What is Gaussian splatting?
Learn how Gaussian splatting turns photos or video into realistic 3D scenes, when it works well, and where it differs from meshes, point clouds, and Neural Radiance Fields.
Short version
Gaussian splatting is a way to represent a reality capture scene with hundreds of thousands to millions of soft, semi-transparent 3D ellipsoids sometimes also called splats. Each splat stores a position, shape, colour, and opacity. Each splat is just a building block (a rendering primitive) and together they form a coherent appearance of the original space.
A Gaussian splat stores an optimised visual field that's both photorealistic and contains the original geometric structure.
What a splat contains
A splat is a 3D Gaussian, which means its visual contribution falls off smoothly from its centre instead of stopping at a hard edge. A splat can be pictured as a tiny grain of rice without a hard edge.
Each splat usually contains these values:
- Position: the centre of the splat in 3D space
- Scale: how large the splat is along each axis
- Rotation: which direction the ellipsoid points
- Opacity: how much the splat blocks or blends with splats behind it
- Colour: the colour the splat contributes to the final image
- Spherical harmonics: optional parameters that contribute towards the view-dependent colour
How Gaussian splatting is created
One of the main advantages of 3DGS models is that they can be created from simple images or videos of a scene. The training process optimises the splats until rendered views closely match the original capture.
The common capture and training flow is:
- Estimate each camera position with Structure from Motion (SfM), a technique that recovers camera poses from matching image features
- Use the sparse point cloud from SfM as the starting positions for splats
- Render the scene from the known camera positions
- Compare each render with the original image from that position
- Adjust splat position, shape, opacity, and colour to reduce the difference, and repeat until all splats fall into the right spots
- During training, add, split, or remove splats as the optimisation process discovers where the scene needs more or less detail
This optimisation process is only possible because splats are continuous (therefore differentiable) in space. In comparison with point clouds or meshes, splats have no edge that would form a discontinuity in space. Therefore standard machine-learning algorithms like SGD or similar can be used to find the optimal parameters for all splats.
Advantages of Gaussian splatting models
3DGS vs mesh
The main advantage of 3DGS in comparison with meshes is that 3DGS can be reconstructed from images and videos. Since mesh models are discontinuous functions in 3D space, meaning meshes have edges where 3D models end, they cannot be easily end-to-end optimised from visual data.
3DGS vs point clouds
3DGS models are essentially point clouds with extra parameters. Instead of representing the space with discrete points, each point gains shape, size, opacity, and a view-dependent colour. This way we can represent large spaces with higher visual accuracy and fewer primitives. We're, however, giving up some geometric accuracy.
3DGS vs Neural Radiance Fields
Neural Radiance Fields (NeRFs) are a family of algorithms and methods for photorealistic reconstruction of 3D models, similar to Gaussian splatting. In comparison to explicit splats, NeRFs represent the scene using a neural network. Because of this representation, NeRFs require millions of neural network inferences per render, making them too computationally expensive for most hardware.
Where Gaussian splatting works well
Gaussian splatting works best when visual realism matters more than editable geometry or survey-grade measurement. It is strongest for captured places and objects that viewers need to explore from multiple angles.
Common uses include:
- Real estate tours and property walkthroughs
- Museums, cultural heritage sites, and exhibits
- Film, VFX, and virtual production references
- Construction, infrastructure, and facilities documentation
- Product scans and object showcases
- Drone, mobile, and handheld reality capture
- Visual inspection workflows where context matters more than exact measurements
Use Gaussian splats when you want a scene to look and feel like the original capture.
Where Gaussian splatting does not work well
Gaussian splatting is not a replacement for every 3D format. It trades editability and measurement confidence for visual fidelity.
Use cases where Gaussian splatting models are suboptimal:
- CAD-grade dimensions or legal survey measurements
- Clean topology for modelling, rigging, or animation
- Physics, collision, or gameplay interaction
- Material editing, relighting, or style changes
- Correct unseen backsides or hidden areas
- Long dynamic motion without a dynamic splatting method
- Precise transparent or mirror-like surfaces
- High-performance solutions (like games) that depend on manual optimisation of assets
For those workflows, combine splats with meshes, point clouds, depth data, or manual reconstruction.
Gaussian splatting in Blurry
Blurry helps you store, process, share, and present Gaussian splatting scenes in the browser. Upload a splat, prepare the scene in the editor, add annotations or waypoints, and publish a link when it is ready.
Use Blurry for visual communication around captured spaces: walkthroughs, project reviews, inspections, showcases, and guided presentations.
Key takeaways
Gaussian splatting represents captured scenes as optimised translucent ellipsoids. It gives you realistic real-time viewing, but it does not replace every 3D workflow.
- Gaussian splats are visual primitives, not triangle meshes
- Training adjusts splat position, shape, opacity, colour, and density
- Real-time rendering comes from projecting and blending splats directly
- Splats are strong for visual exploration and sharing
- Point clouds and meshes still matter for measurement, editing, animation, and physics
- Capture quality controls the ceiling for final splat quality