GitHub - Adversarr/Adversarr

Hi there 👋

🌱 My current research focuses on: AI for Computer Graphics, especially AI for accelerating geometric computing and numerical solvers.
📫 How to reach me: Contact me via email zherui_yang@mail.ustc.edu.cn. Zhihu
🚀 I'm working on gsx, an efficient, multiplatform 3DGS rendering/training library.

"Learning Sparse Approximate Inverse Preconditioners for Conjugate Gradient Solvers on GPUs" (NeurIPS 2025): code It explores the possibility that GNNs learn to improve & accelerates Sparse Approximate Inverse (SPAI) preconditioners, with practical wall-clock acceleration and GNN-SPAI architectural alignment.
"Learning Laplacian Eigenspace with Mass-Aware Neural Operators on Point Clouds" (SIGGRAPH 2026 Conference): code It use neural operator to learn the eigen-space of Laplacian operator (which defines "Fourier modes" on manifolds), achieving high efficiency while maintaining accuracy and robustness.
"Simple yet Effective: Low-Rank Spatial Attention for Neural Operators" (ICML 2026): code By analyzing SOTA neural operators, we propose a simple attention block that removes inductive bias and improves overall performance.

mathprim: tiny header-only tensor library, along with many optimization/solver implementations (Cholesky, ICPCG, L-BFGS, and more).
ssim: tiny & simple simulator, built on mathprim. Reach out my zhihu blog to see how fast it is!
tinygs: Blazing fast 3D Gaussian Splatting (3DGS) implementation, pure C++/CUDA, with async dataloading, tile-based rendering, and highly optimized CUDA rendering kernels.

Name		Name	Last commit message	Last commit date
README.md		README.md