Jiajun Liang

AI 2.0 (GenAI): Currently I build and deploy RLHF systems for multimodal video generation at Kling, covering preference data pipelines, reward modeling, and scalable and stable reinforcement learning.

AI 1.0 (Perception & Recognition): Previously, I led an algorithm team at Megvii and Jiiov, focusing on visual perception systems, including face, hand, finger, and human-centric 2D & 3D understanding and recognition, deployed over billion mobile devices for everyday real-world usage. In parallel, I have extensive research experience with publications at top-tier AI and vision conferences; see my Google Scholar.

I am looking for self-motivated interns for mulimodal understanding and generation research! Drop me an email if you are interested.

News

[2025.12] Kling-Omni Technical Report is released. [Link]
[2025.12] Kling Launch Week has come! Including Omni, Sound, Avatar and more! [Youtube]
[2025.10] GRPO-Guard for tackling over-optimization in flow matching is released. [Project][Paper]
[2025.09] Three papers have been accepted by NeurIPS 2025!
[2025.09] VR-Thinker reward for video generation is released. [Project] [Paper]
[2025.06] EvoSearch for video test-time scaling is released. [Project][Paper]
[2025.05] We release Flow-GRPO, the first method to integrate online policy gradient reinforcement learning (RL) into flow matching models. [Project][Paper]
[2025.05] One paper has been accepted by ICML 2025!
[2025.04] Kling 2.0 is released! [Link]
[2025.01] Video-Align is released. we develop a systematic pipeline that leverage human feedback to improve video generation. [Project][Paper]

About me

I am Jiajun Liang, I was born in Zhongshan, Guangdong Province, the hometown of Sun Yat-sen, and I now live in Beijing.
I received my M.S. degree from Tsinghua University in 2017 and my B.E. degree from Huazhong University of Science and Technology in 2014.
My research interests lie in multimodal understanding and generation, with a particular focus on data curation pipelines and reinforcement-learning based post-training.

Working Experience

Kuaishou Kling, Multimodal Video Generation (2024 – Present)
Megvii, Algorithm Director (2016 – 2024)

Selected Publications

See full list at Google Scholar

▲ Video Generation / RLHF

Flow-GRPO: Training Flow Matching Models via Online RL

J. Liu, G. Liu, Jiajun Liang, Y. Li, J. Liu, X. Wang, P. Wan, D. Zhang, W. Ouyang

NeurIPS, 2025 [Paper] [Code]

GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping

J. Wang, Jiajun Liang, J. Liu, H. Liu, G. Liu, J. Zheng, W. Pang, A. Ma, Z. Xie, X. Wang

Preprint, 2025 [Paper] [Code]

Improving Video Generation with Human Feedback

J. Liu, G. Liu, Jiajun Liang, Z. Yuan, X. Liu, M. Zheng, X. Wu, Q. Wang, M. Xia, X. Wang

NeurIPS, 2025 [Paper] [Code]

Scaling Image and Video Generation via Test-Time Evolutionary Search

H. He, Jiajun Liang, X. Wang, P. Wan, D. Zhang, K. Gai, L. Pan

NeurIPS Workshop, 2025 [Paper] [Code]

VR-Thinker: Boosting Video Reward Models through Thinking-with-Image Reasoning

Q. Wang, J. Liu, Jiajun Liang, Y. Jiang, Y. Zhang, J. Chen, Y. Zheng, X. Wang, P. Wan, X. Yue

Preprint, 2025 [Paper] [Code]

▲ Diffusion Models

LEDiT: Your Length-Extrapolatable Diffusion Transformer without Positional Encoding

S. Zhang, S. Liang, Y. Tan, Z. Chen, L. Li, G. Wu, Y. Chen, S. Li, Z. Zhao, C. Chen, Jiajun Liang, Y. Tang

NeurIPS, 2025 [Paper] [Code]

MegActor-Sigma: Unlocking Flexible Mixed-Modal Control in Portrait Animation with Diffusion Transformer

S. Yang, H. Li, J. Wu, M. Jing, L. Li, R. Ji, Jiajun Liang, H. Fan, J. Wang

AAAI, 2025 [Paper] [Code]

MegActor: Harnessing Diffusion Models for High-Fidelity Human Animation

S. Yang, H. Li, J. Wu, M. Jing, L. Li, R. Ji, Jiajun Liang, H. Fan

Preprint, 2024 [Paper] [Code]

HiDiffusion: Unlocking High-Resolution Creativity and Efficiency in Low-Resolution Trained Diffusion Models

S. Zhang, Z. Chen, Z. Zhao, Y. Tang, Y. Chen, W. Cao, Jiajun Liang

ECCV, 2023 [Paper] [Project]

▲ Knowledge Distillation/Efficient AI

Efficient One-Pass Self-Distillation with Zipf’s Label Smoothing

Jiajun Liang, L. Li, Z. Bing, B. Zhao, Y. Tang, B. Lin, H. Fan

ECCV, 2022 [Paper] [Code]

Decoupled Knowledge Distillation

B. Zhao, Q. Cui, R. Song, Y. Qiu, Jiajun Liang

CVPR, 2022 [Paper] [Code]

Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers

S. Wei, T. Ye, S. Zhang, Y. Tang, Jiajun Liang

CVPR, 2023 [Paper] [Code]

Cumulative Spatial Knowledge Distillation for Vision Transformers

B. Zhao, R. Song, Jiajun Liang

ICCV, 2023 [Paper] [Code]

DOT: A Distillation-Oriented Trainer

B. Zhao, Q. Cui, R. Song, Jiajun Liang

ICCV, 2023 [Paper] [Code]

Asymmetric Decision-Making in Online Knowledge Distillation

Z. Chen, B. Zhao, Y. Ge, Y. Chen, R. Song, Jiajun Liang

ICML, 2025 [Paper]

▲ Vision / Recognition

EAST: An Efficient and Accurate Scene Text Detector

X. Zhou, C. Yao, H. Wen, Y. Wang, S. Zhou, W. He, Jiajun Liang

CVPR, 2017 [Paper] [Code]

Implicit Identity Leakage: The Stumbling Block to Improving Deepfake Detection Generalization

S. Dong, J. Wang, R. Ji, Jiajun Liang, H. Fan, Z. Ge

CVPR, 2023 [Paper] [Code]

A Simple Baseline for Efficient Hand Mesh Reconstruction

Z. Zhishan, Z. Shihao, L. Zhi, Z. Minqiang, T. Yao, Jiajun Liang

CVPR, 2024 [Paper] [Code]