Zhu Jun
Last reviewed
Sources
23 citations
Review status
Source-backed
Revision
v1 · 1,731 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
23 citations
Review status
Source-backed
Revision
v1 · 1,731 words
Add missing citations, update stale details, or suggest a clearer explanation.
Zhu Jun (Chinese: 朱军; born 1983), published in English as Jun Zhu, is a Chinese computer scientist and machine learning researcher. He is the Bosch AI Professor in the Department of Computer Science and Technology at Tsinghua University, deputy director of the university's Institute for Artificial Intelligence, and head of the Tsinghua Statistical Artificial Intelligence and Learning (TSAIL) group.[1][2] His research spans probabilistic (Bayesian) machine learning, adversarial robustness, and diffusion models, where his group produced the Analytic-DPM and DPM-Solver samplers and the U-ViT transformer backbone that underpins the Vidu video generator.[2]
Zhu is the founder and chief scientist of Shengshu Technology, the Beijing startup behind Vidu, and co-founder and chief scientist of RealAI, a Tsinghua-incubated company focused on safe and reliable AI.[3][4] He was elected an IEEE Fellow in 2023, an AAAI Fellow in 2024, and an ACM Fellow in 2025.[4][5][6]
Zhu was born in 1983 in Funan County, Anhui Province, and entered Tsinghua University's Department of Computer Science and Technology in 2001.[7] He received his B.E. in 2005 and his Ph.D. in 2009, both from Tsinghua, where his doctoral advisor was Bo Zhang, a Chinese Academy of Sciences academician and a pioneer of Chinese AI research.[2][7]
After his doctorate he worked at Carnegie Mellon University's Machine Learning Department as a postdoctoral researcher and project scientist in the group of Eric Xing, then returned to Tsinghua as faculty in 2011 at Zhang's invitation.[7][8] He remained an adjunct faculty member at Carnegie Mellon from 2015 to 2018.[1][2]
At Tsinghua, Zhu directs the Tsinghua-Bosch Joint Center for Machine Learning and serves as deputy director of the Institute for Artificial Intelligence, established under Bo Zhang's deanship, where he also leads the institute's Basic Theory Research Center.[1][2] He has served as associate editor-in-chief of IEEE Transactions on Pattern Analysis and Machine Intelligence, on the editorial board of the journal Artificial Intelligence, and more than 20 times as (senior) area chair for NeurIPS, ICML, ICLR, IJCAI, and AAAI.[1][2]
Zhu's early work combined Bayesian modeling with max-margin learning, two traditions that had largely developed separately. With Eric Xing he developed MedLDA, a maximum-margin topic model, and he later generalized the approach into a framework of regularized Bayesian inference that relaxes the classical prior-likelihood formulation by introducing posterior constraints.[8] MIT Technology Review described him as a researcher who committed to Bayesian AI "off the beaten track" before deep learning made probabilistic methods fashionable again.[8]
In 2017 his group released ZhuSuan, an open-source probabilistic programming library named after the Chinese abacus and built on TensorFlow, designed to combine Bayesian methods with deep learning and to scale inference across GPUs.[9] The group also maintains Tianshou, a widely used PyTorch reinforcement learning library.[1]
Zhu's team is known for work on adversarial examples. In the NIPS 2017 Adversarial Attacks and Defences Competition organized by Google Brain researchers, the TSAIL team (including Yinpeng Dong, Fangzhou Liao, Ming Liang, Tianyu Pang, Xiaolin Hu, and Zhu) won first place in all three tracks, using momentum-based iterative attacks for the two attack tracks and a high-level representation guided denoiser (HGD) for the defense track; both methods were published at CVPR 2018.[10] The group later released Ares, an adversarial robustness benchmarking platform presented as a CVPR 2020 oral paper, and in 2024 co-developed MultiTrust, a benchmark that evaluates the trustworthiness of multimodal large language models across truthfulness, safety, robustness, fairness, and privacy.[1][11]
From around 2021 Zhu's group became one of the most influential diffusion-model labs outside the United States. Analytic-DPM, led by his student Fan Bao, derived an analytic estimate of the optimal reverse variance in diffusion probabilistic models and received an Outstanding Paper Award at ICLR 2022.[12] DPM-Solver, a dedicated high-order ODE solver that generates high-quality samples in roughly 10 to 20 network evaluations instead of hundreds, was accepted at NeurIPS 2022, and together with its successor DPM-Solver++ became a standard fast-sampling option in open-source Stable Diffusion tooling.[13]
In September 2022 the group posted U-ViT ("All are Worth Words: A ViT Backbone for Diffusion Models"), which replaced the conventional U-Net backbone with a Vision Transformer that treats time, condition, and noisy image patches uniformly as tokens with long skip connections; it was accepted at CVPR 2023.[14] The preprint appeared roughly three months before the Diffusion Transformer (DiT) paper that later underpinned OpenAI's Sora, and U-ViT became the architectural basis for Shengshu's video models.[3][14] In March 2023 the group scaled the design into UniDiffuser, a single transformer that fits marginal, conditional, and joint distributions across text and images.[15] Zhu's 2024 Tan Kah Kee Young Scientist Award citation recognized this line of work on efficient inference and large-scale training of multimodal diffusion models.[2]
Zhu founded Shengshu Technology in March 2023 with a core team drawn from Tsinghua's Institute for Artificial Intelligence; he serves as chief scientist while remaining at Tsinghua.[3][16] Fan Bao, lead author of the U-ViT and UniDiffuser papers, became chief technology officer; Tang Jiayu, a Tsinghua computer science graduate, was the company's first chief executive and is now its president, with Luo Yihang serving as CEO as of 2026.[3][16][17]
On April 27, 2024, Shengshu and Tsinghua unveiled Vidu at the Zhongguancun Forum in Beijing, with Zhu presenting the system on stage. Built on the U-ViT architecture, it generated 16-second 1080p clips from text and was promoted as China's first long-duration, high-consistency video model and a domestic answer to Sora.[17] The product opened to global users in July 2024, followed by successive Vidu Q1, Q2, and Q3 model generations; the company says Vidu Q3 ranked first in China and second worldwide on Artificial Analysis video leaderboards as of February 2026.[3][18][19] In December 2025, Shengshu and Tsinghua jointly released and open-sourced TurboDiffusion, an acceleration framework aimed at real-time video generation.[20]
Shengshu announced a Series A+ round of more than RMB 600 million in February 2026, and in April 2026 raised an approximately RMB 2 billion (about US$290 million) Series B led by Alibaba Cloud to develop general world models that connect video generation with robotics and autonomous driving.[3][21]
Zhu frames much of his agenda around what Bo Zhang and colleagues call third-generation AI: in his formulation, first-generation AI was knowledge-driven and second-generation AI data-driven, while the third generation must combine both to become safe and reliable.[22] He has argued publicly that problems such as adversarial attacks, data poisoning, and privacy leakage require systematic responses at the technical, legal-regulatory, and societal levels.[22]
In July 2018 he co-founded RealAI with his former doctoral student Tian Tian, who serves as CEO; incubated from Tsinghua's Institute for Artificial Intelligence, the company has been described by Tsinghua as China's first high-tech enterprise dedicated to safe and reliable AI.[4][7][23] RealAI commercializes Zhu's adversarial robustness research, including a facial anti-spoofing "firewall" deployed by banks and detection tools for deepfakes and other AI-generated content, and has served more than 200 government and financial clients.[23] In 2019 Zhu also co-launched the Secure AI Challenger Program with Alibaba, a long-running attack-and-defense competition series that has drawn over 10,000 participants.[23] For translating safe-AI research into practice he received the 25th Qiushi Outstanding Young Achievement Transformation Award in 2023.[23]
| Year | Honor |
|---|---|
| 2013 | IEEE Intelligent Systems "AI's 10 to Watch"; NSFC Excellent Young Scholar; CCF Young Scientist Award [1] |
| 2015 | National Youth Top-notch Talent Support Program [1] |
| 2017 | MIT Technology Review Innovators Under 35 (China) [2][8] |
| 2022 | ICLR Outstanding Paper Award for Analytic-DPM [2][12] |
| 2023 | IEEE Fellow, for contributions to machine learning and its applications [4] |
| 2023 | Qiushi Outstanding Young Achievement Transformation Award [23] |
| 2024 | AAAI Fellow, for significant contributions to the theory and practice of machine learning [6] |
| 2024 | Tan Kah Kee Young Scientist Award in information sciences [2] |
| 2025 | ACM Fellow, for contributions to the theory and methods of probabilistic machine learning [5] |
He has also received the Tencent-backed Xplorer Prize for young scientists.[4]