Echo-3 turns a few minutes of iPhone video into a full-scale 3D environment

Echo-3 transforms an iPhone capture into a navigable, real-scale 3D environment by combining geometric reconstruction and AI generation.

An apartment, a museum, or an industrial building can now be reconstructed in three dimensions from a simple capture taken with an iPhone. With Echo-3, SpAItial AI combines video footage recorded by the user with a generative model capable of completing the missing parts of an environment. The result takes the form of a navigable 3D space, whose dimensions are intended to correspond to those of the original location.

The system accompanies a new mobile application designed to simplify the capture of real-world locations. SpAItial also offers exports that can be used in 3D software, as well as API access for enterprise users. The envisioned applications range from real estate visualization to visual effects, simulation, and robotics. Reconstructing what the camera saw, generating what it missed

3D reconstruction methods generally seek to recover the geometry and appearance of an environment from images taken from multiple angles. However, a handheld capture inevitably presents gaps: some walls are only partially visible, objects obstruct others, and certain surfaces are never filmed.

Echo-3 addresses this difficulty by combining reconstruction and generation.

The model relies on the information present in the capture to establish the structure of the scene, then uses the visual knowledge acquired during its training to complete the insufficiently observed regions.

This approach differs from a reconstruction based exclusively on measured data. A surface absent from the recording can be generated to ensure the visual continuity of the environment.

The objective is to obtain a consistent space in which walls, furniture, and objects maintain their positions when changing viewpoints.

This capability also presents a fundamental limitation: a part generated by AI is not necessarily an accurate representation of a genuinely hidden part of the building. For uses requiring geometric or documentary precision, it is therefore necessary to distinguish observed regions from elements completed by the model. Real dimensions rather than a simple depth illusion

SpAItial emphasizes the metric-scale nature of the produced environments. Unlike a visually convincing scene whose scale might be arbitrary, Echo-3 aims to preserve the distances and dimensions of the captured location.

This opens up different uses from those of a classic generative image. A full-scale reconstruction can be used to explore the layout of a home, understand the volumes of a commercial space, or prepare a simulation environment.

The platform also targets spaces more complex than a single room: houses, public facilities, laboratories, factories, equipment, and multi-story buildings. Outdoor scenes are also among the presented examples.

The demonstrations published by SpAItial allow users to explore several categories of reconstructed spaces.

However, the company does not provide independent dimensional error measurements—expressed, for instance, in centimeters—in its announcement. Compatibility with metrology, regulatory architectural surveys, or industrial inspection uses cannot therefore be inferred solely from the mention of metric-scale geometry. A capture of a few minutes, without specialized equipment

The SpAItial app on iPhone allows users to record a walkthrough of a location, transmit the data to the service, and retrieve a reconstructed 3D environment.

According to the company, a few minutes are enough to perform a capture, without setting up a tripod or using a specialized camera.

On iPhone Pro models equipped with a LiDAR sensor, the app can display a preview of the point cloud as well as coverage indicators during recording. Compatible devices without LiDAR can also perform a capture, albeit with less preview information.

Reconstruction is carried out after data upload. It is therefore not an entirely local and instantaneous generation on the phone.

The resulting environments can be viewed and shared via a link openable in a browser.

SpAItial notes that an Android app is in development, without announcing a release date. Exports in Gaussian Splatting, meshes, and depth maps

One of the technical advantages of Echo-3 lies in the ability to retrieve the scene in several 3D representations.

SpAItial notably mentions SPZ and SOG formats, used to store scenes in Gaussian Splatting, as well as meshes, surface normals, and depth maps.

Gaussian Splatting represents an environment through a collection of 3D primitives whose visual properties make it possible to render the colors, textures, and appearance of a scene from different viewpoints. This representation is particularly interesting for producing immersive tours and real-time photorealistic renderings.

Geometric meshes and depth data address other needs, including integration into modeling tools, simulation engines, or spatial analysis systems.

The availability of these different representations is an asset for teams wishing to reuse a capture beyond SpAItial’s viewer.

However, it remains necessary to verify which formats are actually accessible depending on the subscription plan and the interface used. The announcement does not detail exhaustively how exports are distributed between the mobile app, the web platform, and the API. A comparison with Polycam, KIRI Engine, and Scaniverse

To illustrate its progress, SpAItial filmed the same environment using Echo-3 alongside three existing 3D capture solutions: Polycam, KIRI Engine, and Niantic Scaniverse.

The company states that it followed similar paths and compared the reconstructions from corresponding viewpoints.

In this demonstration, Echo-3 produces cleaner surfaces and fewer floating fragments—those artifacts that sometimes appear as pieces of geometry suspended in mid-air.

This comparison remains a demonstration produced by SpAItial. No independent benchmark covering an extensive range of scenes, dimensional deviations, or processing times accompanies the published results.

It is therefore not possible to establish an overall ranking among these