Ask a 3D artist how long it takes to model a working processing plant for VR, and the honest answer is usually measured in weeks. Every pipe and handrail has to be built and compiled by hand. 3D Gaussian splatting flips that process around. Instead of modelling a place, you film it, all of it, from every angle, every nook and cranny, and specialised AI software rebuilds it as a photorealistic 3D scene you can walk through on a computer, tablet or even in a headset.
For businesses that need people to understand a real location before they get there, that’s a big shift. A site induction can show the actual site. A showroom can be the real showroom. A design review can happen inside a faithful copy of the space being changed. And a remote plant room can be explored from an office in Perth.
It isn’t a shortcut to good VR, though. Here’s where it adds real value, and what to watch out for.
What Is 3D Gaussian Splatting?
3D Gaussian splatting is a way of recreating real places in 3D from specialised photos or video. It was introduced in 2023 by researchers at Inria and Université Côte d’Azur in France, working with the Max Planck Institute for Informatics. They presented it at SIGGRAPH, the world’s leading computer graphics conference.
Traditional 3D models are built from a series of tiny triangles called “polygons”. A Gaussian splat scene works differently. It’s made of millions of soft, semi-transparent points, each with its own position, shape, colour and opacity. Blended together, they produce an image that looks much more like a photograph than a computer model.
The breakthrough was speed. Earlier radiance field methods, such as NeRF, could produce similar realism but struggled to render complete scenes in real time. The original paper showed scenes rendering at 30 frames per second or more at 1080p resolution. Image quality matched the best earlier methods. That real-time performance is what makes the technique practical for VR.
How Does 3D Reality Capture Become a VR Experience?
The process starts with 3D reality capture: walking a camera through the space and filming it from every possible angle. Anything from a phone to a drone can do the job. Insta360 and Splatica, for example, now offer a pipeline that turns five minutes of 360 video into an explorable splat within hours. At the end of the day it is less about the hardware and more about the techniques involved in the capture process itself.
A specialised AI then studies those images and works out where millions of splats need to sit to reproduce every view. This step, often called training, is mostly automated, although the capture still needs care. Missed angles become blurry patches, and moving people or vehicles leave ghostly marks. Good captures come from slow, steady movement with plenty of overlap between views, filmed in one session so the lighting stays consistent. Our mining site inductions are built from stereoscopic 360 video filmed on site, sometimes from a heavy-lift drone. The same capture discipline applies to splats.
Next comes clean-up. Raw scenes are large, so they’re trimmed and compressed for the headset that will run them. Not to mention, they are almost always very messy.
Then the scene becomes an experience. Interaction, guidance, audio and scoring are layered on top, so trainees or customers can actually do something in the space rather than only look around.
Why Are Businesses Turning to Gaussian Splatting VR?
Businesses are adopting Gaussian splatting VR because it makes photorealism far cheaper to produce. Every surface used to be modelled or photographed and textured by hand, so realistic VR was reserved for big budgets. That changes what’s worth building.
Big platforms are leaning in. Meta’s Horizon Hyperscape lets Quest 3 and 3S owners scan a room with the headset and revisit it as a photorealistic VR scene. The heavy processing happens in the cloud. Apple has made moves too: in visionOS 27, its RealityKit framework can render 3D Gaussian splats natively on Vision Pro.
That doesn’t make older techniques obsolete. Back in 2017, Viewport XR built a VR showroom for OzGrind with every polished concrete finish recreated in 8K detail. Customers could get right down to floor level without losing quality. Splats add a new option to that toolkit, especially for spaces that are slow to model, texture and light by hand.
They also fill a gap that 360 video can’t. 360 video is still one of the fastest ways to put people on site, but viewers can only look around from where the camera stood. A splat scene is fully volumetric, so people can crouch down or walk across the room, and the view updates as they move. For training and site familiarisation, that freedom makes a real difference.
Splats are also starting to move. At Laval Virtual 2026, several companies showed animated 4D splats that capture people and motion, gradually taking over from volumetric video. For training, that opens the door to capturing an expert at work and letting trainees walk around them as they watch.
Where Do Photorealistic VR Environments Add the Most Value?
Photorealistic VR environments earn their keep when recognising the real place matters to the outcome.
Site Inductions and Safety Training
New workers learn faster when the induction looks like the site they’ll walk onto. Our VR mining induction for Rio Tinto at Parker Point replaced physical site tours of up to six hours with a 30-minute VR session. Annual induction costs fell by about 75%. Captured environments can make that kind of program even more familiar, because trainees see the actual layout and signage.
Design Reviews and Stakeholder Engagement
Changes to an existing facility are easier to judge inside a faithful copy of it. Teams can review a proposed upgrade in context, and community members can see how a project fits a place they know. Several people can explore the same scene together, so everyone is working from the same picture.
Remote and Hard-to-Reach Places
Some locations are costly or risky to visit, from remote Pilbara mine sites to restricted plant rooms. A captured scene lets specialists and new starters get to know the space before they ever fly in. For FIFO operations, that can mean new crew who already know where the muster point is on their first day.
What Are the Limits of 3D Gaussian Splatting?
3D Gaussian splatting is powerful, but it has clear limits, and planning around them is what separates a good project from a disappointing one.
A splat records a single moment. Lighting and weather appear exactly as they were on the day. Relighting is still limited, although some rendering tools have started to add it. Splats also struggle where the camera never looked. In a hands-on review of Meta’s Hyperscape, UploadVR noticed distortion on small text and underneath furniture, in spots the headset never saw during the scan.
Performance is the other challenge. Large scenes demand a lot from standalone headsets, which is why Meta streams Hyperscape from the cloud. Compression helps, and PlayCanvas reports its web format shrinks splat files by 15 to 20 times, but big sites still need careful preparation.
Then there’s what the camera picks up. Every 3D Gaussian splatting capture records everything in view, including screens and faces. On mining or healthcare sites, agree what can be filmed before anyone picks up a camera, and remove anything sensitive during clean-up.
Lastly, a splat looks real without behaving like real objects. It doesn’t know where the floor ends or which valve can turn. That’s why the strongest projects combine a captured scene with interactive elements built by our custom app development team.
One rule we hold to on every build: capture what people need to recognise, then design everything they need to do.
How Does Immersive Site Visualisation Fit with Your Other Spatial Data?
Splats are becoming part of the wider geospatial world, and standards are catching up. In February 2026, the Khronos Group released a candidate extension for storing splats in glTF, the widely used format for delivering 3D content. Esri, Cesium, Niantic Spatial and the Open Geospatial Consortium all worked on it.
Bentley Systems, which owns Cesium, has said splats will be a key feature of 3D Tiles 2.0. That’s the Open Geospatial Consortium’s upcoming standard for large-scale 3D data. That means immersive site visualisation won’t sit in a silo. It’ll increasingly live alongside maps and survey data.
That matters for anyone managing infrastructure or land. A photorealistic capture shows what a place looks like today. Accurate positioning shows exactly where things are. Viewport XR’s GEAAR, an augmented reality geospatial platform, already pins 3D models to real locations within centimetres using satellite and visual positioning. Bringing that kind of accuracy together with captured scenes is where this work is heading.
Is 3D Gaussian Splatting Right for Your Next VR Project?
If your project depends on people recognising a real place, the answer is probably yes. Splatting can get you a believable environment faster and cheaper than modelling it by hand.
The capture is getting easier every year. Knowing what to do with it isn’t. A beautiful scene with nothing to do inside it won’t change how anyone works.
At Viewport XR, we see splats as a powerful starting point, not a finished product. Some parts of a site are best captured, while others are better built by hand so they can respond to what people do. We make those calls based on what your people need to learn or decide. Pair a captured site with SASHIMI, our multi-user platform for design reviews and team training, and a whole team can review it together.
Talk to our team about turning a real location into a VR experience that does more than look good.
FAQ
They’re good at different things. Photogrammetry produces a solid 3D mesh, which suits digital twins and physical interactions. Gaussian splatting usually looks more lifelike, particularly with soft materials and complex lighting, but it doesn’t produce clean solid surfaces on its own. However both methods are not exclusive and a single project can use both capture methods.
Yes. Meta’s Horizon Hyperscape shows photorealistic splat scenes on Quest 3 and 3S, and web-based viewers can display splats on headsets that support WebXR. Large scenes need compressing or cloud streaming to run smoothly on standalone headsets. Outside of this however, it is best to get the help of a 3D studio like Viewport XR to clean, optimise and deploy your splat for you.
No. A 360 video lets viewers look around from a fixed point where the camera stood. A Gaussian splat scene is fully 3D, so people can walk around inside it and see the space from new angles.
A single room can be filmed in minutes, while a large site may take a few hours of careful capture. Processing then usually takes hours rather than weeks. Cleaning up the scene and building interaction on top is often the longest part.
Yes, within limits. Unwanted splats can be removed and colours adjusted, which is how stray objects and sensitive details are cleaned up. Reshaping or moving objects is much harder than with a traditional 3D model, so interactive elements are usually built separately.