Gaussian splatting: photorealistic 3D leaves the lab

For a long time, digitising a place in 3D meant picking a side. You could model everything by hand, with a clean result but a heavy budget and an output that is always an interpretation. Or you could use photogrammetry, which is more faithful but produces meshes that are often noisy, hard to optimise and struggle with shiny or thin surfaces. Since 2023, a third option has taken hold at a speed rarely seen in computer graphics: Gaussian splatting.

The method was published by a team at Inria in Sophia Antipolis, France, and presented at SIGGRAPH 2023. Three years on, it is built into consumer apps, 3D engines and web viewers. It is a good time to look at what it really brings, and what it cannot do yet.

How it works, without the maths

You start with a set of photos or a video that circles an object or moves through a space. Software first works out where the camera was for each image. Then, instead of rebuilding surfaces out of triangles, it fills the space with millions of small, soft, ellipse-shaped blobs: the Gaussians. Each one has a position, a size, an orientation, a colour that changes with the viewing angle, and a level of transparency.

Training means adjusting those millions of parameters until the scene, seen from each camera position, looks as close as possible to the original photos. Once that is done, display is very fast: the graphics card projects the Gaussians onto the screen and blends them, with no heavy computation. The result is a scene you can explore in real time, with reflections, foliage and fine detail that classic photogrammetry handled poorly.

We no longer rebuild surfaces, we rebuild appearance. That shift in thinking is what makes it look so real.

Why everyone jumped on it

Before Gaussian splatting, the reference for this kind of rendering was NeRF, a neural network approach that appeared in 2020. The results were stunning, but each frame took seconds to render. Splatting delivered the same visual quality with smooth playback, including in a web browser or a virtual reality headset. Moving from the lab to real time is what set everything off.

Today, a decent capture can be done with an ordinary smartphone. Apps such as Polycam, Scaniverse or Luma can scan a room or an object in a few minutes, and more professional tools give control over every step when production quality is the goal.

What it changes for real projects

• Virtual tours: a museum, an industrial site or a property can be explored freely, not just hopping from viewpoint to viewpoint as in a classic 360° tour.
• Heritage: a monument or an exhibition room is captured in a precise state, close to a photograph, for archives as well as public engagement.
• VR training: learners can be placed in the real workshop, with its real visual landmarks, rather than in a rough reconstruction.
• Marketing and sales: a product or showroom can be presented in 3D on a web page, with nothing to install.

The clearest gain is the cost of entry. Convincingly digitising an existing place no longer takes weeks of modelling. Capture takes half a day, processing a few hours, and the team can focus on what matters: the user journey, the interactions and the message.

Limits to know before you start

The technique is not magic, and it is better to know that before promising a result. A faithful scene quickly weighs several hundred megabytes, so it has to be compressed and split up for the web or a standalone headset. Lighting is baked into the capture: you cannot relight the scene like a classic 3D model. Poorly filmed areas produce blurry artefacts, and anything that moves during the shoot leaves traces.

Another point: a Gaussian scene is not a mesh. For physics, collisions or detailed editing of an object, splatting is often combined with simplified geometry. That is what the most polished projects do: photorealistic rendering for the eye, a lightweight model underneath for interaction.

Our take

At EVAVEO, we see Gaussian splatting as one more tool in the box, not a universal replacement. For a virtual tour, a showcase digital twin or a training experience set in a real place, it is often the best option today. For a product that needs to be configured, animated or lit differently, classic modelling still has the edge. The right answer depends on the use case, and that is the first question we ask.

In short

Gaussian splatting turns a few minutes of video into a photorealistic 3D scene you can explore in real time. Born at Inria in 2023, it is already production-ready for virtual tours, heritage and training. Its limits are well known (file size, baked lighting, interaction handled separately) and they are easy to manage when planned for from the start.