Here's the question that embarrasses people new to ML: if you don't have enough data, can't you just... make more? Surprisingly, sort of — and that's data augmentation. NVIDIA's grant US12651480B2, "Data set generation and augmentation for machine learning models" (issued June 9, 2026), is a claim on methods for exactly that — but with a sharper twist than the textbook version.

The textbook version is the easy part: you take your real training examples and produce variations — rotate an image, change its lighting, occlude part of it — or generate entirely synthetic examples that resemble the real distribution. The model then sees a richer, more varied dataset and learns the underlying pattern rather than the quirks of the specific images it was given.

A machine learning model (MLM) may be trained and evaluated. Attribute-based performance metrics may be analyzed to identify attributes for which the MLM is performing below a threshold when each are present in a sample.— U.S. Patent No. 12,651,480 source

What the grant actually claims: targeted, not blind

Read the claims and a more specific mechanism emerges, and it is the interesting part. This is not blind augmentation that throws more variety at a model and hopes. The method is a closed loop driven by where the model is already failing. Per the independent claims, the system obtains the model's predictions, then evaluates them to compute "performance metric values" for a subset of attributes — the specific characteristics of the data on which the model is weak. It then computes how many new samples to make, and the claim is precise about the rule: the quantity "increases with one or more distances between the one or more performance metric values and one or more corresponding threshold values." In plain terms, the worse the model does on some attribute, the more synthetic examples of that attribute it generates.

Those examples come from a generative model. The grant describes applying inputs to "one or more generative MLMs" — including a "compositional generator" that builds samples as combinations of several weak attributes at once — and then "updating parameters" of the original model using the new data. Then it repeats. The patent frames the synthetic data as a second, smaller training set used to refine a model first trained on a larger one, rather than as a wholesale replacement.

The disclosed examples make the idea concrete. One claim enumerates attributes for a person in an image — "an age of the person, an ethnicity of the person, a hair length... whether the person is wearing glasses... whether the person is wearing a mask," lighting and facial-expression conditions, even "a blink rate" and "blink duration." If a face model underperforms whenever, say, glasses and a mask appear together, the system can generate more images with exactly that composition and retrain on them. Another set of claims extends the same logic to time: the metrics can target a "temporal pattern" across "a plurality of video frames" — a "frequency, an amplitude, a velocity, or a duration" of an event — so the generator produces synthetic video sequences that depict the under-covered scenario.

Why the boring step decides the outcome

Because models memorize whatever you let them. Train on too few, too-similar examples and the model aces those and fails everything else — overfitting. Augmentation is the cheapest defense: instead of collecting more real data (slow, expensive, sometimes impossible), you manufacture useful variety from what you have. The grant's contribution is to make that manufacture aimed: measure the gaps, generate to fill them, remeasure, repeat "until one or more criteria are satisfied."

One analogy, then I'll drop it: it's flashcards versus understanding. If a student only ever sees the exact same ten problems, they memorize the answers. But this method goes further than just shuffling the deck — it grades the student first, finds the topics they keep missing, and then prints new flashcards specifically on those topics until the scores come up. Augmentation here rotates the flashcards and reads the report card.

The grant also lists where such a system might live: a "perception system for an autonomous or semi-autonomous machine," a "system implemented using a robot," or one "implemented at least partially in a data center." That is consistent with the computer-vision focus and points at the use cases — driver-monitoring and autonomy perception — where rare combinations of conditions are exactly the dangerous edge cases you cannot easily collect by hand.

The claims also quantify the loop in a way worth dwelling on, because it is what separates this from a blunt "make more data" recipe. The number of new samples is not fixed; it is computed per weak attribute, and one claim describes allocating "a fixed computed number of samples" across multiple weak attributes — a budget split among the model's failure modes in proportion to how badly it is failing each one. Another claim ties the metric directly to "inference accuracy of the at least one MLM for a combination of the attributes," meaning the system can target not just single weak features but specific combinations that trip the model up. The compositional generator named in the claims exists precisely to manufacture those combinations on demand.

The sector point: data is the constraint everyone hits, and methods to stretch it are quietly strategic — which is why a company like NVIDIA, whose chips run the training, also patents the data plumbing that feeds them. The architecture gets the headlines; grants like this one are about the unglamorous step that frequently decides whether the headline model actually works — and this one is specifically about doing that step where the model is weakest, not everywhere at once.