I am a researcher at ByteDance Seed, working on multimodal generation.
I obtained my Ph.D. from Fudan University, advised by Prof. Rui Feng. I was also a visiting student at MMLab, Nanyang Technological University, advised by Prof. Ziwei Liu.
From 2020 to 2024, I worked on visual perception and created the Recognize Anything Model (RAM) family, a series of open-source image perception models that exceed OpenAI's CLIP by more than 20 points in fine-grained perception. This work was done at OPPO Research Institute and IDEA, together with Youcai Zhang, Prof. Yandong Guo, and Prof. Lei Zhang.
From 2024 to 2025, I co-led the development (pre-training, SFT, and RL) of large multimodal models at TikTok AI Innovation Center.
Seedream 5.0 Pro is a multimodal image creation model that reasons before it draws. It plans logical layouts for text-dense infographics, grounds point, lasso, sketch, and color signals for pixel-level interactive editing, decomposes an image into 10+ editable transparent layers, and natively renders text in over ten languages.
High-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning
ACL 2026MGPO lets LMMs iteratively focus on key image regions through automatic grounding, achieving superior performance on high-resolution visual tasks without any grounding annotations.
Open-Set Image Tagging with Multi-Grained Text Supervision
CVPR 2024, Multimodal Foundation Models WorkshopRAM++ is the next generation of RAM, which recognizes any category with high accuracy, covering both predefined common categories and diverse open-set categories.
Recognize Anything: A Strong Image Tagging Model
CVPR 2024 Workshop on Multimodal Foundation Modelsproject page / arXiv / demo / code
RAM is an image tagging model that recognizes any common category with high accuracy.
Tag2Text: Guiding Vision-Language Model via Image Tagging
ICLR 2024project page / arXiv / demo / code
Tag2Text is a vision-language model guided by tagging, which supports tagging and comprehensive captioning simultaneously.
Multi-label learning with missing labels (MLML) is a challenging problem. We propose two simple yet effective methods via robust loss design.
The Recognize Anything Model (RAM) family, demonstrating superior image recognition ability across common and open-set categories.
The RAM family married with Grounded-SAM, which can automatically recognize, detect, and segment anything in an image.
Reviewer for CVPR, ICCV, ECCV, ICLR, NeurIPS, IJCV, WACV, and ACM MM.