Do I need LiDAR for Gaussian splatting on an iPhone?

No. Gaussian splatting works from regular photos or video. Camera positions come from motion tracking and are refined mathematically. Scantic runs on iPhones with and without LiDAR.

How long does it take to create a Gaussian splat on a phone?

With on-device processing in Scantic, a typical scan takes a couple of minutes from the end of capture to a finished walkable scene. Larger rooms take longer than single objects, and newer iPhones are faster.

What is the difference between Gaussian splatting and photogrammetry?

Photogrammetry builds a mesh of triangles with photo textures, which works well for solid objects but struggles with thin structures, glass, shiny surfaces and foliage. Gaussian splatting represents the scene as millions of translucent colored blobs and reproduces those difficult materials much more faithfully, including how they change with viewing angle.

Is Gaussian splatting the same as a NeRF?

They solve the same problem differently. A NeRF stores the scene inside a neural network and needs heavy computation for every frame it renders. A Gaussian splat is an explicit list of blobs that a GPU can draw directly, so it renders in real time, even in a phone or browser.

What file formats do Gaussian splats use?

The most common formats are PLY, the original research format storing every Gaussian's parameters, and SPZ, a compressed format that is much smaller at similar quality. Scantic exports both, and they open in tools like SuperSplat or Blender with a splat add-on.

Can I create Gaussian splats for free?

Yes. Scantic is a free iPhone app that captures and computes Gaussian splats entirely on the phone, offline, with no account.

What is Gaussian splatting?

Explained by playing with it · last updated 15 August 2026

The thing below is not a photo and not a video. It is a Gaussian splat: a statue in a Vienna church, captured by walking around it with an iPhone. Drag it. Go around the back. No camera ever stood in most of the places you are looking from.

3D scan of a Madonna statue in the Otto Wagner church in Vienna

Scanned and computed entirely on an iPhone with Scantic, in 76 seconds. Pruned for the web: what you see here is 333,000 of the original 416,000 blobs.

So how can a phone build that? The honest answer fits on this page, and you can play with every step of it. There is no magic and, surprisingly, no AI guessing involved: just one simple ingredient, multiplied by a few million, plus patience.

The only ingredient: a blob

This is a Gaussian. One single blob: solid in the middle, fading softly to nothing at the edges. It has a position, a width, a height, a rotation, a color and a transparency. That is the complete list. Play with it:

Width

Height

Rotation

Opacity

Color

Not impressive. Neither is a single pixel, or a single triangle. The trick is never the ingredient. It is what arranges it.

Stack a few thousand of them

Below are hundreds of these blobs. Not painted by hand: a computer arranged them, and we will get to how in a second. Slide from one blob to all of them and watch a photograph condense out of colored mist:

Blobs

The big fuzzy blobs block in walls and robe; the small sharp ones carry edges and detail. A real 3D splat works the same way, just with millions, in three dimensions.

Nobody places the blobs. They learn.

Here is the part that sounds like cheating but is just honest mathematics. The computer starts with rough blobs scattered where the photo has detail, each taking the color it happens to sit on. Then it repeats one dumb loop: render the blobs, compare with the photo, nudge every blob a tiny bit so the difference shrinks. Position, size, rotation, color, transparency, all at once, thousands of times. That nudging is called gradient descent, and it is the same engine that trains neural networks. Scrub through the real optimization:

Training

Every notch is a real saved state of the optimization that fitted the blobs above, not an artist's impression. Every scan in Scantic goes through exactly this process, in 3D, on your phone.

That is the whole secret. A Gaussian splat of a room is this exact idea in three dimensions: the blobs become tiny translucent ellipsoids floating in space, the photo becomes all your photos at once, and the nudging must make the blobs match every photo from its own viewpoint. When it converges, the blob cloud reproduces the room, including viewpoints you never photographed, which is what you experienced with the statue at the top.

Why not just use triangles?

Classic 3D scanning (photogrammetry) wraps the world in a mesh of textured triangles. For a solid matte object that works well, and if you need something to 3D print, a mesh is still the right tool. But point it at the real world and it breaks: plant leaves, railings, hair, glass, a polished floor. A mesh must commit to where the surface is, and for a bush or a window there is no good answer, so it tears holes or melts everything into mush.

Blobs never have to commit. A translucent blob can hover in the volume of a bush or sit as a moving highlight on glossy stone. Blobs can even shift their color slightly depending on where you look from, which is how splats reproduce shine and reflections that travel with you. That is why walking through a good splat feels like being there, in a way textured meshes rarely manage.

How your iPhone pulls this off

Until recently this took a desktop with a gaming GPU: hours of figuring out camera positions, then more hours of blob nudging. A phone changes the recipe twice. While you walk, its motion sensors and camera tracking already know roughly where every frame was taken, so the expensive pose puzzle collapses into a quick mathematical clean-up (bundle adjustment, about a second). And Apple's chips share one memory pool between CPU and GPU, which is precisely what blob optimization at room scale needs.

Scantic runs all three stages directly on the iPhone, offline, in a couple of minutes. The statue above took 76 seconds of optimization on an iPhone 17. Nothing is uploaded; the finished splat lands in your library, ready to walk through, share as an end-to-end encrypted link, or export.

How to capture a good splat

The optimization can only reproduce what your photos show, so capture decides quality. Move in a slow circle around your subject and cover every side; surfaces no photo saw come out as holes or fog. A second pass higher or lower fills tops and undersides. Light is quality: dim rooms make blurry, noisy splats. And slow beats smooth: move your feet instead of swinging the phone, because the method needs parallax, not panning.

Frequently asked questions

Do I need LiDAR? No. Splatting works from regular photos. Camera positions come from motion tracking plus mathematical refinement. Scantic runs on iPhones with and without LiDAR.

How long does it take on a phone? With Scantic, a couple of minutes from the end of capture to a finished scene. Objects are faster than rooms, and newer iPhones are faster than older ones.

Splatting vs. photogrammetry? Meshes for solid matte objects and when you need printable geometry; splats for faithful, walkable reproductions of real places with difficult materials.

Splatting vs. NeRF? Same goal, different representation. NeRFs compute every pixel through a neural network and are slow to display. Splats are explicit geometry that GPUs draw in real time, like the one at the top of this page.

Which file formats? PLY (the original research format) and SPZ (compressed, much smaller at similar quality). Scantic exports both, and they open in SuperSplat, Blender with a splat add-on, and a growing list of engines.

What does it cost? Scantic is free on the App Store for iPhones running iOS 17 or later.

The best explanation is your own sofa. Get Scantic free on the App Store, walk once around it, and ninety seconds later you will be standing inside this page.