Samsung Patents a Flat Camera Lens That Sends Scrambled Light Directly to AI
Samsung is working on a camera system that replaces the curved glass stack inside your phone with a nearly flat chip-like lens, then trains an AI to decode the strange, scrambled images that lens produces. The goal is object recognition that skips the usual image cleanup steps entirely.
What Samsung's meta lens camera system actually does
You're holding a phone and its camera is trying to figure out what it's looking at: a person, a stop sign, a package on your doorstep. Today that job requires a stack of carefully curved glass lenses to focus light properly, followed by software that cleans up the image before any AI recognition can even begin.
Samsung's patent describes a different approach. Instead of curved glass, it uses a meta lens, a nearly flat surface covered in microscopic structures (tiny pillars or pins, each a different shape or height) that bend and scramble light in a controlled way. The image the sensor captures looks nothing like a normal photo. An AI then reads that coded image directly and identifies what's in the scene, no cleanup step required.
The clever part is how Samsung trained that AI. Rather than gathering thousands of meta-lens photos with labels, engineers fed normal photos into a computer model of how the meta lens distorts light, generated simulated scrambled images, and trained the AI to recognize objects in those. That sidesteps an enormous data-collection problem.
a meta lens comprising a plurality of pillars or a plurality of pins provided at a surface of the meta lens and having at least one different shapes, different heights, or different widths …
Translation: A flat lens covered in tiny pillars or pins that scrambles incoming light.
How the AI learns to read light Samsung's lens scrambles
A meta lens is a flat optical surface etched with arrays of tiny structures, here described as pillars or pins, each with slightly different dimensions. Those structures bend incoming light waves in ways that a conventional curved lens would not, encoding depth and phase information directly into the raw sensor data. The image sensor captures this encoded signal and passes it on as a coded image, which looks distorted or scrambled to the human eye but carries structured information the AI can interpret.
The AI model at the core of the system is a neural network trained to take a coded image as input and output a label: a classification result identifying what object is present. The training pipeline is the key innovation. Engineers take ordinary RGB images (standard color photos) and run them through a software model that simulates the optical behavior of the meta lens, producing fake coded images that look like what the real hardware would produce.
The neural network is then trained on those simulated coded images, with the correct label taken from the original RGB photo. During training, the system minimizes a loss value (a numerical score measuring how wrong the AI's guess is) that includes a similarity metric between the original photo and the simulated version, keeping the simulation honest.
The result is an AI that can work directly with the unusual output of a flat meta lens without needing a separate image-reconstruction step, potentially reducing both the physical thickness of the camera module and the computing overhead per recognition task.
… input the coded image into an artificial intelligence model, and obtain a label indicating a perception result of an object through inference using the artificial intelligence model …
Translation: The AI analyzes the scrambled image directly to identify the object without forming a normal photo first.
What this could mean for thinner, faster Samsung cameras
For you as a consumer, the most tangible payoff would be thinner devices. The multi-element glass lens stacks in modern phones are one of the main reasons camera bumps keep growing. A flat meta lens that offloads optical work to AI could, in theory, shrink that bump or free up internal space for a larger battery.
There is also a speed angle. Skipping the image-reconstruction step that current computational photography pipelines require means the device could reach a recognition result faster and with less processing power, which matters for always-on features like object detection in augmented reality or real-time scene labeling. Samsung's ongoing filings in computational optics suggest this is a longer-term platform investment rather than something arriving in the next Galaxy handset.
This is the 109th Samsung filing we've tracked in our camera sensor push watchlist since May, following work on sharpening low-res photos and picking the steadiest dark shot.
Samsung is describing a camera that can be physically thinner because its lens works differently, and software trained on simulated distorted images figures out what the camera is actually looking at. For someone tired of a phone that no longer fits comfortably in a pocket because of its camera bump, that is the promise on the table.
The catch is that this is a patent, not a product. The person who would benefit, the one who wants a slim phone without sacrificing a good camera, is not holding that phone yet.
If this technology matures, the moment they would notice is unremarkable in the best way: they pick up a phone, it is thin, the camera works, and nothing about the experience feels like a compromise. That invisible outcome is exactly what makes the underlying engineering matter.
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
16 drawing sheets from US 2026/0270579 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →