Shaoteng Liu

Research Scientist @ Adobe Research | Ph.D. @ CUHK | Visiting @ BAIR, UC Berkeley

Recent Progress

At Adobe, I focus on pre-training and next-generation model architectures, including linear models and state space models.

I am collaborating with NVIDIA on post-training and infrastructure for long-video generation, releasing LongLive-2.0.

During my Ph.D., I focused on post-training and data creation across image/video editing, generation, and VLMs — including EditVerse, GenProp, and Mini-Gemini. In turn, these methods themselves became engines for generating new data.

I have also developed inference-time methods such as Rolling Sink, Jenga, Video-P2P, and Direct Inversion. Efficient methods that extend a model's length, accelerate its speed, and unlock new behaviors.

I have seen some of my work gradually shipped into industry products, including Project Frame Forward and Firefly Image-to-Video.

Recent Events

About

Shaoteng Liu

I am a Research Scientist at Adobe Research (San Jose). My work centers on fundamental performance of visual generation models and enhancing AI creativity. I have been trying to broaden my experience across pre-training, post-training, data, infrastructure, and research, hoping to grow into a well-rounded individual contributor.

I completed my Ph.D. at CUHK, advised by Prof. Jiaya Jia, working on visual generation and multimodal LLMs. Before that, I was a research assistant at BAIR, UC Berkeley, contributing to the early-stage Test-Time Training method Tent. I received my B.Eng. from XJTU, studying Automation (Qian Xuesen Class).

I co-organize the HiGen Workshop at CVPR 2026 and ICCV 2025, and serve as a reviewer for CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML, AAAI, SIGGRAPH, TPAMI, and IJCV.

We are hiring self-motivated and creative interns for Adobe internships and university collaborations. Reach out at shaotengl@adobe.com.

Recent Publications

arXiv 2026

LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation

Yukang Chen, Luozhou Wang, Wei Huang, Shuai Yang, Bohan Zhang, Yicheng Xiao, Ruihang Chu, Weian Mao, Qixin Hu, Shaoteng Liu, Yuyang Zhao, Huizi Mao, Ying-Cong Chen, Enze Xie, Xiaojuan Qi, Song Han

LongLive-2.0 treats algorithm and infrastructure as one system: Balanced SP and NVFP4 make long-video autoregressive fine-tuning practical, while W4A4 execution, NVFP4 KV cache, parallel dequantization, and asynchronous VAE decoding improve end-to-end inference throughput.

arXiv Project Code
arXiv 2026

Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

Haodong Li, Shaoteng Liu, Zhe Lin, Manmohan Chandraker

Rolling Sink scales autoregressive video synthesis to ultra-long durations (5–30 min) at test time with consistent subjects, stable colors, and smooth motion.

arXiv Project Demo Code
Oral Presentation

ICLR Oral 2026

EditVerse: Unifying Image and Video Editing and Generation with In-Context Learning

Xuan Ju, Tianyu Wang, Yuqian Zhou, He Zhang, Qing Liu, Nanxuan Zhao, Zhifei Zhang, Yijun Li, Yuanhao Cai, Shaoteng Liu, Daniil Pakhomov, Zhe Lin, Soo Ye Kim, Qiang Xu

EditVerse unifies a diverse range of generation and editing tasks for both images and videos within a single, powerful model.

arXiv Project Code
CVPR 2025

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

A generative propagation framework that addresses multiple video tasks in a unified way by leveraging the generative power of diffusion models.

arXiv Project Video Data Firefly Adobe News
ICML 2026

Both Semantics and Reconstruction Matter: Making Representation Encoders Ready for T2I Generation and Editing

Shilong Zhang, He Zhang, Zhifei Zhang, Chongjian Ge, Shuchen Xue, Shaoteng Liu, Mengwei Ren, Soo Ye Kim, Yuqian Zhou, Qing Liu, et al.

PS-VAE introduces a semantic-pixel reconstruction objective that compresses both semantic information and fine-grained details into a compact representation.

arXiv Project
CVPR 2026

HBridge: H-Shape Bridging of Heterogeneous Experts for Unified Multimodal Understanding and Generation

Xiang Wang, Zhifei Zhang, He Zhang, Zhe Lin, Yuqian Zhou, Qing Liu, Shiwei Zhang, Yijun Li, Shaoteng Liu, Haitian Zheng, Jason Kuen, Yuehuan Wang, Changxin Gao, Nong Sang

An asymmetric H-shaped architecture that bridges heterogeneous experts through mid-layer semantic connections for unified multimodal understanding and generation at lower training cost.

arXiv
NeurIPS 2025

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

Jenga accelerates HunyuanVideo by 4.68× to 10.35× through dynamic attention carving and progressive resolution generation.

arXiv Project Code
TPAMI 2025

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

Mini-Gemini scales open-source vision-language models from 2B to 34B for high-resolution image understanding, strong OCR, and multimodal-reasoning-grounded image generation.

arXiv Project Code
Oral Presentation

NeurIPS Oral 2024

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

A slow agent decomposes the task and decides which actions to learn, while a fast agent writes code and RL configurations for low-level execution.

arXiv Project
Most Influential CVPR Papers (Paper Digest)

CVPR 2024

Video-P2P: Video Editing with Cross-attention Control

Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, Jiaya Jia

A video inversion process, cross-attention control, and decoupled guidance enable prompt-based video editing with strong fidelity and controllability.

arXiv Project Code
ICLR 2024

Direct Inversion: Boosting Diffusion-based Editing with 3 Lines of Code

Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, Qiang Xu

Rethinking the inversion process makes diffusion-based editing substantially stronger with a remarkably lightweight change.

arXiv Project Code