Homeglossary

Gaussian Splatting

Gaussian splatting is a 3D scene representation and rendering method that recreates a real object or space from photos or video using many soft, translucent visual points called Gaussians. It can produce realistic, interactive views without building a conventional triangle-based 3D model. In practice, it is especially useful for viewing captured environments, but it is not a complete replacement for editable geometry.

What Is Gaussian Splatting?

Gaussian splatting is a 3D scene representation method that recreates an object or space from photos or video using soft, translucent visual points called Gaussians. Instead of triangle surfaces, it uses colored, semi-transparent ellipsoids in 3D space to create realistic views from new camera positions.

A Gaussian is a smooth, cloud-like mathematical shape, while a splat is its projected footprint on a screen. Overlapping splats reproduce texture, edges, depth, reflections, and visual variation. Also called 3D Gaussian splatting, the technique is mainly for realistic viewing of captured scenes rather than editable engineering models.

Why Gaussian Splatting Became Important

Splatting methods have existed in computer graphics and volume visualization for decades. Interest accelerated after the 2023 research paper 3D Gaussian Splatting for Real-Time Radiance Field Rendering, which showed a practical way to reconstruct and render detailed scenes from multiple images at interactive speeds.

The key shift was combining photographic realism with responsive navigation. Earlier neural radiance field methods could create convincing new views but often took longer to train or render. Gaussian splatting made it more practical to explore reconstructed rooms, objects, and outdoor locations from freely moving virtual cameras.

How 3D Gaussian Splatting Works

The workflow starts with images, though good results depend on systematic capture and optimization. Software estimates camera positions, then adjusts soft 3D marks until rendered views resemble the original images.

  1. Capture sharp, overlapping photos or video frames from many positions around the subject.
  2. Estimate camera positions and orientations with structure from motion, which finds matching features across images.
  3. Create a sparse point cloud of estimated 3D locations from matched features.
  4. Turn points into initial Gaussians with location, size, angle, transparency, and color.
  5. Render trial images from known camera positions and compare them with source photos.
  6. Optimize Gaussian properties so trial images better match the captured scene.
  7. Add splats in detailed areas and remove weak ones, called densification and pruning.
  8. For a new viewpoint, project relevant splats onto the screen and blend them in depth order.

The Key Components of a Gaussian Splat

A splat is more than a colored dot. Position places it in the scene, while scale, rotation, and covariance define its stretched ellipsoid-like shape. One splat can cover a narrow edge, broad wall area, or textured feature.

Opacity controls how strongly a splat contributes when it overlaps others, and color records visible appearance. Many implementations use spherical harmonics, a compact way to vary color by viewing direction. This helps reproduce shiny surfaces appearing brighter from certain angles. Together, these settings make splats soft, angled patches of appearance rather than rigid points.

Gaussian Splatting vs. Photogrammetry, Point Clouds, Meshes, and NeRF

Gaussian splatting is not the same as photogrammetry, although both can begin with the same photographs and camera-solving process. The best choice depends on whether a project needs convincing viewing, editable surfaces, or reliable measurement.

MethodRepresentationBest useMain trade-off
Gaussian splattingOriented, semi-transparent 3D ellipsoidsPhotorealistic interactive explorationWeak for editing and precise geometry
Photogrammetry meshTriangle surfaces with texturesEditing, asset creation, inspectionCan need extensive cleanup and retopology
Point cloudIndividual measured 3D pointsSurveying, scanning, spatial analysisOften looks sparse without processing
NeRFNeural model of scene appearanceHigh-quality novel-view synthesisMay be slower or less convenient to render and deploy

Gaussian splatting is not automatically better than a neural radiance field. It often balances quality and real-time rendering, while a NeRF may suit workflows that prioritize a learned continuous scene model. Neither replaces a mesh when downstream tools need clean surfaces.

Where Gaussian Splatting Is Used

The technology is useful when people need to look around a captured place or object with a sense of presence.

  • Architecture and real estate walkthroughs of existing properties.
  • Cultural heritage documentation of fragile sites and objects.
  • Virtual production and visual effects reference for immersive location records.
  • Digital twins, remote inspection, and immersive training for facilities, machinery, and work areas.
  • Product visualization, tourism previews, and location-based storytelling.
  • Mapping and robotics research, where splats add visual context to geometric sensor data.

For CAD editing, manufacturing, safety-critical clearance checks, or formal surveying, teams usually need meshes, LiDAR, or other measured data in addition to, or instead of, a splat.

Benefits of Gaussian Splatting

Its main advantage is prioritizing what viewers see. A well-captured splat can communicate a place more effectively than a geometrically perfect model that looks artificial or incomplete.

  • It can preserve rich visual detail from ordinary camera captures.
  • It supports smooth, real-time navigation through reconstructed scenes.
  • It avoids much of the manual mesh cleanup needed for viewing-focused work.
  • It can capture a location faster than manually modeling visible surfaces.
  • It represents soft visual transitions naturally because splats blend instead of forming only hard polygon boundaries.

Practical Limits and Challenges

Gaussian splatting does not solve every 3D capture problem. Its output reflects what cameras saw, including blind spots and visual ambiguity.

  • Splats are difficult to edit as exact, watertight geometry.
  • Dimensions and measurements may be unreliable without calibrated survey or scan data.
  • Hidden areas create holes, especially behind furniture, under objects, and above ceilings.
  • Glass, mirrors, shiny metal, water, and transparent materials can confuse reconstruction.
  • Moving people, vehicles, shadows, or changing lights can leave artifacts or ghosting.
  • Detailed scenes may require substantial storage, graphics memory, and compatible viewers.
  • Capturing homes, workplaces, or public locations can expose faces, documents, screens, and other sensitive information.

How to Capture Better Gaussian Splats

Capture quality has more influence on results than most software settings. Treat recording like a careful visual survey of everything viewers may later want to see.

  • Use sharp, steady imagery with adequate light and minimal motion blur.
  • Maintain generous overlap between consecutive images or video frames.
  • Move slowly around the subject and avoid sudden direction changes.
  • Capture objects from different heights and distances when close-up detail matters.
  • Lock exposure and focus where possible so appearance remains consistent.
  • Keep people and moving objects out of the scene when practical.
  • Cover floors, ceilings, corners, doorways, and other easily missed angles.
  • Get permission before recording and remove sensitive information from the area.

Choosing Gaussian Splatting for a Project

Choose Gaussian splatting when the goal is realistic exploration and visual fidelity matters more than editable structure. Choose a mesh when a team must modify surfaces, manufacture an object, run collision tests, or depend on stable dimensions. Choose LiDAR, surveying, or calibrated scanning when measurement accuracy is the priority.

Gaussian splatting does not inherently require LiDAR and can be built from photos or video alone. However, LiDAR or other scan data can supplement a splat when a project needs convincing presentation and stronger spatial reference.

Is Gaussian Splatting AI?

Gaussian splatting is principally a computer graphics and scene-reconstruction technique, not generative AI. It uses optimization to tune scene parameters from image evidence and may appear in AI-assisted capture or editing workflows. Its core purpose is to represent and render a specific observed scene, not generate content from a text prompt or broad training dataset.

Formats, Viewing, and Long-Term Project Planning

Tools, viewers, and file formats vary across the Gaussian splatting ecosystem, so teams should test intended web, desktop, or immersive viewers early. Keep original photos or video, camera-calibration data, capture notes, and any source point cloud. These materials make it possible to retrain, repair, or export projects as software changes.

Blender and other 3D applications may support splat imports, plug-ins, or conversion workflows, but capability differs by version and pipeline. Before committing to a deliverable, confirm whether the recipient needs a live splat viewer, rendered video, conventional mesh, or a combination of formats.

Start Building
on Emergent today
Start Building