I am a PhD candidate at BUAA and Shanghai AI Laboratory, advised by Prof. Yu Qiao and Prof. Xiao Sun. My research focuses on spatial intelligence—from efficient 3D Gaussian Splatting to agentic 3D understanding—and autoregressive 3D generative models.
Previously I was a research assistant at ETH Zurich, where I built technology to efficiently prune tokens in ViT. I received my bachelor’s degree from the University of Science and Technology of China.
TPAMI, 2026
A masked rasterization supporting neural, dynamic 4D, and large-scale distributed Gaussian Splatting. A log-linear scaling law between the number of primitives and rendering quality, with 2x more efficient scaling than Grendel-GS.
ICML, 2026 Oral
We build an automatic pipeline that turns raw videos into accurate and semantically aligned 3DGS and constructs instance- and camera-level spatial QA data; finetuning on it can improve open-source VLMs across multiple public spatial benchmarks.
CVPR, 2026 Oral, Best Paper Award Candidate
We render a depth map from a proxy mesh to cull occluded Gaussians, eliminating long-overlooked occlusion redundancy in 3DGS. This hybrid rendering strategy leads to large speedups for large-scale rendering under heavy occlusions.
arXiv, 2025
A web-native WebGPU platform for real-time Gaussian splatting, mesh rendering, and per-frame ONNX inference.
CVPR, 2025
We propose a learnable mask for selecting 3D Gaussians during training by incorporating it into the strict rendering equation of 3DGS, and write a custom backward CUDA kernel to avoid vanishing gradients for zeroed Gaussians. This leads to 2x-3x faster rendering with almost no quality degradation.
WACV, 2024
We first adapt token pruning to dense prediction tasks, where directly discarding tokens hurts dense feature maps. By dynamically reactivating unselected tokens, we achieve high pruning rates with little performance loss.
