GaussianCompressionOptions
Options & Configuration · AiDotNet.Models.Options
Per-technique configuration for the GaussianSplatting post-training compression pass (#1835 excellence goal #3).
For Beginners
Compression takes the raw Gaussian cloud your training produced and makes it smaller / faster to render without hurting quality much:
- Prune — drop Gaussians that ended up nearly transparent.
- Merge — combine Gaussians that overlap almost entirely into one.
- Quantize — represent each color coefficient with 8 bits instead of 32.
All three are on by default; toggle individual techniques off if you don’t need them.
How It Works
Every field is defaulted to a paper-friendly setting so the user can enable
CompressOnBuildComplete and get a meaningful
compressed cloud without touching this class. Advanced users tune per-technique on/off +
thresholds.
Constructors
| Constructor | Summary |
|---|---|
GaussianCompressionOptions() |
Creates a new instance with paper-quality defaults. |
GaussianCompressionOptions(GaussianCompressionOptions other) |
Deep-copies every field from other. |
Properties
| Property | Summary |
|---|---|
bool MergeOverlapping { get; set; } |
Merge nearly-identical overlapping Gaussians into a single Gaussian. |
bool PruneLowOpacity { get; set; } |
Prune Gaussians whose opacity has collapsed below the configured threshold. |
bool QuantizeSphericalHarmonics { get; set; } |
Quantize per-Gaussian color / SH coefficients into a small integer range. |
double MergeOverlapThreshold { get; set; } |
Overlap fraction above which two Gaussians are candidates for merging. |
int QuantizationBits { get; set; } |
Bits-per-coefficient after quantization. |