Efficient Generation
Image · Video · 4D
Make generation faster and cheaper, and less dependent on costly paired data and 3D labels.
What we cut
Biography
I am a Professor at the School of Information Science and Technology, University of Science and Technology of China (USTC), working in Prof. Yongdong Zhang's group. Before joining USTC, I was a postdoctoral researcher at the National University of Singapore (NUS), working with Prof. Angela Yao and Prof. Tat-Seng Chua. I received my Ph.D. in Computer Science from NUS, advised by Prof. Tat-Seng Chua.
My group works on Efficient AI. Generative models, foundation models and agents have mostly become stronger by spending more: more paired data, more parameters, more sampling steps, longer reasoning and more model calls. We ask how to keep the capability while spending less, in three directions: efficient generation of images, videos and 4D scenes; efficient post-training of LLMs and VLMs; and efficient agents for spatial understanding and social simulation.
My earlier research focused on video understanding, including video question answering, causal and grounded video reasoning, and 3D affordance. Our Invariant Grounding for Video Question Answering was a CVPR 2022 Best Paper Finalist. I am a recipient of the NSFC Excellent Young Scientists Fund (Overseas).
Research: Efficient AI
Same capability, smaller bill. Each direction cuts a different part of the cost.
Efficient Post-training
LLMs · VLMs · RL · On-policy distillation
Train models on tasks without a ground-truth verifier, and turn every unit of training compute into learning.
What we cut
Efficient Agents & Applications
Spatial understanding · Social simulation
Agents that reason reliably about 3D space and simulate how large populations respond, at a cost that scales.
What we cut
One thread: smaller models · fewer sampling steps · shorter reasoning · experience in the weights · fewer calls.
Doing More with Less
Recent results in efficient generation. Struck-through numbers are what a standard approach spends; amber is what ours needs.
Parameters at FID 1.42 on ImageNet
675M177M
Pixel-aware post-training
25 minon 8×A100 · 1% of the data
Hand-made image editing pairs
0learns editing from ordinary videos
Paired editing videos
0image edits + ordinary videos
VINCIE-NExTNeurIPS 2026
Recasts video editing as editing one frame and propagating it. Best overall score among open-source video editors in our evaluation (3.08 vs. 2.49).
Gaussians in feed-forward 4D reconstruction
100K+/frame8K/video
Time to animate a mesh from a 32-frame video
60 min5 min
Shape Flow GuidanceECCV 2026
A pretrained 3D generator edits each frame's mesh into the next, adding the 3D supervision that 2D losses lack. Preferred by users 84.8% of the time.
Topics & Selected Papers
Efficient Generation
- Smaller generators: [ICLR 2026 reAR] [CVPR 2026 VA-π] [2026 VPG]
- Fewer sampling steps: [2026 TFRD]
- Learning from data that already exists: [ICLR 2026 VINCIE] [NeurIPS 2026 VINCIE-NExT]
- Compact and fast 3D / 4D: [NeurIPS 2026 GGQR] [ECCV 2026 SFG] [ICLR 2026 Interp3D] [ICML 2026 RelaxFlow · Spotlight] [AAAI 2026 AnchorDS]
- Egocentric and world generation: [ICLR 2026 EgoTwin] [2026 DynFlowDrive]
Efficient Post-training
- Shorter reasoning: [2026 RLCPR]
- Experience into weights: [2026 E2L]
- Small teaches large: [2026 Weak-to-Strong RL Transfer]
Efficient Agents & Applications
- Spatial understanding: [2026 Spatial Harness]
- Long and streaming video: [2026 DRR] [2026 Sekai]
Earlier Work: Video Question Answering
- Causal and grounded VideoQA: [CVPR 2022 IGV · Best Paper Finalist] [TPAMI 2023 TIGV] [ICCV 2023 TranSTR] [ACM MM 2023 RaFormer] [ACM MM 2022 EIGV]
- Surveys, studies and tutorials: [IJCV 2025 VideoQA in the Era of LLMs] [EMNLP 2022 VideoQA Survey] [ACM MM 2025 Tutorial]
News
-
2026 / 11
Won the Gold Medal at iENA 2026 in Nuremberg.
International Trade Fair “Ideas – Inventions – New Products” (纽伦堡国际发明奖金奖), for a safe and controllable video generation system.
-
2026 / 03
Joined USTC as a Professor.
School of Information Science and Technology, University of Science and Technology of China.
-
2026 / 01
Four papers accepted at ICLR 2026.
Join Us
Students and postdocs. I am looking for motivated Ph.D. and Master's students, undergraduate interns and postdocs who want to make generative models, foundation models and agents cheaper to train and run. Please email me your CV and a few lines on what you would like to work on.
欢迎对高效生成模型、大模型后训练、智能体方向感兴趣的同学（博士、硕士、本科实习）和博士后加入，请将简历和感兴趣的方向发送至邮箱。
Industry collaboration. We welcome joint projects on cutting the training and inference cost of generative models, LLMs and agents.
Honors & Awards
- 2026Gold Medal, iENA 2026 International Trade Fair “Ideas – Inventions – New Products”, Nuremberg, for A Safe and Controllable Video Generation System 纽伦堡国际发明奖金奖
- 2025NSFC Excellent Young Scientists Fund (Overseas) 国家自然科学基金优秀青年科学基金项目（海外）
- 2023Dean's Graduate Research Excellence Award, National University of Singapore
- 2022Best Paper Award Finalist, CVPR 2022 (33 of 8,161 submissions)
- 2022Research Achievement Award, National University of Singapore
Service
- Area Chair: CVPR 2027
- Keynote Speaker: ACM Multimedia Workshop
- Tutorial Organizer: Video Question Answering and Beyond (ACM MM 2025); tutorial on video question answering (ACM MM 2026)
- Tutorial Speaker: ACM MM 2025, ACM MM 2026
- Grand Challenge Organizer: ACM Multimedia, three times, including VidVRD 2021
- Reviewer: major computer vision, machine learning and multimedia venues, including ECCV
Experience & Education
- Professor, University of Science and Technology of ChinaSchool of Information Science and Technology
- Postdoctoral Researcher, National University of Singaporewith Prof. Angela Yao and Prof. Tat-Seng Chua
- Ph.D. in Computer Science, National University of SingaporeNExT++ Research Center · Advisor: Prof. Tat-Seng Chua
- M.S., Columbia UniversityRobotics
- B.Eng., Huazhong University of Science and TechnologyMechanical Design, Manufacturing and Automation