Xin Dong | 董鑫

I am a third-year Ph.D. candidate in the IVG@SZ Lab at Tsinghua University, supervised by Prof. Yansong Tang. Prior to that, I obtained a Master's degree in Computer Science and Technology from Ningxia University, where my research focused on face attribute analysis, under the supervision of Prof. Hao Liu. I also hold a Bachelor's degree from the School of Computer Science at Sichuan University.

Currently, I am passionate about high-fidelity visual generation, ranging from 2D human image generation and physical 3D reconstruction to 4D combustion synthesis in real-world settings. I hope my work can serve as a critical data backbone for 3D world models, thereby facilitating embodied policy training and enabling highly immersive experiences in video generation, VR and interactive gaming.

Email  /  Github  /  Scholar

* indicates equal contribution

This work addresses the problem of recovering complete, simulatable object geometry from reconstructed real-world scenes, enabling physics-based interaction with objects embedded in the scene. Specifically, we propose SAM3D-Phys that integrates scene reconstruction with generative 3D priors of SAM3D to recover physically simulatable objects.

We propose a 3D Pose Conditional Diffusion model (3DPCD) that leverages a human parametric model to integrate comprehensive and adjustable 3D control into forward–backward diffusion steps.

We propose a mask-free method for lossless multi-object 3D reconstruction from sparse and occluded real-world views, enabling physically plausible robotic interaction via Material Point Method (MPM) simulation. Our key insight is that object coupling stems from occlusion and limited viewpoints, which we address by formulating multi-object decoupling as a sparse-view reconstruction problem.

We propose an end-to-end open set face anti-spoofing (OSFA) approach for unseen attack recognition.

We propose the HFM approach, which integrates statistical and causal feature modulation for stable face anti-spoofing in unseen domains and unseen attacks.

Video moment retrieval and highlight detection are crucial tasks in video analysis that aim to localize specific moments and estimate clip-wise relevance based on a given text query. Existing approaches often neglect the rich visual information related to the text query within individual frames. To address this limitation, we propose a Comprehensive Spatial-Temporal Representation Learning Framework (CoSTL), which captures both fine-grained image-level information and temporal dynamics.

In this work, we focus on zero-shot 3D style transfer that can generate multi-view consistent stylized views of the 3D scene given an arbitrary style image. Our method combines feature Gaussian splatting and deferred stylization, enabling high-quality stylization with the data-sufficient decoder network while ensuring view consistency by unifying view-dependent operations into a view-invariant process.

We propose LAMP, a framework for pose-accurate and visually coherent multi-person image generation across various occlusion scenarios and interactions.

We propose a dynamic graph learning method for robust facial age estimation, which enforces causal regularization to discover an attentive feature space while preserving age label dependencies.

We propose a meta descent learning method (MDL) for class imbalanced age estimation while preserving the relative ordinal information.

Website Template, Thanks to Jon Barron