Gaussian splats on the iPhone?

For those who don’t know the tool: Scantic is an iPhone capture and reconstruction application developed by Sebastian Beyer that turns camera footage of objects, rooms and environments into 3D Gaussian splats. It puts acquisition, reconstruction, and inspection on the phone rather than sending source imagery to a cloud service, then provides sharing and raw-file export to move the result downstream.

The workstation has become suspiciously small

Scantic takes one of the more computationally interesting parts of a Gaussian-splatting workflow and puts it in a place normally occupied by messages, maps and alarming battery percentages. The iPhone application captures camera imagery and reconstructs an interactive 3D Gaussian splat on the device itself.

Phones have been useful acquisition devices for photogrammetry and spatial capture for years. Scantic’s workflow decision is that the reconstruction also happens locally. The developer states that capture, processing and viewing can operate entirely offline, without an account or cloud processing. That changes the data path substantially. Instead of recording imagery, uploading it, waiting for remote computation and retrieving the result, Scantic keeps the capture in its on-device library and performs reconstruction there. The finished scene can then be inspected in the application’s viewer, before you leave the location. The developer’s website describes processing as taking “a couple of minutes”, actual processing time is also likely to be a particularly slippery benchmark for this class of application because capture size, phone generation and thermal conditions can all matter. Scantic’s own support documentation explicitly says newer iPhones process scans noticeably faster.

Capture first, orbit politely

The acquisition procedure is intentionally simple. The user starts a scan, points the iPhone at a room, object or other environment and moves around the subject. Scantic guides the capture while recording the imagery needed for reconstruction. The developer recommends moving slowly, covering the subject from multiple sides and providing plenty of light. Its support documentation notes that dim environments can produce blurrier scans. These are familiar acquisition constraints, Gaussian splats may look novel in a viewer, but the camera still appreciates photons and coverage.

After capture, Scantic offers a choice between faster processing and its finest-detail mode. The capture remains in the application’s library, so the user can process it again rather than having to repeat acquisition each time.

That separation between acquisition and processing is useful in practical field work. A usable capture can be retained and reconstructed again later, which makes the phone more than a one-shot scanner interface. Scantic does not, however, document the individual reconstruction parameters exposed by those processing choices, so it would be inappropriate to assume that users have the detailed optimisation controls found in research or desktop 3DGS implementations.

Training stays on the device

Local reconstruction is the consequential part – the complete capture-to-scene process runs on the iPhone and does not require an internet connection. Scans remain in the application’s local library unless the user explicitly chooses to share them.

For productions dealing with unreleased sets, locations, props or visual references, that architecture is potentially more interesting than raw speed. Keeping source capture and reconstruction local removes the automatic upload stage inherent in a cloud reconstruction service. It does not automatically turn a consumer phone into an approved studio security environment, of course, but it does give pipeline and security teams a materially different data flow to assess.

There is also a practical location benefit. A production user can capture and reconstruct where connectivity is poor or nonexistent (Most of Germany, for example….). That may be useful for scouts, reference acquisition and quick spatial documentation where waiting for an upload is inconvenient or impossible.

Gaussian, but not a mesh

Scantic generates a 3D Gaussian splat rather than presenting its reconstruction as a conventional polygonal model. That distinction matters downstream. 3D Gaussian Splatting represents a captured scene using spatially positioned Gaussian primitives whose appearance contributes to the rendered image.

For artists, the practical consequence is that a visually convincing captured view is not automatically a production-ready polygon asset. A Gaussian splat and a textured mesh have different editing, rendering and interchange characteristics. A studio considering Scantic for VFX, virtual production, environment work or reference capture therefore needs to decide whether the desired downstream product is the splat itself or whether further conversion and asset preparation will be required.

That makes Scantic particularly easy to understand as an acquisition and reconstruction endpoint. It can get reality into a 3DGS representation quickly, but it does not claim to replace the subsequent asset-building work required when a pipeline ultimately needs conventional topology, UVs, riggable geometry or another representation.

Export is where production begins

The Scantic support documentation confirms that a user can share a scan through a browser link or export a raw file (PLY). The report also says SPZ export is planned. Scantic’s current public product and support documentation does not list SPZ as an available export option, so it should not be treated as a shipping capability. For production evaluation, the raw export is consequently the hand-off worth testing most carefully.

Sharing takes a different route

Although reconstruction itself is local, Scantic also provides optional network sharing. That distinction is worth making because “fully on-device” describes the capture and reconstruction workflow, not every optional function in the application. When the user creates a share link, Scantic can upload the scan for browser-based viewing. The developer says shared scans are stored on a server in Nuremberg, Germany. Private and password-protected links are encrypted on the iPhone before upload, with the key carried in the URL rather than sent to Scantic’s server.

Share links can be configured to expire after seven days, 30 days, one year or not to expire, and the user can disable an active link. A download switch controls whether the share page exposes a convenient download button. Scantic explicitly warns that this is not copy protection because the scan data necessarily reaches the recipient’s browser for display.

The App Store lists Scantic as an iPhone application requiring iOS 17.0 or later and an A12 Bionic chip or later. The supported-device list therefore reaches back to the iPhone XS and XR generation.

Free entry, unclear commercial tiers

The US App Store currently lists Scantic as free. The base application includes a standard-quality export and that higher-quality export can be purchased either through a €29.99 annual subscription or for €4.99 per individual scan.

https://scantic.app/
https://apps.apple.com/us/app/scantic/id6795293934
https://scantic.app/support/
https://github.com/graphdeco-inria/gaussian-splatting