The marketing name is "portrait lighting" or "relighting"; the filing name is photo relighting using deep neural networks. Google's granted patent US11776095B2, "Photo relighting using deep neural networks and confidence learning," issued October 3, 2023, claims changing the lighting in a photo after it was taken, using a neural network plus a confidence-learning step. Its CPC tags pair image-enhancement (G06T 5/008) with rendering-illumination (G06T 15/506) — the seam where a photo editor meets a 3D renderer.

On the record, this is the moment computational photography crosses into the generative. Earlier pipeline steps — fusion, tone mapping, deghosting — reconstruct a faithful image from real captured frames. Relighting goes further: it synthesizes lighting that was never physically present. The abstract is explicit that the network does not merely brighten pixels; it learns and applies a lighting model, and the claims describe predicting an "original light model indicative of a location of one or more environmental light sources with reference to an object in the input image." The system has to infer where the light was before it can decide where to put light that never existed.

"Apparatus and methods related to applying lighting models to images of objects are provided. A neural network can be trained to apply a lighting model to an input image."— U.S. Patent No. 11,776,095 source

The claims sharpen what "lighting model" means. Claim 16 enumerates the data a model carries: "a color," "an intensity," "an albedo," "a light direction," "a surface normal," or "one or more light sources," with at least one source at "a different location between the original light model and the particular lighting model." That list is essentially the vocabulary of physically based rendering — albedo and surface normals are how a graphics engine describes how a surface reflects light. The patent imports that machinery into a photo-editing pass: estimate the scene's reflectance and geometry from a single photograph, then re-render it under a chosen light. Claim 12 confirms the hard version of the task, "wherein the particular input image of the particular object is a single image" — no multi-shot light stage, just one frame.

Why the confidence-learning detail matters: a relighting model that guesses wrong produces uncanny, obviously-fake results — a nose lit from an impossible angle, a cheek that glows where it should fall into shadow. Claim 1 ties training to "confidence learning that is based on light predictions and prediction confidence values," and claim 4 pushes it down to "a light prediction... for a portion of the input image and a prediction confidence of the lighting prediction for the portion." The model is trained not only to predict the light but to know where its own prediction is trustworthy. That per-region reliability estimate is the difference between a usable feature and an artifact generator, and it is where the claim's specificity concentrates.

The claims also reveal the intended targets and the deployment shape. Claim 6 scopes "an object that diffusely reflects light"; claim 7 names the obvious case, "a face of a person" — relighting is, in practice, a portrait feature. Claim 5 describes training "using a plurality of images of the object" under "a plurality of lighting models," the supervised setup a controlled capture rig produces. And the system can run either way: claim 13 keeps inference on the capturing device, while claim 15 sends "the request to apply the particular lighting model" to "a second computing device" holding the trained network — on-device or in the cloud, the claim covers both.

The training setup the claims describe is what makes the single-image inference possible at all. Claim 17 specifies training "with a dataset comprising a plurality of images comprising one or more objects under a plurality of different lighting conditions," and enumerates those conditions: "lighting provided from different directions," "of varying intensities," "with light sources of different colors," or "with different numbers of light sources." That is the signature of a controlled capture rig — the same subject photographed under a methodical sweep of lights — which gives the network paired examples of how a face looks under known illumination. Having learned that mapping in the lab, the deployed model can run the inference backward on a single uncontrolled photo: estimate the light that was present, then substitute a chosen one. The data collection is the unglamorous, expensive half of the invention, and the claim language quietly admits it.

The deployment flexibility in the claims also signals product intent. Because the network can run on the capturing device (claim 13) or be queried as a service on a "second computing device" (claim 15), Google preserves the option to ship the feature on-device for speed and privacy, or to offload heavier relighting to the cloud — without leaving the protected scope. Claim 9's plain step of "providing the output image using the computing device," and claim 11's option to apply "a plurality of particular lighting models" to produce "a plurality of output images," read like the UI affordances of a real photo editor: show me the result, and let me try several lighting presets. The patent is written around a feature someone intends to ship, not merely a theorem.

Novel, or just renamed? Single-image portrait relighting was an active research subject before this grant, and Google's own research teams published in the area. What the patent claims is the particular combination — a network that predicts an environmental light model from one image and is trained with per-region confidence so it can hedge where it is unsure — packaged for a shipping pipeline. The category, neural image manipulation, is crowded; the line between "enhancement" and "generation" is exactly where the interesting IP now sits, and confidence-aware relighting plants a flag on the generative side of it.

Scope, stated carefully: this is a granted patent to Google on a specific neural-relighting method, within a fast-moving landscape of generative image techniques from many players. It marks where Google's computational-photography pipeline went generative, not a fence around the entire idea of relighting.

Follow the filing, not the gallery. When a phone offers to change the lighting on a portrait after the fact, it is running something like this 2023 grant — a neural network that infers a scene's light sources, reflectance, and surface geometry from a single frame, then synthesizes illumination that was never there, while tracking how confident it is at every patch. Computational photography has become, in part, computational invention.