Haipeng Fang 房海鹏

I am a 2nd-year PhD candidate under the supervision of Prof. Sheng Tang and A/Prof. Fan Tang at Institute of Computing Technology, Chinese Academy of Sciences (ICT, CAS). My research interests include:

  • Efficient AI; Model Compression and Acceleration
  • Image and Video Generation; AIGC Technologies
  • Image Understanding: Classification, Detection and Segmentation

Contact: fanghaipeng21s AT ict dot ac dot cn

GitHub    /    Google Scholar    /    ORCID       

✍   Hightlights

- I am currently seeking opportunities for visiting research or postdoctoral positions. I am always available and would be truly grateful if you could offer me a valuable opportunity.

News
12/2025 I received the Lenovo PhD Scholarship.
06/2025 I give a talk on Diffusion Acceleration at the AI Time rehearsal session.
05/2025 I give a talk on Model Compression and Acceleration at the AI Future Forum on Foundation Models and Frontier Technologies.
02/2025 One first-authored paper got accepted by CVPR 2025.
Full List
Publications
In Process
[In Process] FD--: Less Tokens and Blocks Make AI-Generated Image Detection Faster and More Generalizable
Haipeng Fang, Yixing Lu, Yu Li, Qiang Sheng, Juan Cao, and Sheng Tang
In Process
TL;DR: We improve AIGI detection generalization through subtractions by progressively pruning tokens and exiting blocks early, guided by an early pressure loss to reduce shortcut learning. This yields a 2.3x speedup and a 6.3% generalization gain.
[CVPR - 334(444)] ResCa: Residual Caching for Diffusion Transformers Acceleration
Haipeng Fang, Yu Li, Fan Tang, Yixing Lu, Juan Cao, and Sheng Tang
In Process
TL;DR: We observe that many tokens follow similar denoising trajectories. We introduce a “proxy denoising” method that denoises 16 proxy tokens and uses their denoising trajectories to simulate 4096 tokens, achieving up to 5.5x speedup with lossless quality.
2025
[TCSVT 2025] MoAnimate: Bridging the Motion-Oriented Latent Representation Gaps in Human Animation
Haipeng Fang, Fan Tang, Zhihao Sun, Ziyao Huang, Juan Cao, Sheng Tang, and Yongdong Zhang
IEEE Transactions on Circuits and Systems for Video Technology (TCSVT)
Paper TL;DR: We address unguided initialization and inefficient interactions in consistency modeling with MoAnimate, a dual-stream motion framework that improves entity consistency across benchmarks and generalization scenarios.
[CVPR 2025] Attend to Not Attended: Structure-then-Detail Token Merging for Post-training DiT Acceleration
Haipeng Fang, Sheng Tang Juan Cao, Enshuo Zhang, Fan Tang, and Tong-Yee Lee
Proceedings of the 42nd IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2025)
Preprint / Paper /
TL;DR: We analyze the diffusion prior in DiT, identify the locations and degrees of redundancy in DiT, and design a “structure-then-detail” token merging method for post-training diffusion transformer acceleration.
[ICASSP 2025] FR2ViT: Finetuning-free Token Reduction for Dense Prediction Through a Refinement-Reactivation Architecture
Haipeng Fang, Ziheng Wu, Xinyi Zou, Jun Huang, Juan Cao, and Sheng Tang
Proceedings of the 50th IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2025)
Paper
TL;DR: We analyze the conflict between token reduction and dense prediction tasks, and design a Refinement-Reactivation Architecture for dense prediction.
2024
[ACMMM 2024] Rethinking Image Editing Detection in the Era of Generative AI Revolution
Zhihao Sun, Haipeng Fang, Juan Cao, Xinying Zhao, and Danding Wang
Proceedings of the 32nd ACM International Conference on Multimedia (ACMMM 2024)
Paper
TL;DR: We propose the GRE dataset to advance detection of generative regional editing, featuring diverse methods, a multi-modal pipeline, and comprehensive benchmarks to fill gaps in existing research.
Academic Services
Conf. Reviewer/PC Member CVPR, ICASSP, Neural Networks, BMVC
Work Experience
2022.2-2023.3 Research Intern - PAI, Alibaba Cloud
Provided smart cabinet services for AUCMA, Haier, Dahuang'e, and EasyTouch
  • Designed SKU decision logic with static-dynamic analysis and generative augmentation
  • Integrated PAI-EAS, OSS, Redis for data flow and PaaS support

    • Contributed to PAI-EasyCV development and testing
  • Benchmarked detectors with PAI-Blade and TorchAcc
  • Extended EfficientFormer, MLP-Mixer, Shuffle Transformer, Hydra Attention
    		•
    Education
    2021-now Ph. D Candidate in Computer Applied Technology
    University of Chinese Academy of Sciences (Cultivation Unit: Institute of Computing Technology, CAS)
    2017-2021 B. Eng in Software Engineering (Ranked 1/88)
    School of Computer Science and Cyberspace Security, Xiangtan University
    2014-2017 Zibo No.4 High School
    Competitions
    2021 3rd/1107 in Amap POI, CCF Big Data and Intelligent Computing Competition
    2020 First Prize, National University Green Computing Competition
    2020 Honorable Mention, Mathematical Contest in Modeling (MCM)
    2019 First Prize, High Education Club Cup National Undergraduate Mathematical Modeling Contest
    2019 First Prize, Hunan Collegiate Programming Contest
    Honors and Awards
    2026 Lenovo PhD Scholarship
    2022, 2023 First-Class Academic Scholarship, University of Chinese Academy of Sciences
    2022, 2023, 2024 Merit Student, University of Chinese Academy of Sciences
    2021 Outstanding Graduate, Hunan Municipal Commission of Education
    2020 National Inspirational Scholarship, Ministry of Education of China
    2018 National Scholarship, Ministry of Education of China

    Design: Jon Barron, Limin Yang and Qiang Sheng
    Last updated:    #Visit: (2025/4/2~)