fVDB v0.6.0 Moves Gaussian Splatting API to Reality Capture

NVIDIA has released fVDB v0.6.0, which removes the high level Gaussian splatting Python API from fvdb-core and hands it to fVDB Reality Capture.

fVDB v0.5.0 rebuilt that pipeline in pure Python around one user facing object, fvdb.GaussianSplat3d. Both fvdb/gaussian_splatting.py and fvdb/_gaussian_autograd.py are present at tag v0.5.1 and absent at v0.6.0, so code importing them from fvdb breaks on upgrade.

Callers move to fvdb_reality_capture.GaussianSplat3d, ProjectedGaussianSplats, gaussian_render_jagged and evaluate_spherical_harmonics, and the ShOrderingMode, RollingShutterType, CameraModel and ProjectionMethod enums now live there too. fvdb.viz keeps its splat rendering, with Scene.add_gaussian_splat_3d accepting a core owned GaussianSplatViewData tensor contract and the older six property model interface held temporarily behind a deprecation warning.

The compiled CUDA kernels stay in fvdb-core, and v0.6.0 extends them. Multi-GPU Unscented Transform projection and world space rasterization arrive for 3DGUT style training on distorted camera models, building on the single GPU unscented transform added in the 0.4 release ahead of GTC. Ellipse-tile intersection tests replace bounding box tests, which NVIDIA reports cuts intersection records by 30 to 40 percent. Gaussian intersection offsets and indices widen to 64 bits, which NVIDIA describes as preventing overflow and corrupt training results on large scenes.

NVIDIA measured spherical harmonics backward 23.5 to 57.5 percent faster for 1M to 100M Gaussians and one to four cameras, and projection backward 1.43 to 2.18x faster, on dual RTX 3090s. A channel ordering fix in the fused SSIM prefetch measured 1.76 to 4.67x across 1080p to 8K benchmarks on the same dual RTX 3090s, and a multi-GPU backward rasterization race, where one device could clear gradients accumulated by another, is fixed. tinyply 3.0 loads a 375 million Gaussian PLY in four minutes instead of twelve in NVIDIA's reported test.

Away from splatting, v0.6.0 adds fvdb.nn.Prune for pruning grid topology and aligned features against a per voxel boolean mask, native level set and fog volume rendering through Scene.add_level_set and Scene.add_fog_volume, and slider, number, text and checkbox viewer widgets. retopologize_sdf is renamed rebuild_narrow_band with no compatibility alias.

fVDB v0.6.0 builds against PyTorch 2.13 with CUDA 13.0 and 13.2 and Python 3.10 through 3.15, and its package metadata now requires torch 2.7 or newer. CUDA 12.8 has left the wheel matrix. It is available now.