Here's what actually issued. On November 8, 2022, NVIDIA Corporation was granted US11494976B2, "Neural rendering for inverse graphics generation," with inventors including Sanja Fidler and Antonio Torralba — strong neural-graphics research credentials. The CPC list spans 3D modeling (G06T 17/00), rendering (G06T 15/10), and network training (G06N 3/0454, G06N 3/088).

The mechanism is inverse graphics. Ordinary rendering goes forward: from a 3D scene description to a 2D image. Inverse graphics runs it backward — from a 2D image, recover the underlying 3D shape, materials, and lighting. Neural rendering uses trained networks to make that inversion tractable, learning to predict scene parameters that, when rendered, reproduce the input image. It's how you turn photos into editable 3D content.

“Approaches are presented for training an inverse graphics network. An image synthesis network can generate training data for an inverse graphics network. In turn, the inverse graphics network can teach the synthesis network about the physical three-dimensional (3D) controls.”— U.S. Patent No. 11,494,976 source

The independent claim describes a mutual-training loop that is more specific than "use a neural net for 3D." Claim 1 starts by providing "a two-dimensional image of an object as input to a generative network," which generates "a set of view images of the object represented from different views." So a single photo is expanded by a generative model into multiple synthetic viewpoints of the same object. Those view images, plus "information for the different views," are fed to "an inverse graphics network," which determines "a set of three-dimensional information" for each view. The system then renders, for each view, "a representation of the object using the set of three-dimensional information and the respective view information," compares the rendered representations "against corresponding ground truth training data to determine at least one loss value," and adjusts the inverse graphics network's parameters from that loss. The generative network manufactures the multi-view training data; the inverse graphics network learns 3D from it. This is the mechanism that lets the system learn 3D reconstruction "while requiring little annotation," as the abstract puts it — the multi-view supervision is synthesized rather than hand-labeled.

The reciprocal half — the part the abstract calls the inverse network teaching the synthesis network — lives in the dependent claims and is the distinctive twist. Claim 2 provides "at least a subset of the representations of the object, rendered by the inverse graphics network, as training data to further train the generative network." Claim 3 then trains "the inverse graphics network and the generative network together using a common loss function." So the two networks are not trained in isolation: the inverse network's 3D-aware renderings feed back to improve the generator, and a shared loss couples them. That bidirectional coupling is what makes the generator's synthetic views increasingly 3D-consistent — the inverse network injects "physical three-dimensional controls" the generator otherwise lacks.

The claims also name the generator and the rendering machinery. Claim 4 specifies the generative network "is a style generative adversarial network enabling only camera view-related features to be adjusted for generating the set of view images" — a StyleGAN whose latent space is constrained so that varying it changes the camera viewpoint while holding object identity fixed, which is exactly what you need to manufacture clean multi-view training pairs. Claim 5 uses "a selection matrix to reduce a dimensionality of image features to be included in a latent code," and claim 6 renders the representation from that latent code "using a differentiable renderer" — differentiability being the property that lets the rendering loss backpropagate into the networks at all. Claim 7 notes the latent code "includes camera features for the corresponding view," and claim 8 enumerates the 3D information as "at least one of a shape, texture, lighting, or background." Claim 9 adds that the input image can be annotated with merely "weakly accurate camera information," underscoring the low-annotation selling point.

The claims even spell out the payoff and the deployment context. Claim 16 covers the inference path — "receiving a two-dimensional image" and "generating a three-dimensional representation" using the trained inverse graphics network — so the patent covers both the training scheme and the trained model's use. Claim 15 lists the deployment surfaces: systems for "graphical rendering operations," "simulation operations to test or validate autonomous machine applications," "deep learning operations," edge devices, and data-center or cloud resources. That enumeration ties image-to-3D directly to NVIDIA's businesses — content creation, autonomous-vehicle simulation, and synthetic-data generation.

Strategically this sits exactly where NVIDIA wants to be: the intersection of AI and graphics, where its GPU dominance and its rendering heritage compound. Owning a method that bridges images and 3D through a co-trained generator/inverse-network pair, a view-controllable StyleGAN, and a differentiable renderer — for content creation, simulation, and synthetic-data generation — protects a capability that feeds its hardware and its software stack alike.

On scope, the discipline: granted B2, enforceable, but the claims describe a specific neural-rendering method for inverse graphics — the generator-makes-views, inverse-network-learns-3D loop, the reciprocal co-training under a common loss, the view-controllable StyleGAN, and the differentiable renderer with its latent camera code. They do not cover all 3D reconstruction or all neural rendering, both broad and well-populated fields. Claim 1, and the co-training in claims 2 and 3, are the boundary.

The takeaway: US11494976B2 is NVIDIA patenting the AI-graphics seam again — this time the image-to-3D direction, built on a generator and an inverse-graphics network that train each other through a differentiable renderer — with top-tier research inventors and method-specific claims tied to the rendering pipeline its silicon accelerates.