Shaoteng Liu

I am an Incoming Research Scientist at Adobe Research. I completed my Ph.D. at CUHK and was a research assistant at the BAIR, Berkeley. My research interests lie in VLMs, Agents, and AIGC, including applications such as image and video generation, editing, and manipulation.

We are hiring self-motivated and creative interns. If you are interested in an Adobe internship or a university collaboration, please feel free to contact me.

Selected Research Full List

	Generative Video Propagation Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, Jiaya Jia CVPR, 2025. arXiv / Project Page / Video / Data / Adobe Firefly Adobe News / Twitter / 机器之心 We demonstrate that through a careful design of a generative video propagation framework, various video tasks can be addressed in a unified way by leveraging the generative power of such models.
	Training-Free Efficient Video Generation via Dynamic Token Carving Yuechen Zhang, Jinbo Xing, Bin Xia, Shaoteng Liu, Bohao Peng, Xin Tao, Pengfei Wan, Eric Lo, Jiaya Jia Preprint, 2025, arXiv / Project Page / Code Jenga accelerates HunyuanVideo by 4.68-10.35× through dynamic attention carving and progressive resolution generation.
	Mini-Gemini: Mining the Potential of Multi-modality Vision Language Models Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, Jiaya Jia Preprint*, 2024. arXiv / Project Page / 机器之心 / Demo / Model / Data / Code Mining potential of open-source VLMs! Mini-Gemini is a novel framework ranges from 2B to 34B VLMs for hi-resolution image understanding. It has a impressive OCR capability, and can generate HQ images powered by its multi-modal reasoning ability.
	RL-GPT: Integrating Reinforcement Learning and Code-as-policy Shaoteng Liu, Haoqi Yuan, Minda Hu, Yanwei Li, Yukang Chen, Shu Liu, Zongqing Lu, Jiaya Jia NeurIPS, 2024. Oral arXiv / Project Page The slow agent decomposes the task and determines "which actions" to learn. The fast agent writes code and RL configurations for low-level execution.
	Direct Inversion: Boosting Diffusion-based Editing with 3 Lines of Code Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, Qiang Xu ICLR, 2024. arXiv / Project Page / Video / Data / Code Rethinking the inversion process. Boosting Diffusion-based Editing with 3 Lines of Code.
	Video-P2P: Video Editing with Cross-attention Control Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, Jiaya Jia CVPR, 2024. Most Influential CVPR Papers (Paper Digest ) arXiv / Project Page/ Twitter/ Code Add 'Lego' attribute to the child, an edited video is generated. Powered by a novel video inversion process and cross-attention control. We also find that a Decoupled-Guidance strategy is essential for video editing.
	Tent: Fully test-time adaptation by entropy minimization Dequan Wang, Evan Shelhamer Shaoteng Liu, Bruno Olshausen, Trevor Darrell ICLR, 2021. Spotlight arxiv/ Code Tent equips a model to adapt itself to new and different data during testing.