Hi there!
Welcome to the website of Hanyang Kong (ๅญๆๆธ). ๐
I am honored to be pursuing my Ph.D. at the xML Lab @ NUS, advised by Prof. Xinchao Wang since August 2021. My academic journey in computer science began at Xiโan Jiaotong University, where I earned my masterโs degree under the mentorship of Prof. Qingyu Yang. ๐
My research is centered on AIGC (Artificial Intelligence-Generated Content) and LLMs (Large Language Models), with a growing emphasis on 3D/4D Vision-Language-Action (VLA) modelsโempowering agents to understand, reason, and act in both static and dynamic real-world environments with spatial and temporal depth.
Key research interests include:
- ๐ 3D Generation and Estimation: Advancing digital modeling for realistic scene understanding.
- ๐ก Diffusion Models and LLM Applications: Exploring generative AI for both creative and practical purposes.
- ๐ 3D/4D Vision-Language-Action (VLA) Models: Developing embodied AI systems capable of perceiving, interpreting, and interacting with complex physical real-world environments.
Thank you for visiting!
Iโm open to employment opportunities for Research Scientist/Engineer roles starting in 2026, preferably in Singapore, Beijing, or Shanghai.
๐ฅ News
- 2025.06: ย ๐๐ Our paper RogSplat: Robust Gaussian Splatting via Generative Priors was accepted by ICCVโ25.
- 2025.02: ย ๐๐ Our paper Generative Sparse-View Gaussian Splatting was accepted by CVPRโ25.
- 2024.10: ย ๐๐ Our paper EDGS: Efficient Gaussian Splatting for Monocular Dynamic Scene Rendering via Sparse Time-Variant Attribute Modeling was accepted by AAAIโ25.
- 2024.07: ย ๐๐ Our paper DreamDrone: Text-to-Image Diffusion Models are Zero-shot Perpetual View Generators was accepted by ECCVโ24.
- 2023.07: ย ๐๐ Our paper Priority-centric human motion generation in discrete latent space was accepted by ICCVโ23.
๐ Publications
RogSplat: Robust Gaussian Splatting via Generative Priors
Hanyang Kong, Xingyi Yang, Xinchao Wang
- Make 3DGS Work in the Wild: RogSplat fixes real-world issues like occlusion and motion blur.
- Smart Outlier Cleanup: Detects and inpaints corrupted regions with a generative refiner.
- Reliable in Real Scenes: Outperforms prior methods on complex real-world datasets.
Generative Sparse-View Gaussian Splatting
Hanyang Kong, Xingyi Yang, Xinchao Wang
- Turn Sparse Views into Rich 3D: Boosts 3D/4D Gaussian splatting with diffusion-based novel views.
- Geometry-Aware Consistency: Enforces structural alignment via semantic correspondences.
- More with Less: Matches dense data models using only a few input images.
Hanyang Kong, Xingyi Yang, Xinchao Wang
- Voxelized Time-Variant 3DGS: Introduces deformable Gaussian splatting with unsupervised attribute filtering.
- Kernel-Based Motion Flow: Formulates scene deformation using sparse, interpretable flow.
- Faster, Better Rendering: Achieves faster rendering with superior visual quality.
DreamDrone: Text-to-Image Diffusion Models are Zero-shot Perpetual View Generators
Hanyang Kong, Dongze Lian, Michael Bi Mi, Xinchao Wang
- Zero-shot, Training-Free Scene Creation: Generates perceptual scenes directly from text, without specific training for each scene.
- Click-Guided Dreamscapes Navigation: Allows drone flight control through point selection, offering a visually immersive experience.
- Resource-Efficient Scene Generation: Omits the need for a 3D point cloud, enabling faster scene creation with lower computational demand.
Priority-Centric Human Motion Generation in Discrete Latent Space
Hanyang Kong, Kehong Gong, Dongze Lian, Michael Bi Mi, Xinchao Wang
- Text-to-motion generation in descrete latent space.
- Priority-centric diffusion scheme for the discrete diffusion model.
๐ Educations
- 2021.08 - present: Ph.D. candidate in College of Design and Engineering, National University of Singapore.
- 2017.08 - 2020.06: M.Eng. in Faculty of Electronic and Information Engineering, Xiโan Jiaotong University.
- 2013.08 - 2017.06: B.Eng. in Faculty of Electrical Engineering and Automation, Hefei University of Technology.