Technology

Gaussian Splatting: the future of 3D visualization

Gaussian Splatting represents 3D scenes as millions of Gaussian particles, achieving photorealism impossible with traditional geometry. Artsuka integrates it natively in Vision Pro, macOS and Windows.

What is Gaussian Splatting?

Gaussian Splatting represents a 3D scene as millions of Gaussians —small ellipses with position, color and opacity— rendered in real time via GPU rasterization. Unlike polygonal meshes, it requires no geometry simplification, preserving every visual detail.

Introduced in 2023 by INRIA, the technique was quickly adopted for its unique balance of photorealistic quality and real-time performance. Unlike NeRF, Gaussian Splatting is significantly faster in both training and rendering.

Artsuka supports .ply, .sog and .artsu, a proprietary format that packages splats with cameras, floor plans and HDR environments in a single bundle.

How Artsuka uses Gaussian Splatting

Artsuka Studio imports .ply and .sog directly, letting you configure cameras, points of interest and walkthroughs. The .artsu format adds camera metadata, 2D floor plans and HDR environments in a single portable file.

Rendering uses Metal on Vision Pro and macOS, and DirectX 12 on Windows, achieving 60+ FPS with millions of particles. The entire pipeline works offline.

  • 90+ FPS
    per eye on Vision Pro with Metal
  • 1M+
    Gaussians rendered in real time
  • ±2mm
    lidar capture precision
  • 100%
    offline, no Internet

How Gaussian Splatting works

  1. 1

    Capture

    Take photos or video of the space with a standard camera or phone. 100-200 photos for a room.

  2. 2

    Structure from Motion

    COLMAP analyzes the photos to estimate camera positions and generate an initial point cloud.

  3. 3

    Training

    Each point becomes a Gaussian. Position, covariance, color and opacity are optimized via gradient descent (SGD).

  4. 4

    Densification

    The system adds Gaussians where detail is missing and removes redundant ones, iteratively until convergence.

  5. 5

    Import

    Artsuka Studio imports the resulting .ply or .sog directly, no conversions needed.

  6. 6

    Visualization

    Native rendering on Vision Pro, macOS or Windows at 60+ FPS with GPU rasterization.

Technical aspects of Gaussian Splatting

Gaussian Splatting was presented at SIGGRAPH 2023 by Kerbl, Leimkühler, Drettakis and Paszke in the paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering". The technique represents a scene as millions of 3D Gaussians, each defined by position (XYZ), 3×3 covariance (shape and orientation), spherical harmonic color (view-dependent) and opacity (α).

The pipeline begins with Structure from Motion (SfM) using COLMAP, which estimates camera positions and generates a sparse point cloud. Each point is initialized as a Gaussian with isotropic covariance derived from neighboring point density.

Training uses Stochastic Gradient Descent (SGD) with differentiable Gaussian rasterization: the image is rendered, loss is computed against the real photo, and gradients are propagated to adjust each Gaussian's parameters. Adaptive densification adds Gaussians where loss is high and removes transparent ones.

For rendering, each 3D Gaussian is projected to 2D from the camera perspective, sorted by depth (back-to-front) and composited via alpha blending. This rasterization is analogous to triangle rasterization but with Gaussians, enabling real-time rendering without neural networks at inference.

Unlike NeRF, which requires evaluating an MLP per pixel, Gaussian Splatting stores parameters explicitly and rasterizes them directly on GPU. This enables 60+ FPS with millions of Gaussians, compared to 1-5 FPS typical of NeRF.

Gaussian Splatting vs NeRF vs Photogrammetry

FeatureGaussian SplattingNeRFPhotogrammetry
Visual qualityPhotorealisticPhotorealisticGood, loses fine details
Training speed10-30 minHours to days30-60 min
Rendering speed60+ FPS1-5 FPS60+ FPS (mesh)
Representation3D GaussiansNeural network (MLP)Polygonal mesh + texture
File size50-500 MB50-500 MB10-200 MB
Reflections & transparencyYes, with SHYes, nativeLimited
Sky & open spacesYesYesNo, requires closed geometry
Post-processing editingYes, remove/move GaussiansNo, retrainYes, edit mesh
Rendering hardwareStandard GPUPowerful GPUAny device

Software and tools for Gaussian Splatting

  • COLMAP

    Structure from Motion (SfM) pipeline for generating initial point clouds from photos. Open source, developed by ETH Zürich.

  • INRIA Gaussian Splatting

    Reference implementation of the original paper. Generates .ply with trained Gaussians. Requires NVIDIA GPU with CUDA.

  • Polycam

    Mobile app to capture Gaussian Splats directly from your phone. Exports in .ply format compatible with Artsuka.

  • Luma AI

    Cloud platform to capture and process Gaussian Splats from video. Exports .ply and .sog.

  • Artsuka Studio

    Professional software to import, configure and visualize Gaussian Splats on macOS, Windows and Vision Pro. Supports .ply, .sog and .artsu.

  • Postshot

    Desktop software to train Gaussian Splats locally from photos. Exports compatible .ply.

Gaussian Splatting glossary

Gaussian
3D particle with position, covariance (shape), color and opacity. The basic unit of representation in Gaussian Splatting.
Covariance
3×3 matrix defining the shape and orientation of a Gaussian. Controls how it stretches and rotates in 3D space.
SfM (Structure from Motion)
Technique for estimating camera positions and 3D geometry from multiple 2D images. COLMAP is the reference tool.
Differentiable rasterization
Process of rendering Gaussians to 2D while allowing gradient propagation back to 3D parameters. Essential for training.
Densification
Iterative process that adds Gaussians where detail is missing and removes redundant ones during training.
Alpha blending
Composition of back-to-front sorted Gaussians using opacity. Each Gaussian contributes to the final color based on its α.
Spherical harmonics (SH)
Mathematical functions that allow a Gaussian's color to depend on the viewing angle, enabling reflections and view-dependent effects.
NeRF (Neural Radiance Field)
Prior technique that represents 3D scenes as a neural network. Slower in rendering but similar in quality.
Radiance field
Function describing light radiation at every point and direction in space. Both NeRF and Gaussian Splatting are radiance fields.

Benefits

  • Absolute photorealism

    Every detail of light, texture and reflection is preserved without quality loss.

  • Real-time rendering

    Millions of Gaussians at 60 FPS with Metal and DirectX 12.

  • Standard formats

    Native support for .ply, .sog and .artsu. Import captures without conversions.

  • Offline first

    All processing and rendering is local. No Internet required.

Use cases

  • Architecture

    Present projects captured with photogrammetry at real scale in Vision Pro.

  • Real Estate

    Photorealistic virtual tours that replace physical visits.

  • Cultural heritage

    High-fidelity digitization of monuments and historic spaces.

  • Construction

    Visual documentation of site progress with splat captures.

FAQ

.ply, .sog and .artsu, which packages splats with cameras, floor plans and HDR environments.

Gaussian Splatting produces more photorealistic results by not simplifying geometry. Rendering is more efficient using GPU rasterization. Photogrammetry generates editable meshes but loses detail in fine elements like vegetation or reflections.

Gaussian Splatting is faster in training (minutes vs hours) and rendering (60 FPS vs 1-5 FPS), with comparable or superior quality. NeRF uses a neural network to render, while GS uses direct Gaussian rasterization.

No. You can capture splats with a standard camera or phone using Polycam, Luma AI or Artsuka's pipeline. To train from scratch you need an NVIDIA GPU with CUDA.

Yes. Artsuka Vision renders Gaussian Splats natively in visionOS with Metal, at 90+ FPS per eye.

Completely. All rendering is local. Import your splats and visualize them offline.

Take 100-200 photos of the space from multiple angles. Process with COLMAP + INRIA pipeline, or use apps like Polycam or Luma AI that automate the process.

"3D Gaussian Splatting for Real-Time Radiance Field Rendering" by Kerbl et al., presented at SIGGRAPH 2023 by INRIA researchers. It is the basis for all current implementations.

Related topics

Want to see Gaussian Splatting in action?

Request a demo and discover the photorealism of Gaussian Splatting on Vision Pro.

Request demo