GitHub
Browse our source code and thousands of tests, run dozens of step-by-step tutorials across 3D DL use cases; engage with the team on the issues board.
Overview
We build and maintain high‑performance, production‑ready SDKs that make spatial AI practical at scale. These tools power 3D deep learning, differentiable rendering, reality capture, and physics‑aware simulation across meshes, point clouds, Gaussian splats, and sparse volumes. Below we highlight two pillars used across research and production workflows: Kaolin for 3D deep learning in PyTorch and fVDB for sparse, large‑domain spatial intelligence and reality capture.
Kaolin
NVIDIA Kaolin is a PyTorch library that accelerates research and development in 3D deep learning and AI. It provides differentiable rendering, camera & lighting API, fast conversions between 3D representations, GPU‑optimized geometric operators, structured point cloud / octree acceleration (SPC), convenient mesh and PBR material tensor management. The library bundles cutting-edge techniques, like FlexiCubes for mesh optimization, and Simplicits for simulation, delivering speedups and new features. The library enables interactive UI prototypes over any closed-box renderer, and ships a representation-agnostic physics simulation API, supporting diverse interactive workflows for 3D AI across representations.
Points of Contact:
Clement Fuji Tsang,
Masha Shugrina
Kaolin Capabilities
Differentiable Rendering • Camera API • Differentiable Lighting • Mesh and PBR Material Tensor Management •
GPU Optimized Operations for Meshes, Points, Voxels • Conversions for 3D Representations • I/O and Datasets • Structured Point Clouds / Octrees • FlexiCubes and DMTet Differentiable Mesh Optimization • Gaussian Splat Voxelizer • Quaternions in PyTorch •
Physics Simulation with Collisions for Meshes, Points, Splats • Interactive 3D Visualization
Kaolin Featured Modules
Documentation
Install wheel across CUDA, PyTorch, OS versions. Browse detailed API references, and follow our SIGGRAPH lab session, which fuses multiple NVIDIA technologies.
Developer Page
Get a high-level overview of library features, and applications, from physics simulation to differentiable rendering and visualization.
Resources
fVDB
fVDB is a Python library of GPU‑accelerated data structures and algorithms for large‑domain spatial applications in PyTorch, built on NanoVDB. It provides sparse volumetric grids, Gaussian splat representations, jagged tensors, fast volumetric operations, and a web/Notebook visualizer. The Reality Capture toolbox adds high‑level pipelines for radiance‑field reconstruction and mesh extraction at scale.
Point of Contact:
Francis Williams
fVDB Links
Website
Overview, features, and documentation for fVDB.
Reality Capture
Tutorials and APIs for large‑scale radiance field reconstruction and meshes.