Get the newsletter
Sony · AI / ML · Filed Sep 12, 2025 · USPTO published Aug 20, 2026 · Aug 20 release

Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models

Generating enough training data for AI is one of the quieter bottlenecks in building image recognition systems. Sony's new patent takes a fresh approach by fusing two separate statistical models to guide what an AI produces.

Modified face images displaying distinct changes in pose, age, facial expression, and added eyeglasses. Drawing from patent filing US 2026/0244920 A1.
Modified face images displaying distinct changes in pose, age, facial expression, and added eyeglasses.All 25 drawings ↓
Publication number
US 2026/0244920 A1(US20260244920A1)
Applicant
Sony Semiconductor Solutions Corporation
Filed · Published
Sep 12, 2025 · Aug 20, 2026
Inventors
Namiki HAGIO, Tomoki UEDA
Status when we published
Waiting for an examiner
Document
13 claims
US classification
706/15
Parent application
National Stage Entry of PCT/JP2024/007728 (filed Mar 1, 2024)(PCTJP2024007728)

How Sony's dual-distribution AI generator works

Ever tried to teach a child to draw by showing them only a handful of pictures? AI systems face the same problem: they need enormous amounts of varied examples to learn from, and creating those examples by hand is slow.

Sony's patent describes a system that makes an AI generator much more efficient at producing new images (like face photos) by giving it two sources of guidance at once. One source looks at an image you've already given it and tries to reconstruct what that image "should" look like statistically. The other source pulls from a broader library of training data the AI has already learned from. By blending these two guides, the system produces a richer, more varied set of outputs without needing as many starting images.

The practical idea is that you feed in one piece of data, the system consults both statistical guides, and the AI generator uses that combined signal to produce many new, realistic variations. For engineers building face-recognition or image-analysis tools, that means less manual data collection and faster iteration.

From the filing · CLAIM 1
generates a generation probability distribution used for generation of a latent variable for generating output data in a generation model of the output data that outputs the output data with respect to an input of input data …

Translation: The system calculates the mathematical odds needed to create hidden variables for new data output.

How the two probability distributions shape the latent variable

At the center of this patent is a generative model (an AI system designed to create new data, not just classify existing data). The key challenge it addresses: how do you steer that generator toward outputs that are both faithful to a specific input and varied enough to be useful as training material?

Sony's system solves this by constructing what the patent calls a generation probability distribution, a statistical map the AI uses to decide what to generate. That map is built from two components:

  • First probability distribution: derived by encoding the input image and estimating what range of outputs could plausibly reconstruct it (this is similar to how a variational autoencoder works, where an image is compressed into a statistical description rather than a fixed value).
  • Second probability distribution: derived from the broader pool of learning data the generative model was trained on, reflecting what the model "knows" about the world.

The system merges these two distributions into a single generation distribution, then samples a latent variable (think of it as a random seed drawn from that merged map) from it. The generative model takes that seed and produces output data, such as a face image.

The effect is that each output is anchored to the input but influenced by the model's wider knowledge, producing diverse yet coherent variations efficiently.

From the filing · THE ABSTRACT
The present technology can be applied to, for example, a data generation device that generates data such as a face image.

Translation: This method can be used by AI tools to create realistic digital portraits of human faces.

What this means for AI image generation at scale

For teams building AI systems that depend on large, labeled image datasets, data scarcity is a real constraint. Generating synthetic training data is an established workaround, but naive generation often produces outputs that cluster too tightly around a few examples or drift too far from the desired distribution. Sony's dual-distribution approach directly addresses that tension by keeping outputs tethered to real input data while pulling in variety from learned patterns.

The patent names face images as an explicit use case, which connects directly to Sony's semiconductor business and its image sensors used in cameras and phones. Generating controlled, varied face data has applications in camera autofocus tuning, facial recognition training, and image quality testing. Researchers and developers working in computer vision and Big Tech patent news around synthetic data generation will recognize this as part of a broader effort by sensor and chip companies to own the full pipeline from image capture to AI training.

Sony's 343rd filing we've tracked in our Sony coverage since May joins a run that includes a dual-mode pixel design and glare-based light sourcing.

Source: US 2026/0244920 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 Sony patent news breaks.

There are more where this came from

We read the new patents from Big Tech, carmakers and more and send you the ones worth knowing. Plain English, free, every week.

The drawings · 25 sheets

25 drawing sheets from US 2026/0244920 A1 · click any drawing to enlarge

System architecture for generating training datasets, drawing sheet 1 from US 2026/0244920 A1
Sheet 1 System architecture for generating training datasets
Hierarchical structure of encoder and decoder networks, drawing sheet 2 from US 2026/0244920 A1
Sheet 2 Hierarchical structure of encoder and decoder networks
Convolutional architecture of a residual network block, drawing sheet 3 from US 2026/0244920 A1
Sheet 3 Convolutional architecture of a residual network block
Internal structure of a top-down network block, drawing sheet 4 from US 2026/0244920 A1
Sheet 4 Internal structure of a top-down network block
Conceptual steps for controlling image probability distributions, drawing sheet 5 from US 2026/0244920 A1
Sheet 5 Conceptual steps for controlling image probability distributions
Mathematical formulas for latent variable generation, drawing sheet 6 from US 2026/0244920 A1
Sheet 6 Mathematical formulas for latent variable generation
Visual effect of varying interpolation weights on images, drawing sheet 7 from US 2026/0244920 A1
Sheet 7 Visual effect of varying interpolation weights on images
Reconstruction quality across different network layers, drawing sheet 8 from US 2026/0244920 A1
Sheet 8 Reconstruction quality across different network layers
Show the other 17 sheets
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 9
Sheet 9
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 10
Sheet 10
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 11
Sheet 11
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 12
Sheet 12
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 13
Sheet 13
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 14
Sheet 14
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 15
Sheet 15
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 16
Sheet 16
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 17
Sheet 17
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 18
Sheet 18
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 19
Sheet 19
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 20
Sheet 20
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 21
Sheet 21
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 22
Sheet 22
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 23
Sheet 23
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 24
Sheet 24
Sony Patents a Faster Way to Generate AI Training Data From Two Probability Models patent drawing, sheet 25
Sheet 25
Patent filing page