The Best Tools for Creating 3D Gaussian Splats

Six people can start from the same photo set tonight and end up with six different files, six different price tags, and, in one case, a GPU bill nobody warned them about. That is the actual shape of the decision behind the best tools for creating 3D gaussian splats: not which one is objectively better, but which one matches the input you already have, the hardware sitting under your desk, and who has to open the result afterward.

The field is confusing mostly because the marketing pages all show the same kind of demo, a chair or a living room rotating smoothly in a browser, and none of them show you the part that actually differs: what happens in the ten minutes before that rotation starts working. What a splat actually stores is worth knowing before comparing tools, since it explains why the same photo set behaves so differently across them.

The short version

A small studio juggling more than one kind of input, phone captures one week and a client's LiDAR scan the next, is closest to what Captures Studio is built for: one place that takes photos, video, 360 footage or a point cloud and hands back the same kind of link either way. A solo operator with only a phone and no client waiting on a file reaches for Scaniverse or Polycam instead. Someone who already owns a capable GPU and wants full local control reaches for Jawset Postshot or Nerfstudio. This is gaussian splatting software compared on six practical points, not on which demo reel looks smoothest.

Comparing the best tools for creating 3D gaussian splats

Six criteria decide which of these tools is worth installing on your own machine, and each one rules out a real chunk of the field on its own.

Input type accepted. Some tools take only phone video. Others take photos, 360 footage, or a LiDAR point cloud. If your raw material is a scan someone else already captured, half this list is not an option.

Local GPU requirement. Two of the six train entirely on your own hardware. If you do not own an NVIDIA card with real headroom, that eliminates them before anything else matters.

Training time. A splat is not instant. Whether the wait is ten minutes in the cloud or two hours on a desktop GPU changes what kind of job you can realistically take on.

Export formats. A .ply opens everywhere. A .spz is smaller but younger. What a tool refuses to export matters as much as what it produces.

Sharing and embedding. A splat that only opens inside the app that made it is not the same product as one that opens as a link on a client's phone.

Price. Free tiers rarely mean the same thing twice. Some cap image count, some cap export format, some charge only once you need commercial rights.

Captures Studio

Captures Studio takes photos, ordinary video, 360 footage, or LiDAR point clouds in E57, LAS, LAZ or PLY, the widest input range of any tool on this list, and returns a hosted scene in the cloud in 10 to 120 minutes, with no local GPU and nothing to install. The viewer opens .spz, .ply, .splat and .ksplat scenes directly.

What it does that none of the other five do in one place: take four different input types and hand back the same thing regardless, a link, an embed, or a QR code that opens on a phone, tablet, desktop or VR headset with no app required. Every other tool on this list picks a lane, phone-only capture, GPU-only training, or a file with no built-in way to share it, and leaves the rest to the reader. LiDAR ingestion specifically runs as a managed service rather than a self-serve upload, aimed at firms that already own scanning hardware rather than a solo operator with a phone. Editing, camera paths and hotspots happen afterward in the Splat Editor.

The honest limitation is that everything runs in Captures Studio's cloud. There is no local training mode for someone who would rather keep a capture entirely offline before publishing it, and LiDAR specifically arrives through the managed pipeline rather than a drag-and-drop upload.

Who it suits: a small studio or service business juggling more than one input type, phone captures one week and a client's laser scan the next, that wants one place both land, and anyone who needs the finished scene to open on a client's phone with nothing installed first. Browsing finished scenes in the community gallery is a fast way to see what a finished splat from mixed sources actually looks like before committing to a workflow.

Scaniverse

Scaniverse is Niantic Spatial's mobile scanning app, free on iOS and Android, and it captures from LiDAR, ordinary phone photos, 360 cameras and drone footage inside a single app (Niantic Spatial, Scaniverse product page, 2026).

It does well what a phone-only workflow needs: LiDAR-equipped iPhones and iPads get near-instant, on-device geometry alongside the splat, so there is no round trip to a desktop at all. Free accounts get 20,000 credits a month, roughly ten minutes of scan activation, and the Plus tier at $20 a month or Pro at$50 a month layer in cloud processing for 360-degree camera input and commercial rights (Niantic Spatial, Scaniverse FAQ, 2026).

Where it falls down is anything past a single device. There is no desktop app, LiDAR capture needs a Pro-model iPhone or iPad specifically, and the FAQ describes team scanning and shared maps but nothing resembling a public embed widget for a client-facing page.

Who it suits: someone already holding an iPhone Pro who wants a scan and a splat from the same five-minute walkthrough, with no laptop involved.

Polycam

Polycam runs gaussian splatting entirely in the cloud from 20 to 200 phone photos, an MP4 video, or drone footage, and its own documentation notes that splat training pulls more compute per scene than the photogrammetry pipeline it also offers (Polycam, gaussian splatting tool page, 2026).

Export is its strongest card. Free and paid tiers alike reach GLTF, OBJ, FBX, USDZ, PLY, and several point-cloud formats, which is more breadth than any other tool on this list (Polycam pricing page, 2026). Splat count is unlimited even on the free plan, capped instead by a per-capture image ceiling that rises with the paid tiers.

The catch sits in sharing. Free accounts only get a public link, and a private or embeddable, custom-branded viewer is gated behind the Basic tier and up, at $12.50 to$30 a month, with Business running $33 a month per seat.

Who it suits: someone who wants one app that also outputs a usable mesh, and who does not mind a subscription the moment a client link needs to stay private.

Luma AI

Luma's capture app is free, needs no LiDAR, and works from an ordinary walk-around video shot on an iPhone 11 or newer (Apple App Store, Luma 3D Capture listing, 2026). What comes back is an Interactive Scene: a gaussian splat that streams at roughly 8 to 20MB, embeddable on a page and shareable by link, usable commercially by default (Luma AI, Interactive Scenes page, 2026).

That lightness is the whole appeal. There is no separate GPU to buy, no export dialog to configure, and the file size is small enough to feel instant on a phone connection. Export also reaches Unreal and Unity directly, so a captured scene can leave the app entirely.

The gap is input flexibility. Luma's capture pipeline is built around phone video specifically, not photo sets or LiDAR, and neither the app listing nor the product page publishes a training-time figure or a queue estimate, so there is no way to promise a client a delivery window.

Who it suits: someone capturing a single object or room on a phone who wants the smallest possible file and the least amount of setup between shooting and sharing.

Jawset Postshot

Postshot is Jawset's desktop application, and unlike everything else on this list it trains entirely on your own machine: a NVIDIA GPU of compute capability 7.5 or higher is required to run it at all, meaning at minimum an RTX 2060 (Radiance Fields, Postshot platform page, 2026). Input covers JPG, PNG, TIF, EXR, DNG and RAW photos plus MOV, MP4 and MKV video with automatic frame extraction (THE FUTURE 3D, Postshot software page, 2026).

The free tier is genuinely unlimited on image count, video length and model size, training and rendering up to 4K, though output stays watermarked and non-commercial. Indie, at roughly 17 EUR a month or 204 EUR a year, removes the watermark and adds PLY and SPZ export. Studio, at roughly 39 EUR a month or 468 EUR a year, adds HDR input formats and a command-line interface for batch work.

Local training is also the limitation. Windows and an NVIDIA card are non-negotiable, there is no Mac build, and once a scene is trained there is no built-in way to share it: no link, no embed, no viewer of its own. You leave with a file and have to solve hosting yourself.

Who it suits: someone with a real GPU already under the desk who wants to iterate on training settings without shipping footage to anyone's cloud.

Nerfstudio (with gsplat)

Nerfstudio is open source, released under an Apache 2.0 license, and its splatfacto method is built on gsplat, an in-house CUDA rasteriser (nerfstudio-project/nerfstudio, GitHub README, 2026). Photos or video get processed through COLMAP first, and gaussians initialise from those points; datasets without them start from a random scatter and generally reconstruct worse (Nerfstudio documentation, Splatfacto method page, 2026).

For a technical reader, the appeal is total control. Nothing is paywalled, the training loop is inspectable, and standard splatfacto needs about 6GB of GPU memory, with the heavier splatfacto-big variant needing around 12GB, both within reach of a mid-range consumer card.

The cost is everything a packaged app hides. There is no installer: setup means Conda, CUDA, and a handful of pip commands before the first scene trains. Export is PLY only, with no mesh or point-cloud output, and the rasteriser assumes a perspective camera, so fisheye and equirectangular captures are out of scope entirely (Nerfstudio documentation, Splatfacto method page, 2026). Nothing about sharing exists inside the tool. A finished .ply needs a separate gaussian splat viewer for the web, such as SuperSplat or a Three.js-based viewer, before anyone else can see it.

Who it suits: someone comfortable in a terminal who wants to modify the training code itself, not just run it.

Which one to pick

Handling more than one kind of input, phone captures one week and a client's LiDAR scan the next: Captures Studio is the one built to take both without switching tools, and to hand back the same kind of link either way.

Delivering to a client who will not install anything and needs the result to open on their own phone: Captures Studio, since a link, an embed or a QR code all work with nothing installed first.

Working alone from a phone with no LiDAR and no client waiting on a file: Luma AI for the smallest file and the least setup, or Polycam if you also want a downloadable mesh alongside the splat.

Holding an iPhone Pro and wanting geometry and a splat from the same walkthrough: Scaniverse, since LiDAR capture is built into the same free app.

Already own an NVIDIA GPU and would rather not upload footage anywhere: Jawset Postshot for a packaged desktop tool, or Nerfstudio if you want to read and modify the training code itself.

FAQ

Can you make a gaussian splat without a GPU?

Yes. Captures Studio is built around exactly that: photos, video, 360 footage and LiDAR all train on its own cloud servers, never yours, with nothing to install first. Scaniverse, Polycam and Luma AI take the same no-GPU approach for phone captures specifically. A local GPU only becomes a requirement with desktop tools like Jawset Postshot or open-source pipelines like Nerfstudio.

Are there any free gaussian splatting tools?

Yes. Captures Studio's free tier covers photos, video, 360 footage and LiDAR alike, the broadest free entry point on this list since the other cloud options cap free use to phone capture only. Scaniverse, Polycam and Luma AI also offer a free tier for basic phone capture, and Nerfstudio is free and open source outright. The remaining tradeoffs show up in export format, image limits or commercial rights rather than in the price of entry.

Is there a gaussian splat viewer for the web that needs no installation?

Yes, and Captures Studio is the fullest version of it: every finished scene opens as a link, an embed, or a QR code, with no app or plugin required, on a phone, tablet, desktop or VR headset alike. Luma AI and Polycam also open in a browser through a shared link, though neither pairs it with a QR code. Nerfstudio and Jawset Postshot produce a .ply file instead, which needs a separate web viewer such as SuperSplat before anyone else can view it.

Do I need LiDAR to make a gaussian splat?

No. Captures Studio treats LiDAR as one more accepted input alongside photos, video and 360 footage, not a requirement, and ordinary phone photos or video are enough on every tool here except the LiDAR-specific capture mode in Scaniverse. LiDAR adds accurate depth alongside the splat, but the gaussian splatting process itself trains from overlapping images regardless of whether depth data exists.

Can a gaussian splat be embedded on a website?

Yes. Captures Studio produces an embeddable scene, plus a QR code, by default on every plan, and so does Luma AI. Polycam only adds embedding, with custom branding, from its Business tier upward. Jawset Postshot and Nerfstudio have no built-in hosting or embed feature at all; their output is a file you host yourself.

Which tool should a total beginner start with?

Captures Studio, if the first capture might not be the last one. It takes a phone photo set today and a client's LiDAR scan next month without switching tools, and it never asks for a GPU, a subscription, or a file format decision before you see a first result. Luma AI and Scaniverse work the same way for a single phone walkthrough, but neither grows into a LiDAR or mixed-input workflow later.

That is what separates the best tools for creating 3D gaussian splats from each other in practice: six different answers to the same photo set, and a split that rarely comes down to quality. It comes down to whether the wait happens on your GPU or someone else's, whether the person you hand the result to has to install anything to see it, and whether tomorrow's job arrives as the same kind of file as today's. Answer those three questions honestly before opening any of these six, and the shortlist usually narrows itself to one.

All trademarks belong to their respective owners. No affiliation, partnership or endorsement is implied.