The independent claim of US12675908B2, a patent granted on July 7, 2026 and assigned to Snap Inc., is directed to turning ordinary 2D images into a structured 3D scene that a system can edit. What distinguishes the claimed method is not that it produces 3D geometry, but how the geometry is organized and how the model that produces it was trained. This brief walks the granted claim in plain terms, notes its CPC classification, and places it in the cohort it issued alongside. This is an issued patent, not a pending application.

Claim 1 recites three steps. It receives, by one or more processors, a set of 2D images representing a real-world environment comprising a scene. It then samples a 3D scene representation from a posterior distribution conditioned on a single or multiple sets of images drawn from a distribution different than a training distribution. Finally, a machine-learning model generates the 3D scene representation, and the claim specifies exactly how that representation is structured: it explicitly and separately defines a 3D shape and appearance of a background of the scene and a 3D position, 3D shape and appearance of each object of the scene. The claim closes on the training method, requiring that the model have been trained in an unsupervised approach from a dataset of images and their camera poses. Here is the claim in full:

A method comprising: receiving, by one or more processors, a set of two-dimensional (2D) images representing a real-world environment comprising a scene; sampling a three-dimensional (3D) scene representation from a posterior distribution conditioned on a single or multiple sets of images drawn from a distribution different than a training distribution; and generating, by a machine learning model, the 3D scene representation of the set of 2D images, where the 3D scene representation explicitly and separately defines a 3D shape and appearance of a background of the scene and a 3D position, 3D shape and appearance of each object of the scene depicted in the set of 2D images, where the machine learning model has been trained in an unsupervised approach from a dataset of images and their camera poses.— Estimating 3D scene representations of images, US12675908B2

Reading the claim in plain terms

Two phrases in the claim carry most of its scope. The first is explicitly and separately defines. Many methods that lift images into 3D produce a single fused representation of everything in view. This claim instead requires the model to keep the background and each object apart, with each object assigned its own 3D position, shape, and appearance. That separation is what makes the recovered scene editable rather than merely viewable: an object defined on its own terms can be moved, replaced, or composited against without disturbing the rest of the scene. The supporting disclosure describes how that structure is realized, using per-object latent variables modeled as diagonal Gaussians and object positions chosen among candidate locations, with one network mapping the latents to a 3D representation and a second rendering a 2D image from points to color and density in a NeRF-like fashion.

The second load-bearing phrase is trained in an unsupervised approach from a dataset of images and their camera poses. The claim requires only images and their camera poses for training, and the disclosure is explicit that this means no manually labelled annotations such as depth maps, segmentation masks, or object poses. The sampling step reinforces the same generative framing: the representation is drawn from a posterior distribution, and the claim expressly contemplates input images drawn from a distribution different than the training distribution, so the recited method is aimed at scenes it was not trained on. A dependent claim in the record extends the recovered representation to placing AR or VR virtual elements into video, which connects the reconstruction method to the augmented-reality use it enables. The supporting disclosure also describes a two-stage training procedure, learning object-level structure before scene-level composition, and represents candidate object positions with a categorical one-hot selection, details that sit behind the claim's compact recitation of an unsupervised approach.

Classification and cohort

The record carries CPC classification G06T 7/75, the subclass for determining position or orientation of objects using feature-based methods, together with G06V 10/82 (visual recognition using neural networks) and G06T 2207/20084 for artificial neural networks applied to image analysis. That pairing places the grant at the meeting point of 3D pose estimation and neural image understanding, which tracks a claim that recovers per-object 3D position with a learned generative model.

US12675908B2 issued within a same-week Snap cohort of granted patents spanning the AR stack. On rendering and motion, US12675840B2 claims late warping to minimize latency of moving objects by time-warping rendered content to an updated pose, and US12675930B2 is directed to a state-space system for pseudorandom animation. Interaction carries two grants: US12675169B2 covers gesture-based shared AR session creation by matching observed to captured motion, and US12675157B2 claims pausing device operation based on facial movement, such as when a user's eyes look away. The optics side appears in US12674986B2, an adjustable display arrangement that switches the virtual field of view of an extended-reality device.

Read together as a matter of coverage, the granted claims describe a Snap AR pipeline in which scene reconstruction, rendering, interaction, and display optics each hold issued patents, and the reconstruction grant reported here is the perception member of that set. The claim is factual coverage as issued: a generative, label-free method that recovers a 3D scene while keeping background and objects separately defined, granted July 7, 2026, and classified under the pose-estimation and neural-image CPC subclasses noted above.