NVIDIA's 3DGRUT v1.1.0 now exports a trained scene to five formats, USD, USDZ, PLY, NuRec, and lightfield from AOUSD, with a mesh injection tool and roundtrip tests holding the conversions together.
NuRec is the generally available Omniverse library that ingests real camera, lidar, and drive log data and renders it back as interactive 3D Gaussian scenes inside OpenUSD for Isaac Sim, AlpaSim, and CARLA. 3DGRUT is where those scenes get trained and emitted, and v1.1.0 turns the earlier single OpenUSD ParticleField path into a multi format exporter that also reaches PLY for the wider splat ecosystem and NuRec and lightfield containers for the simulation side.
v1.1.0 adds PPISP, Physically-Plausible Image Signal Processing, a learned post-processing stage that models the photometric variation real multi camera rigs introduce and the release pairs it with a USD export path so the corrected result travels with the asset.
Feeding those captures in got a dedicated path too. v1.1.0 adds an NCore-v4 dataset loader, NCore being NVIDIA's sensor-data format, alongside accelerated simplejpeg decoding and GPU cached camera rays, plus multi-camera support and image mask supervision. These target rig and drive log data specifically. Many cameras, large frame counts, and masked regions, rather than the single camera object captures the original 3DGS pipeline assumed.
The new version also brings an MCMC densification and pruning strategy and a Sparse Adam optimizer, both drawn from the 3DGS MCMC that reaches target quality with fewer, better-placed Gaussians and trims per step cost, and it adds a fine grained warp per pixel rendering path with load balancing. These land across both the 3DGRT ray tracer and the 3DGUT rasterizer that handles fisheye, 360, and rolling shutter cameras.
A new Playground 2.0 path tracer now shades inserted meshes with a Cook Torrance physically based model, lights them from HDR environment maps, supports BRDF and BTDF for transmissive surfaces with a configurable bounce cap, and ships with a Viser web GUI and a headless mode for batch runs.
Builds against CUDA 12 and 13 and Blackwell GPUs, runs on DGX Spark and other ARM64 systems, and adds Windows and uv-based installs, on top of a move to Git LFS and the project's first formatted, CI-gated codebase.
Learn more about the updates here. 3DGRUT continues to be open source and permissibly licensed.
