storytold/spark
2
0
Fork 0
mirror of https://github.com/storytold/spark.git synced 2026-10-09 00:09:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Expand internal splat scale encoding to e^[-12,9]. Lower zero cut-off to e^-30. Fix background color selection in editor. (#110)

Browse Source
This commit is contained in:
Andreas Sundquist
2025-07-11 10:48:10 -07:00
committed by GitHub
parent c8394665ec
commit dc07e02a33
10 changed files with 33 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/docs/packed-splats.md
+2 -2
View File
@@ -1,6 +1,6 @@
# PackedSplats
A `PackedSplats` is a collection of Gaussian splats, packed into a format that takes exactly 16 bytes per splat to maximize memory and cache efficiency. The `center` xyz coordinates are encoded as float16 (3 x 2 bytes), `scale` xyz as 3 x uint8 that encode a log scale from e^-9 to e^9, `rgba` as 4 x uint8, and quaternion encoded via axis+angle using 2 x uint8 for octahedral encoding of the axis direction and a uint8 to encode rotation amount from 0..Pi.
A `PackedSplats` is a collection of Gaussian splats, packed into a format that takes exactly 16 bytes per splat to maximize memory and cache efficiency. The `center` xyz coordinates are encoded as float16 (3 x 2 bytes), `scale` xyz as 3 x uint8 that encode a log scale from e^-12 to e^9, `rgba` as 4 x uint8, and quaternion encoded via axis+angle using 2 x uint8 for octahedral encoding of the axis direction and a uint8 to encode rotation amount from 0..Pi.
## Creating a `PackedSplats`
@@ -115,7 +115,7 @@ The center x/y/z components are encoded as float16, which provides 10 bits of ma
### Splat scales encoding
The XYZ scales are encoded independently using the following mapping: Any scale values below e^-20 are interpreted as "true zero" scale, and encoded as `uint8(0)`. Any other values quantized by computing `ln(scale_xyz)`, mapping the range e^-9..e^9 to uint8 values 1..255, rounding, and clamping. This logarithmic scale range can encode values from 0.0001 up to 8K in scale, with approximately 7% steps between discrete sizes, and has minimal impact on perceptible visual quality.
The XYZ scales are encoded independently using the following mapping: Any scale values below e^-30 are interpreted as "true zero" scale, and encoded as `uint8(0)`. Any other values quantized by computing `ln(scale_xyz)`, mapping the range e^-12..e^9 to uint8 values 1..255, rounding, and clamping. This logarithmic scale range can encode values from 0.0001 up to 8K in scale, with approximately 7% steps between discrete sizes, and has minimal impact on perceptible visual quality.
### Splat orientation encoding
docs/docs/spark-renderer.md
+1 -1
View File
@@ -56,7 +56,7 @@ const spark = new SparkRenderer({
| **preUpdate** | Controls whether to update the splats before or after rendering. For WebXR this *must* be false in order to complete rendering as soon as possible. (default: `false`)
| **originDistance** | Distance threshold for `SparkRenderer` movement triggering a splat update at the new origin. (default: `1.0`) This can be useful when your `SparkRenderer` is a child of your camera and you want to retain high precision coordinates near the camera.
| **maxStdDev** | Maximum standard deviations from the center to render Gaussians. Values `Math.sqrt(5)`..`Math.sqrt(9)` produce good results and can be tweaked for performance. (default: `Math.sqrt(8)`)
| **enable2DGS** | Enable 2D Gaussian splatting rendering ability. When this mode is enabled, any `scale` x/y/z component that is exactly `0` (minimum quantized value) results in the other two non-zero axes being interpreted as an oriented 2D Gaussian Splat instead of the usual approximate projected 3DGS Z-slice. When reading PLY files, scale values less than e^-20 will be interpreted as `0`. (default: `false`)
| **enable2DGS** | Enable 2D Gaussian splatting rendering ability. When this mode is enabled, any `scale` x/y/z component that is exactly `0` (minimum quantized value) results in the other two non-zero axes being interpreted as an oriented 2D Gaussian Splat instead of the usual approximate projected 3DGS Z-slice. When reading PLY files, scale values less than e^-30 will be interpreted as `0`. (default: `false`)
| **preBlurAmount** | Scalar value to add to 2D splat covariance diagonal, effectively blurring + enlarging splats. In scenes trained without the splat anti-aliasing tweak this value was typically 0.3, but with anti-aliasing it is 0.0 (default: `0.0`)
| **blurAmount** | Scalar value to add to 2D splat covariance diagonal, with opacity adjustment to correctly account for "blurring" when anti-aliasing. Typically 0.3 (equivalent to approx 0.5 pixel radius) in scenes trained with anti-aliasing.
| **focalDistance** | Depth-of-field distance to focal plane (default: `0.0`)