Gaussian splats: now part of our stack

If you've explored modern 3D technology recently, you've probably come across the term Gaussian Splats or 3D Gaussian Splatting (3DGS). This relatively new technique has rapidly become one of the most exciting developments in 3D visualization, offering photorealistic results with significantly faster rendering than traditional methods.

Whether you're creating digital twins, virtual tours, product visualizations, or immersive web experiences, Gaussian Splats are changing how real-world environments are captured and displayed.

Yes, but what are they?

A Gaussian Splat is a tiny, semi-transparent 3D "blob" that represents a small portion of a real-world scene. Instead of building a model from millions of polygons, Gaussian Splatting reconstructs a scene using millions of overlapping ellipsoids.

Each splat contains information such as:

• Position in 3D space
• Size and shape
• Color
• Opacity
• Orientation

When rendered together, these splats blend seamlessly to create an incredibly realistic representation of the original environment.

Think of it as painting a 3D scene with millions of soft particles instead of constructing it with triangles.

The process begins with capturing hundreds—or sometimes thousands—of photographs or video frames from different viewpoints around an object or environment.

The typical workflow looks like this:

1. Capture Images

A camera, drone, or smartphone photographs the subject from multiple angles.

More overlap between images generally produces better results.

2. Camera Alignment

Photogrammetry software determines exactly where every photograph was taken and calculates the camera positions.

3. Neural Optimization

Instead of generating a polygon mesh, machine learning algorithms create millions of tiny Gaussian primitives.

Each Gaussian is optimized to match the appearance of the original photographs.

4. Real-Time Rendering

Specialized rendering techniques display these Gaussians efficiently, allowing users to navigate the scene smoothly in real time.

The result is a highly detailed digital replica with impressive lighting, reflections, and surface detail.

Traditional 3D models approximate reality using flat polygons and textures.

Gaussian Splats instead represent the actual light captured in photographs.

Because every splat contributes soft color and transparency, the renderer naturally recreates subtle visual details like:

• Fine textures
• Complex vegetation
• Hair and fur
• Thin objects
• Soft edges
• Lighting variation
• Reflections

Scenes often look remarkably close to real photography.

Incredible Visual Quality

One of the biggest strengths of Gaussian Splatting is its realism.

Fine details that are difficult to model traditionally—such as leaves, cables, grass, or rough stone—are reproduced naturally.

Faster Than Traditional Neural Rendering

Earlier neural rendering techniques like NeRFs produced beautiful results but often required slow rendering.

Gaussian Splats are designed for real-time viewing, making them practical for interactive applications.

Efficient Data Representation

Rather than storing massive polygon meshes with multiple texture maps, Gaussian Splats represent visual information compactly while maintaining excellent quality.

Excellent for Real-World Scanning

If the goal is to preserve an existing object or location exactly as it appears, Gaussian Splats often outperform traditional modeling workflows.

Examples include:

• Historic buildings
• Museums
• Construction sites
• Real estate
• Industrial facilities
• Cultural heritage sites

Smooth Web Experiences

Modern viewers can stream Gaussian Splats efficiently, making them suitable for websites without requiring users to download massive files.

Despite their impressive capabilities, Gaussian Splats are not the perfect solution for every project.

Difficult to Edit

Traditional 3D models allow artists to move vertices, modify geometry, or animate objects.

Gaussian Splats represent captured appearance rather than editable geometry.

Making significant changes after capture can be challenging.

Limited Animation

Animated characters and moving objects remain difficult.

Gaussian Splatting works best for static scenes.

Large Capture Requirements

High-quality results require many overlapping photographs taken under consistent lighting conditions.

Poor image capture leads to lower-quality reconstructions.

Not Ideal for CAD or Engineering

Engineering workflows require precise geometry, measurements, and clean surfaces.

Polygon meshes or CAD models remain the better choice for manufacturing, simulation, and design.

The technology is already being adopted across numerous industries.

Real Estate

Interactive property walkthroughs with photographic realism.

Construction

Progress documentation and digital site records.

Architecture

Visualizing completed buildings exactly as they exist.

Tourism

Virtual exploration of landmarks and attractions.

Museums

Digitally preserving artifacts and exhibition spaces.

Industrial Inspection

Capturing facilities for remote review and maintenance planning.

Film & VFX

Fast environment capture for virtual production workflows.

Gaming

Photorealistic environmental assets and immersive world building.

Education

Interactive learning experiences that allow students to explore real-world locations remotely.

Since its introduction in 2023, Gaussian Splatting has become one of the fastest-growing areas of computer graphics research. Improvements in compression, rendering speed, editing tools, and mobile support are arriving rapidly.

As hardware becomes more powerful and software continues to mature, Gaussian Splats are expected to become a standard method for capturing and sharing real-world environments on the web.