Samsung
Samsung · AI / ML · Filed Feb 10, 2026 · USPTO published Nov 12, 2026 · This week's release
Samsung Patents an AI Grader That Circles the Flaws in AI-Made Images
Samsung has filed a patent for an AI that grades other AI's pictures, giving each one a quality score and shading the exact spots that look wrong, such as distorted features or blurred regions.
- Publication number
- US 2026/0359962 A1(US20260359962A1)
- Applicant
- SAMSUNG ELECTRONICS CO., LTD.
- Filed · Published
- Feb 10, 2026 · Nov 12, 2026
- Inventors
- Yi WEI, Suren KUMAR, Luowei ZHOU
- Status when we published
- Waiting for an examiner
- Document
- 20 claims
- US classification
- 345/619
- Parent application
- Claims priority from a provisional application 63802911 (filed May 9, 2025)
What Samsung's AI image grader actually does
Every time you type a description into an AI image generator, you are rolling the dice. Sometimes the picture is great. Sometimes a hand is warped, a texture is smeared, or the thing you asked for is missing entirely.
Samsung's patent describes a second AI that acts as a grader. It looks at your written request and the finished picture together, then gives the image a score for things like how well it matches your words or how good it looks. It also highlights the exact spots that look wrong, a bit like a teacher circling mistakes in red pen.
Those scores and highlights can then be used to retrain the image generator, so it makes fewer of those mistakes next time. The goal is pictures that come out right on the first try, with fewer do-overs for you.
… deriving, using a first decoder, an evaluation score of the first image based on the first type token; obtaining, using a second decoder, an artifact region information based on the second type token and the image embedding; and providing the evaluation score and the artifact region information.
Translation: The system calculates a quality score and pinpoints the exact flaws in the picture.
How the grader produces a score and a flaw map
The filing describes a grading model built on a multi-modal large language model (an AI that reads text and looks at images together). It takes in the original prompt, the generated image, and a task instruction naming the grade wanted: text-image match, visual appeal, overall quality, or plausibility.
The model then outputs two special markers called tokens, one for scoring and one for flaw-finding. The first goes to a score decoder, a small neural network that turns it into a number. The second goes to a mask decoder, which pairs it with features from an image encoder to produce a binary mask (a black-and-white map of the flawed pixels).
The claims also cover using the results to improve the generator:
- A reward instruction built from the score and flaw map is used to fine-tune the generator.
- Training adjusts the loss function (the formula that tells an AI how wrong it was) based on the score or flawed regions.
- A later image can have pixels replaced in areas flagged as artifacts.
The grading model itself is fine-tuned with LoRA (a lightweight method that adjusts a small add-on layer instead of the whole model), while both decoders are trained in full.
Why fewer redo prompts matter for AI image tools
You will probably never see this system directly. It works behind the scenes, grading an image generator's output before the generator reaches you. The payoff is fewer pictures with warped details and fewer times you have to retype and resubmit a request. The filing itself points to wasted computing power and repeated back-and-forth with users as costs of weak generators.
It also tackles a cost problem. The description says collecting large amounts of human preference ratings takes significant resources, and that none of the existing quality scores is enough on its own. A single grader that switches between score types with a task instruction, and adapts to new ones with little extra data, would cut that work. The filing lists phones, tablets, TVs, and AR glasses among the devices that could run it.
Samsung's 16th filing we've tracked since July on AI models working together builds on earlier applications like one AI for many edits and a voice assistant picking AI size.
Source: US 2026/0359962 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.
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 Samsung patent news breaks.
Add Patentlyze on Google
One click.
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 · 9 sheets
9 drawing sheets from US 2026/0359962 A1 · click any drawing to enlarge
Show the other 1 sheets
More from Samsung
All 1,629 Samsung patents →
More in AI image & video
All 223 AI image & video patents →
Google
Google Files Patent for an Image AI That Sends Each Photo Piece to a Specialist
Google
Google Patents a Two-Model AI System That Grades Its Own Image Drafts
Microsoft
Microsoft Files Patent for AI Design Tool That Puts Art Where You Ask
Adobe
No comments yet