What Is Gaussian Splatting? Real-Time Radiance Field Rendering, Explained

What is Gaussian splatting and how does it work?

Gaussian splatting — formally 3D Gaussian splatting (3DGS) — is a technique in computer vision and graphics for turning a set of images of a place into a photorealistic, explorable 3D scene that renders in real time. The algorithm was introduced in the 2023 SIGGRAPH paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" by Kerbl, Kopanas, Leimkühler and Drettakis, and it has reshaped 3D reconstruction and novel view synthesis faster than almost any method before it — going from research paper to state-of-the-art in production tools within a year.

The core idea: instead of using a mesh of triangles or a dense volumetric grid to represent the scene, the scene is represented as millions of 3D Gaussians — soft, semi-transparent ellipsoids, each with a position, size, orientation, colour and opacity. Think of each one as a tiny airbrushed blob ("splat") floating in space; mathematically, each projects to a 2D Gaussian when rendered to screen. Individually they look like nothing; rendered together, in their millions, they reproduce surfaces, reflections and fine detail with startling, high-quality fidelity.

The pipeline runs in four stages:

1. Capture a dataset. A set of overlapping 2D images of the scene — anywhere from dozens to thousands of photos, or frames pulled from video.
2. Recover camera positions. A structure from motion tool (almost always COLMAP) works out where every photo was taken from and produces a sparse point cloud as a starting skeleton.
3. Optimization. Each point is seeded as a 3D Gaussian, and an optimization loop repeatedly renders the splats against the original photos, measures the difference, and adjusts position, shape, colour and opacity to optimize the match — splitting and pruning Gaussians as it goes. This is gradient-descent training, like a neural network, but the thing being trained is the explicit set of splats rather than network weights. Using Gaussian primitives this way is what separates 3DGS from earlier radiance field methods: the representation stays explicit and fast to render.
4. Rasterization. The finished scene renders through a fast, GPU-friendly rasterizer that sorts and blends the splats for any viewpoint — which is what makes real-time rendering possible.

The result is a radiance field: a representation that captures how light leaves the scene in every direction, enabling novel view synthesis — rendering the scene from viewpoints no camera ever stood at — including view-dependent effects like reflections on glass and polished floors that traditional 3D models handle badly.

It's worth unpacking what the 3DGS paper actually claimed, because every phrase earned its place. Radiance field methods had recently revolutionized 3D scene reconstruction from multiple photos or videos — but slowly. The authors' contribution was threefold: represent the scene with 3D Gaussians that preserve desirable properties of continuous volumetric radiance fields while avoiding wasted computation in empty 3D space; optimize that set of 3D Gaussians from the sparse points produced during camera calibration; and develop a fast visibility-aware rendering algorithm that supports anisotropic splatting, which both accelerates training and allows realtime rendering. The headline result: state-of-the-art visual quality and real-time rendering on several established datasets (including the synthetic Blender dataset), while maintaining competitive training times. In plainer English: Gaussian splatting is a rasterization technique that made high-quality real-time rendering of photorealistic scenes learned from photos practical on ordinary hardware — millions of 3D Gaussians as a representation of 3D space, rendered fast enough for 3D computer vision to leave the lab.

How does Gaussian splatting compare to NeRF and photogrammetry?

Three technologies get mentioned together, and they solve the problem in genuinely different ways:

• Photogrammetry builds an explicit textured mesh from photos. It produces measurable, editable geometry — the right tool for CAD and survey deliverables — but struggles with reflective, transparent and thin objects, and its output looks like a 3D model rather than a photograph.
• NeRF (neural radiance field) stores the scene inside a neural network. Quality is excellent, but every pixel of every frame requires querying the network, so rendering is slow without heavy optimisation — historically seconds per frame, not frames per second.
• 3D Gaussian splatting keeps NeRF's photorealism but stores the scene as explicit splats that rasterize like ordinary graphics primitives. The original paper demonstrated real-time radiance field rendering at 1080p (≥30 fps, and over 100 fps on a strong GPU) with training measured in minutes rather than days — which is why 3DGS displaced NeRF in most practical applications within about a year.

The honest comparison for commercial buyers: photogrammetry and LiDAR for measurement, Gaussian splatting for presence. A splat scene is not a survey-grade model — but nothing else makes a space feel as real on a screen. Our view on where that fits commercial property marketing is in our Gaussian splatting virtual tours post.

How do you create a Gaussian splat?

The accessible route in 2026, in order of effort:

1. Phone apps and hosted services — Polycam, Luma AI and similar tools let you capture video on a phone and receive a trained splat in minutes. Quality is capture-dependent but the barrier to entry is essentially zero.
2. Open-source pipeline — COLMAP for camera poses plus a trainer such as the original INRIA implementation, Nerfstudio's gsplat, or Postshot. You'll want an NVIDIA GPU with at least 8–12 GB of VRAM for small scenes; large scenes benefit from 24 GB+.
3. Professional capture — for large or complex spaces, capture quality dominates results. We capture with a LiDAR scanner that records 134-megapixel panoramas alongside the point cloud, and our free Galois M2 E57-to-COLMAP converter turns that survey data into a COLMAP-ready dataset — giving the trainer geometrically accurate camera poses instead of estimated ones.

Editing is improving fast: tools like SuperSplat and Postshot allow cropping, cleaning floaters and compressing scenes, and splats can be georeferenced by aligning them to surveyed control points or a registered point cloud — which is exactly where a LiDAR-first capture pays off.

What are the benefits and limitations of Gaussian splatting?

Benefits:

• Photorealistic output — including reflections and view-dependent lighting that meshes can't reproduce
• Real-time rendering — smooth exploration in a browser or engine, no pre-rendered video
• Fast training — minutes to low hours on consumer GPUs, versus the days early NeRFs needed
• Explicit representation — splats can be edited, merged, compressed and streamed more easily than a neural network's weights

Limitations:

• Not survey geometry — you can't take reliable millimetre measurements off splats alone; pair with a point cloud survey where dimensions matter
• File size — millions of Gaussians produce large files, though compression formats (such as SOG and .spz) are shrinking them rapidly
• Capture sensitivity — moving people, changing light and sparse coverage produce artefacts ("floaters")
• Dynamic scenes remain research territory — 4D splatting exists but isn't routine production yet

What is Gaussian splatting used for?

The applications have spread well beyond research demos:

• Property and venue marketing — explorable, photoreal walkthroughs of real spaces
• Film, TV and VFX — digital backlots and virtual production environments scanned from real locations (one reason film location scouting is converging with 3D capture)
• Games and real-time engines — splat renderers now exist for Unreal Engine, Unity and Blender, plus WebGL/WebGPU viewers for the browser
• Heritage documentation — photoreal records of spaces that meshes flatten
• Robotics and simulation — photoreal environments for training perception systems

Where is the technology heading?

Three visible directions: aggressive compression and streaming (making splats as portable as video), dynamic scenes (capturing motion, not just spaces), and integration with measurement workflows — splats aligned to survey-grade point clouds so one capture serves both presence and precision. That last one is our world: the same visit that produces a measurable LiDAR dataset can now produce the training data for a photorealistic splat.

Want a photorealistic capture of your space?

If you're curious what Gaussian splatting looks like applied to a real venue — or you need a capture that delivers both survey-grade LiDAR data and splat-ready imagery from one visit — we're a London-based team that does exactly that, with 134-megapixel capture and delivery in 3–5 working days. Get in touch or see our portfolio.