Agents for science · Preprint, 2025
FMIRAgent
FMIRAgent connects visual perception, reasoning, and restoration strategies for microscopy.
Explore the codeRuian He何瑞安
Ruian He 何瑞安
Researcher & engineerShanghai, China
I am an Algorithm Engineer at Alibaba Cloud and a Postdoctoral Researcher at Shanghai Jiao Tong University. I am currently advised by Prof. Linfeng Zhang and Dr. Xiantao Zhang.
I received my Ph.D. in Computer Science from Fudan University in June 2026, advised by Prof. Bo Yan, and my bachelor's degree in Computer Science from Fudan in 2021.
My research interests span computer vision, multimodal language models, and agentic AI.
2026 — Present
Alibaba CloudAlgorithm Engineer · Since October
2026 — Present
Shanghai Jiao Tong UniversityPostdoctoral Researcher · Since October
Jun — Oct 2026
Alibaba CloudIntern
Mar — Jun 2025
Microsoft Research AsiaResearch Intern · Shanghai System and Engineering Group
Chain-of-thought reasoning for multimodal language models
2021 — 2026
Fudan UniversityPh.D. in Computer Science · September 2021 – June 2026
2016 — 2021
Fudan UniversityBachelor's degree in Computer Science
Selected research
Scientific imaging · Nature Methods, 2024
UniFMIR
UniFMIR uses pretraining to address a range of fluorescence microscopy image restoration tasks.
Read the paper
Efficient visual computing · AAAI, 2024
Space-time supersampling
Space-time supersampling brings frame extrapolation and supersampling into one low-latency framework.
Read the paperPublications & preprints
23 papers · * Equal contribution.
Published papers 15
MicroWorld: Empowering Multimodal Large Language Models to Bridge the Microscopic Domain Gap with Multimodal Attributed Property Graph
Structured knowledge and visual retrieval to support multimodal reasoning in microscopy.
Non-Uniform Class-Wise Coreset Selection for Vision Model Fine-tuning
Difficulty-aware selection allocates data budgets across classes for efficient model fine-tuning.
Efficient Trajectory Space-Time Super-Resolution for Fast Live-cell Imaging
We propose Trajectory Space-Time Super-Resolution (T-STSR), a method designed to boost frame rates and resolution in fast subcellular imaging while significantly reducing computational overhead.
A data-efficient strategy for building high-performing medical foundation models
We show that pretraining on approximately one million synthetic images allows a retinal foundation model trained on only 16.7% of 904,170 real images to match or exceed the performance of RETFound trained on the full dataset across nine tasks.
Efficient Online Training for Zero-Shot Time-Lapse Microscopy Denoising and Super-Resolution
We introduce MDSR-Zero, a zero-shot online learning method designed for plug-and-play noise suppression and super-resolution of microscopy videos.
FacialFlowNet: Advancing Facial Optical Flow Estimation with a Diverse Dataset and a Decomposed Model
A diverse facial optical-flow dataset and a decomposed model for facial motion estimation.
Pretraining a foundation model for generalizable fluorescence microscopy-based image restoration
We provide a universal fluorescence microscopy-based image restoration (UniFMIR) model to address different restoration problems, and show that UniFMIR offers higher image restoration precision, better generalization and increased versatility.
Low-latency Space-time Supersampling for Real-time Rendering
A unified approach to frame supersampling and extrapolation for low-latency real-time rendering.
SAMFlow: Eliminating Any Fragmentation in Optical Flow with Segment Anything Model
Through theoretical analysis, we find the pre-trained large vision models are helpful in optical flow estimation, and SAM is suitable for solving the fragmentation problem.
Context-Aware Iteration Policy Network for Efficient Optical Flow Estimation
A learned iteration policy adapts the computation budget to each optical-flow sample.
Uncertainty-Guided Spatial Pruning Architecture for Efficient Frame Interpolation
We develop an Uncertainty-Guided Spatial Pruning (UGSP) architecture to skip redundant computation for efficient frame interpolation dynamically.
MVFlow: Deep Optical Flow Estimation of Compressed Videos with Motion Vector Prior
We propose an optical flow model, MVFlow, which uses motion vectors to improve the speed and accuracy of optical flow estimation for compressed videos.
Motion Matters: Difference-based Multi-scale Learning for Infrared UAV Detection
Multi-frame differences and multi-scale features improve infrared UAV detection.
Fine-grained Blind Face Inpainting with 3D Face Component Disentanglement
We propose a novel fine-grained blind face inpainting framework and build up a new dataset called CelebO-3D.
Co-completion for occluded facial expression recognition
We propose a task-specific framework which first combines occlusion discarding and feature completion together to reduce the interference of occlusions on instance level.
No published papers match your search.
Preprints 8
Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents
Long-horizon browser agents designed for tasks in real-world web environments.
MicroVQA++: High-Quality Microscopy Reasoning Dataset with Weakly Supervised Graphs for Multimodal Large Language Model
A microscopy visual-question-answering corpus with graph-based quality control and human screening.
Self-Explained Thinking Agent for Autonomous Microscopy Restoration
An agent that reasons about image degradation, selects restoration strategies, and explains its decisions.
Unifying Segment Anything in Microscopy with Vision-Language Knowledge
Vision-language knowledge guides Segment Anything toward cross-domain microscopy segmentation.
Subcellular behavior model enables highly precise temporal super-resolved live-cell imaging
Behavior-aware particle matching improves interpolation for fast live-cell imaging.
A Generative Framework for Self-Supervised Facial Representation Learning
A 3D-aware generative framework disentangles facial identity and expression without paired supervision.
Instruct-NeuralTalker: Editing Audio-Driven Talking Radiance Fields with Instructions
Text instructions edit audio-driven neural talking faces while preserving lip synchronization.
Feature Pyramid Network for Multi-task Affective Analysis
Hierarchical representations connect valence-arousal estimation, expression recognition, and action-unit detection.
No preprints match your search.
Service & awards
Academic service
- Conference reviewing & program committees
- CVPR (2022, 2024, 2025), NeurIPS (2024), ICCV (2023), ECCV (2024), AAAI (2023–2025), IEEE VR (2023), ACM Multimedia (2024).
- Journal reviewing
- ACM TOG, IEEE TPAMI, and IEEE TCSVT (2024).
- Teaching at Fudan
- Computer Graphics A (Spring 2023–2025) and Compiler (Autumn 2022).
Awards & challenges
- 2026Shanghai Outstanding Graduate 上海市优秀毕业生
- 2024, 2025National Scholarship Top 1%
- 2024Jinrirencai Scholarship 10 students at Fudan University
- 20232nd place · 3rd Anti-UAV Challenge CVPR Workshops
- 20233rd prize · Fast-Motion VFI Challenge IACC
- 2021, 2023Outstanding Ph.D. Student Award
- 2018–2020Outstanding Undergraduate Student Award
- 2017Boxue Scholarship Top 10%