FVDB: Large scale GPU reality capture from Nvidia (fvdb.ai)

NVIDIA released fVDB-Reality-Capture, an open-source reality-capture toolbox built on top of fVDB that accelerates large-scale 3D reconstruction workflows. It supplies high-level APIs for loading sensor data, reconstructing radiance fields (via 3D Gaussian Splatting), extracting meshes/point clouds, visualizing results, and exporting to formats like PLY and USDZ. The library is positioned as a domain-specific companion to fVDB—analogous to torchvision for PyTorch—designed for production use with a focus on robustness, minimal dependencies, multi-GPU scaling, and extensibility. Demonstrations include reconstructing 100 million 3D Gaussians from 400 high-res images and high-quality reconstructions from hundreds of views. For the AI/ML community this matters because fVDB-Reality-Capture makes high-fidelity radiance-field reconstruction tractable at much larger scales and lower memory cost. Underlying fVDB reportedly achieves ~50% better throughput and ~30% lower runtime versus gsplat in end-to-end training benchmarks while producing higher-quality outputs. Key technical features include out-of-core sensor loading (works with data larger than RAM), composable transforms with caching, optimized 3D Gaussian Splatting, mesh/point extraction, browser-capable renderings, and automatic multi-GPU scaling on a single machine. Released under Apache 2.0, it invites community contributions and integration into existing pipelines, lowering the barrier for scalable NeRF/radiance-field research and production workflows.