Introducing fVDB: Deep Learning Framework for Generative Physical AI with Spatial Intelligence

fVDB (Early Access) is a GPU-optimized deep learning framework for sparse, large-scale, high-performance spatial intelligence. It builds NVIDIA accelerated AI operators on top of NanoVDB to enable reality-scale digital twins, neural radiance fields, 3D generative AI, and more. fVDB is the infrastructure for generative physical AI with spatial intelligence. Apply for the fVDB Early Access Program: This video demonstrates various techniques fVDB infrastructure enables, including triangle mesh reconstruction from point clouds, large-scale neural radiance field training, high-resolution simulation upsampling, and even fully AI-generated city models. 0:00 - Digital Twins at Reality Scale 0:50 - Introducing fVDB 1:30 - Triangle Mesh from Point Clouds 1:44 - City-Scale NeRF 1:54 - Large-Scale 3D Generative AI 2:07 - Physics Super-Resolution 2:25 - Conclusion fVDB is essential for scaling applications in autonomous driving and 3D generative AI. The framework leverages NVIDIA core technologies including NanoVDB, CUTLASS, tensor cores, and CUDA. It’s implemented as a PyTorch extension for easy integration with other libraries and spatial intelligence algorithms. If you’re already using the VDB format, fVDB can read and write existing VDB datasets out of the box. It interoperates with other libraries and tools, such as Warp for Pythonic spatial computing, and the Kaolin Library for 3D deep learning. Adopting fVDB into your existing AI workflow is seamless. Apply for the fVDB Early Access Program: Dive into our announcement blog: Check out the fVDB technical blog: Read the research paper to learn more: #siggraph2024 #NVIDIA #fVDB #graphics #AI #digitaltwin #neuralnetworks #3D #generativeai #nvidiaresearch #siggraph