GaussianSplattingOptions

Options & Configuration · AiDotNet.Models.Options

Configuration options for Gaussian Splatting models (Kerbl et al. 2023, “3D Gaussian Splatting for Real-Time Radiance Field Rendering”).

Inheritance: ModelOptions → GaussianSplattingOptions

For Beginners

If you’re new to 3DGS: leave every field at its default and you’ll get paper-quality behavior. As you learn the model, override individual fields (e.g. bump MaxGaussians for a larger scene, lower PositionLearningRate for a more delicate initial cloud). Every field has a paper-anchored default so no configuration is required to start.

How It Works

Two families of settings live on this class: paper defaults for the model itself (opacity / scale / rotation initialization, per-attribute learning rates, spatial-index tuning) and densification-schedule knobs added by #1835 for finer-grained control over the split/prune loop that grows the initial random cloud into millions of specialized Gaussians during training.

Constructors

Constructor Summary
GaussianSplattingOptions() Creates a new instance with paper-quality defaults.
GaussianSplattingOptions(GaussianSplattingOptions other) Deep-copies every field from other.

Properties

Property Summary
GaussianCompressionOptions CompressionOptions { get; set; } Per-technique on/off + tuning for the post-training compression pass.
IDensificationSchedule DensificationSchedule { get; set; } Swappable densification-schedule strategy.
SplitChildLearningRateScales SplitChildLearningRateScales { get; set; } Per-attribute LR multipliers for split children.
bool CompressOnBuildComplete { get; set; } Run a post-training compression pass (prune + merge + SH quantize).
bool EnableDensification { get; set; } Whether the split/prune densification loop runs during training.
bool EnableSpatialIndex { get; set; } Whether to maintain a spatial index for accelerated per-pixel Gaussian queries.
bool UseSphericalHarmonics { get; set; } Whether to use spherical-harmonics color evaluation (view-dependent color).
double ColorLearningRate { get; set; } Base learning rate for Gaussian colors (or SH coefficients when SH is enabled).
double DefaultPointSpacing { get; set; } Default 3D spacing for cloud initialization when no point cloud is supplied.
double InitialNeighborSearchScale { get; set; } Initial scale factor for neighbor search during construction.
double InitialScaleMultiplier { get; set; } Initial scale multiplier for newly-created Gaussians.
double MinScale { get; set; } Minimum Gaussian scale — floor to prevent collapse during optimization.
double OpacityLearningRate { get; set; } Base learning rate for Gaussian opacities.
double PositionLearningRate { get; set; } Base learning rate for Gaussian positions.
double PruneOpacityThreshold { get; set; } Legacy opacity threshold below which a Gaussian is pruned.
double RotationLearningRate { get; set; } Base learning rate for Gaussian rotations (quaternion updates).
double ScaleLearningRate { get; set; } Base learning rate for Gaussian anisotropic scales.
double SplitGradientThreshold { get; set; } Threshold on the mean screen-space positional gradient (the norm of the loss gradient with respect to the projected mean in normalised device coordinates, averaged over the views a Gaussian was visible in) above which a Gaussian is densifie…
double SplitOpacityFactor { get; set; } Opacity factor applied to split children.
double SplitOpacityMax { get; set; } Maximum post-sigmoid opacity a split child can start at.
double SplitPositionJitter { get; set; } Position jitter magnitude applied to split children.
double SplitScaleFactor { get; set; } Scale factor applied to split children.
double? GradientNormThreshold { get; set; } Threshold on the mean screen-space (NDC) positional gradient norm above which a Gaussian is densified (cloned when small, split when large).
double? OpacityPruneThreshold { get; set; } Opacity below which a Gaussian is culled during pruning.
int DensificationInterval { get; set; } Iterations between fixed-interval densification firings.
int MaxGaussians { get; set; } Hard ceiling on total Gaussian count during training (OOM guard).
int ShDegree { get; set; } Spherical-harmonics degree (0-3).
int SpatialIndexRadius { get; set; } Neighborhood radius (in tiles) for spatial-index lookups.
int TileSize { get; set; } Splat rendering tile size (pixels per tile side).
int? DensificationEndIteration { get; set; } Iteration index after which densification stops.
int? DensificationStartIteration { get; set; } Iteration index at which the split/prune loop first activates.
int? GradientAccumulationWindow { get; set; } Iterations of gradient magnitude averaged before a split decision.
int? MaxGaussianCount { get; set; } Hard ceiling on total Gaussian count during densification.