jin, Dongyang Liu, ..., Xiangpeng Yang, ..., Peng Gao, Steven Hoi

  arXiv Preprint, 2026

  TL;DR: On-policy self-distillation enables continuous fine-tuning of few-step diffusion models without sacrificing their original few-step generation capability.

  Paper Project Code
• 
  Z-Reward: Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions

  Xin Jin, Huanqia Cai, ..., Xiangpeng Yang, ..., Steven C. H. Hoi

  Technical Report, 2026

  TL;DR: A reasoning-internalized reward model that distills rubric-aligned score distributions into an efficient VLM for text-to-image post-training.

  Paper Project HF Page
• 
  DGL: Dynamic Global-Local Prompt Tuning for Text-Video Retrieval

  Xiangpeng Yang, Linchao Zhu, Xiaohan Wang, Yi Yang

  AAAI Conference on Artificial Intelligence (AAAI), 2024

  TL;DR: Training only 0.83 MB parameters to achieve better performance than full finetuning.

  Paper Code Video

• [06/2025] "Multimodal SSMs, Multimodal Reasoning, and Multi-Grained Video Editing" at Twelve Labs, hosted by James Le (video).
• [03/2025] "VideoGrain: Exploration and Application of Multi-Grained Video Editing Based on Diffusion Models" at Qingke AI (video).

Peer reviewer