What Is Gaussian Splatting? A Plain-English Guide to 3DGS
Guide · 2026-06-09 · 9 min read · by SplatMart Team
Gaussian splatting is a way to turn photos or video of a real scene into a photorealistic 3D model you can fly through in real time. Here's how it works, what it's used for, and how it compares to NeRF and photogrammetry.
Gaussian splatting is a technique for turning a set of photos or a short video of a real scene into a photorealistic 3D model that you can view and fly through in real time. Instead of building the scene out of triangles like a traditional 3D model, it represents the scene as millions of tiny, soft, coloured blobs — called 3D Gaussians — that blend together to recreate exactly what the camera saw.
The full name is 3D Gaussian Splatting, often shortened to 3DGS. It was introduced in a 2023 research paper ("3D Gaussian Splatting for Real-Time Radiance Field Rendering") and has quickly become one of the most popular ways to capture real places and objects in 3D, because it looks incredibly realistic and runs fast enough for games, VR, and the web.
What is a Gaussian splat, exactly?
A "Gaussian" is just a soft, fuzzy blob — think of a small cloud of colour that is bright in the middle and fades out at the edges. A Gaussian splat scene is made of millions of these blobs, and each one stores a few pieces of information:
- Position — where the blob sits in 3D space.
- Scale and rotation — how big it is and which way it is stretched (they are usually little ellipsoids, not perfect spheres).
- Colour — what colour it is, including how that colour changes depending on the angle you view it from (which is what gives splats realistic, view-dependent reflections).
- Opacity — how see-through it is.
On their own, the blobs look like coloured dust. But when millions of them are layered and blended together — "splatted" onto your screen — they reconstruct the original scene with photographic detail.
How does Gaussian splatting work?
The process happens in three broad stages:
1. Capture
You take a set of overlapping photos or a slow video that moves around the subject, covering it from many angles. The more complete and steady the coverage, the better the result.
2. Camera solving (Structure from Motion)
Software figures out where each photo was taken from by matching features between images. This step, called Structure from Motion (often run with a tool like COLMAP), produces a rough point cloud and the camera positions.
3. Training (optimisation)
Starting from that point cloud, the system places Gaussians and then repeatedly compares its rendered output to the original photos, nudging each blob's position, size, colour, and opacity until the rendered scene matches the photos. After a few minutes to an hour on a GPU, you have a finished splat — typically saved as a .ply or .splat file.
Why is Gaussian splatting such a big deal?
- Photorealism — because it is optimised directly against real photos, it captures soft reflections, transparency, and fine detail that traditional photogrammetry meshes struggle with.
- Real-time speed — a trained splat renders in milliseconds on a modern GPU, so it works in games, VR/AR, and even directly in a web browser.
- Accessibility — you can make a usable splat from an ordinary phone video using free or low-cost apps; you don't need a lidar scanner or a camera rig.
What is Gaussian splatting used for?
- Games and virtual production — realistic environments and backdrops.
- Real estate and architecture — walkable 3D captures of properties and buildings.
- VR and AR — immersive, photoreal spaces.
- E-commerce and product shots — turn an object into an interactive 3D model.
- Film, VFX, and digital twins — capturing real locations for reuse.
How is it different from NeRF and photogrammetry?
Photogrammetry builds a traditional mesh (triangles plus a texture). It is widely supported but tends to look flat on reflective or fuzzy surfaces. NeRF (Neural Radiance Fields) uses a neural network to achieve gorgeous results, but it is typically slow to render. Gaussian splatting sits in the sweet spot: NeRF-level realism with real-time rendering and no neural network at view time. We cover this in depth in our guide to NeRF vs Gaussian splatting.
How do I make — or get — a Gaussian splat?
To make your own, capture a scene and process it with an app like Luma AI, Polycam, or Postshot, or train it yourself with open-source tools. If you would rather skip the capture and processing, you can buy ready-made, professionally captured splats. SplatMart is a marketplace for exactly that — browse photorealistic .splat assets across architecture, interiors, nature, vehicles, and more, ready to drop into your project.
Frequently asked questions
What is Gaussian splatting in simple terms?
It is a way to turn photos or video of a real scene into a photorealistic 3D model made of millions of tiny coloured blobs (3D Gaussians) that blend together. The result looks like the real place and can be explored in real time.
Is Gaussian splatting the same as a point cloud?
No. A point cloud is a set of single-coloured points. Gaussian splatting uses soft, scalable, semi-transparent ellipsoids with view-dependent colour, so it reconstructs a far more realistic, continuous-looking image than raw points.
Is Gaussian splatting AI?
Not in the usual sense. Unlike NeRF, it does not use a neural network to render. It uses an optimisation process (gradient descent) to fit the blobs to your photos, which is a machine-learning technique, but the final splat is just data — no AI model runs when you view it.
What file format is a Gaussian splat?
Splats are most commonly stored as .ply or .splat files. Newer compressed formats such as .spz and .sog are also appearing. Most viewers and game engines accept .ply or .splat.
Do I need a powerful computer to view a Gaussian splat?
Viewing is lightweight — many splats run smoothly in a web browser or on a phone. Training (creating) a splat is the demanding part and usually needs a GPU, which is why many people use cloud apps or buy ready-made splats.