I am an M.Sc. student at the Technical University of Munich and a research intern at Agile Robots SE. I build generative and geometric models for interactive humans, dynamic 3D scenes, and robot learning.
My current work studies camera-controlled human motion video diffusion with Prof. Matthias Nießner and cross-embodiment video generation for dexterous manipulation.
I am seeking Ph.D. opportunities for Fall 2027 in 3D/4D scene representation, video generation, and robot world models.
Previously, I collaborated with Prof. Daniel Cremers, Prof. Nassir Navab, and Prof. Benjamin Busam on 3D reconstruction and tracking. I earned an M.Sc. in Computer Science from Tongji University and a B.Sc. in Information and Computing Science from Nanjing Tech University.
I joined Agile Robots SE as a research intern working on world models for robot dexterous manipulation.
Our paper CSG-Fusion received the Best Paper Award at the ICCV 2025 Workshop E2E3D (workshop track).
I graduated from Tongji University with a Master's degree in Computer Science.
Our paper LiteTracker is accepted by MICCAI 2025.
Our paper RE0 is accepted by ICRA 2025.
I joined ImFusion as a research intern on dense video point tracking.
I joined TUM-CVG and TUM-DI-LAB under the co-supervision of Prof. Yan Xia and Prof. Chuanxia Zheng.
I joined the Technical University of Munich as a Master's student in Electrical Engineering and Information Technology.
Built foundations in long-term tracking, 3D/4D reconstruction, and scene decomposition.
Working on camera-controlled human motion video generation and video world models for robot dexterous manipulation.
Unifying these threads into perception-action models of interactive humans and dynamic scenes.
* equal contribution, † corresponding author
Authors
Overview
3D face tracking-conditioned video diffusion for expressive, identity-preserving portrait animation from a single image, with autoregressive sampling for longer videos. On VFHQ, ViDS ranked first on 8 of 13 reported metrics.
Authors
Overview
Matching-based fusion of sparse-view pointmaps into compact, cross-view-consistent 3D Gaussians. At 90% ScanNet++ overlap, it improved PSNR by 2.8 dB over Splatt3R while using approximately 124K fewer Gaussians.
Authors
Overview
Causal temporal feature reuse with prior-motion initialization for accurate, low-latency online tissue tracking. It ran approximately 7× faster than its predecessor and 2× faster than prior state of the art, reaching 29.67 ms P95 for 1,024 points.
Authors
Overview
Training-free 3D zero-shot instance segmentation from multi-view masks and CLIP semantics.
Contributions
Contributions
Contributions
Contributions
Contributions
Contributions
Overview
Developed a 4D Half-Gaussian splatting pipeline for deformable stereo endoscopic reconstruction with depth-prior initialization, HexPlane spatiotemporal deformation, and edge-aware depth regularization. Achieved 38.1 PSNR on EndoNeRF versus prior endoscopic GS/NeRF baselines; also evaluated on SCARED.
Authors
Overview
A TUM DI Lab report on object-centric 3D reconstruction and decomposition with 3D Gaussian Splatting.
M.Sc. Electrical Engineering and Information Technology
M.Sc. Electrical Engineering and Information Technology
Double-degree program with Tongji University (Tongji M.Sc. awarded 2025).
Thesis: camera-controlled human motion video diffusion at the Visual Computing Group.
M.Sc. Computer Science
M.Sc. Computer Science
Thesis: endoscopic scene reconstruction with 4D half-Gaussian splatting.
B.Sc. Information and Computing Science (Embedded Software)
B.Sc. Information and Computing Science (Embedded Software)
A computing major within the Department of Mathematics.
Focus
AI Infrastructure Handbook
Overview
A static handbook for understanding how modern AI systems train, serve, generate, route, compress, and fail.
Focus
Reusable User Study Infrastructure
Overview
An open-source toolkit for building reusable user study questionnaires and experiment workflows.
Focus
2D Audio-driven Human Avatar Animation
Overview
A lightweight audio-driven 2D avatar solution that runs entirely in the browser using WASM based on Lite-avatar. No backend server required.
Thoughts on research, 3D, video generation, and the occasional in-between.
Receipt Deformation Simulation
Dive into an interactive water simulation where ducks float, collide, and create ripples.
01 / 01
