Nvidia Patents an AI System That Blends Photos Based on How Certain It Is About Each Object
Nvidia has filed a patent for a processor that uses neural networks to decide how to blend two or more photos together, weighting each region by how confident the AI is about what it sees. The approach could change how devices handle tricky shots where one image is clearer than another in specific areas.
What Nvidia's confidence-based image blending actually does
You're taking a photo in mixed lighting, and your camera captures two shots in quick succession: one exposed for the bright window, one for the dark room. The tricky part is combining them so both the sunny view and the shadowy foreground look natural.
Nvidia's patent describes a processor that uses AI to do exactly that kind of blending, but with a twist. Instead of just averaging the two images pixel by pixel, the system uses a neural network to assign a confidence score to different objects or regions in each image. If the AI is very sure about how a face looks in one image but less sure in the other, it leans on the high-confidence version for that area.
The result is a blend driven by certainty, not just arithmetic. Areas where the AI is confident get more weight; areas where it's unsure get blended more carefully. It's a way to let AI judgment, not fixed formulas, decide which parts of which photo to trust.
one or more circuits to use one or more neural networks to blend two or more images based, at least in part, on one or more confidence values below a first threshold value.
Translation: Specialized hardware uses AI to combine multiple photos based on how certain the system is about what it sees.
How the neural network scores and merges image regions
The patent describes a processor with dedicated circuits that runs one or more neural networks to combine two or more images. The core mechanism is the confidence value: a score the neural network assigns to objects or regions within each image, reflecting how reliably it has identified or understood that part of the scene.
When the confidence score for a region falls below a set threshold, the system adjusts how that region contributes to the final blended image. In plain terms, low-confidence areas get treated differently than high-confidence ones, so the blend is not uniform across the frame.
The claim is intentionally broad. It covers:
- Any processor with circuits running neural networks for blending
- Any confidence value below a threshold driving the blend
- Two or more images as input
The patent does not specify a particular type of neural network, a specific kind of image (photo, video frame, synthetic render), or a particular display context, which keeps the scope wide. The threshold value acts as the decision boundary: above it, the AI trusts the data; below it, the system compensates, possibly drawing more from the other image in that region.
Apparatuses, systems, and techniques to blend two or more images based on confidence values of objects within said two or more images.
Translation: The technology combines different pictures using the computer's certainty level regarding specific objects inside them.
What this means for AI cameras and image processing
For anyone using a camera on a phone, laptop, or AR headset, this kind of AI blending is already happening behind the scenes in computational photography. What this patent stakes out is the specific idea of using a neural network's own confidence output as the blending control signal, rather than a fixed rule or a separate depth map. That framing could apply to night-mode photography, video conferencing background processing, or even compositing in creative tools.
The claim as written is quite broad, covering any processor that does this, in any context. If granted in its current form, it would give Nvidia a meaningful position over a technique that sits at the intersection of AI inference and image processing, two areas where Nvidia's run of image-generation filings continues to grow.
Nvidia's 24th filing we've tracked in AI photo editing since May also follows work on shot by shot video compression and fixing lighting flicker in video.
Claim 1 covers any processor that uses any neural network to score how confidently it has identified objects in images, then uses those scores to blend the images together. That description fits an enormous range of everyday products: a phone combining two exposures, a medical scanner merging different scan types, a game engine layering rendered scenes.
If granted, this patent could give Nvidia legal standing against a wide swath of image-processing software and hardware, not just products that copy its specific approach. Nothing in the claim restricts the industry, the type of network, or even what the final blended image is used for.
The practical question is whether any prior research already described confidence-weighted blending in similar terms, which is the hurdle that will determine whether this claim survives examination.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Nvidia patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
61 drawing sheets from US 2026/0289736 A1 · click any drawing to enlarge
Show the other 52 sheets
Sheets 61 to 61 are in the full USPTO filing.
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in