Adobe Patents a Two-Pass AI System for Cutting Out Objects in Phone Photos
Selecting a person or object in a photo on your phone is still surprisingly rough around the edges. Adobe has filed a patent for a neural network that takes two passes at the problem, starting with a quick rough cut and then sharpening it using deeper image data.
What Adobe's mobile photo-selection system actually does
You're trying to cut out a friend from a photo on your phone so you can drop them onto a different background. You tap "select subject," and the outline the app draws is close but not quite right, clipping hair or missing part of a shoulder.
Adobe's patent describes a system designed to fix that two-step. A neural network first scans the image for simple, surface-level details, like edges and color patches, and draws an initial rough selection. Then a second pass takes that rough outline and checks it against deeper, more abstract features the network has also pulled from the image, things like understanding that a shape is probably an arm and not a shadow, and tightens the result.
The key engineering trick is doing all of this on mobile hardware, which has far less processing power than a desktop or cloud server. The design is specifically built to keep the model lightweight enough to run on a phone without sending your photo to the internet.
… a first segmentation head that generates an initial segmentation mask for the digital visual media item by decoding low-level features of the digital visual media item; and a second segmentation head that generates a refined segmentation mask for the digital visual media item by decoding the initial segmentation mask and high-level features of the digital visual media item …
Translation: The AI uses a two-step process to first make a rough cutout and then clean up the edges.
How the two segmentation heads refine a selection
The patent describes a segmentation refinement neural network, meaning a model that figures out which pixels in an image belong to a subject and which belong to the background.
The network has one encoder (the part that reads the image and extracts patterns) and two segmentation heads (output modules that each produce a selection mask, essentially a pixel map marking what's inside the selection and what's outside):
- First segmentation head: Works from low-level features, which are raw, early-stage patterns like edges, colors, and textures. It produces an initial rough mask quickly.
- Second segmentation head: Takes that rough mask and combines it with high-level features, which are more abstract representations the encoder has also extracted, closer to semantic understanding of what objects are present. It outputs a refined, tighter mask.
The architecture is described as transformer-based. Transformer models (the same family of design behind large language models) are good at relating distant parts of an image to each other, which helps with tricky selections like hair against a busy background. The patent's claim is specifically about running this two-head refinement pipeline on a mobile device, implying a compressed model design that fits within a phone's memory and processor limits.
… the disclosed systems generate, utilizing a neural network decoder, an initial segmentation mask of the digital visual media item from the low-level features. Moreover, the disclosed systems generate, utilizing the neural network decoder, a refined segmentation mask of the digital visual media item from the initial segmentation mask and the high-level features.
Translation: The software creates a preliminary selection and then improves it using higher-level image details.
What this means for mobile photo editing apps
For anyone who edits photos on a phone, cleaner automatic selections directly improve the quality of background swaps, portrait effects, and object removals, without requiring manual touch-up. If Adobe ships something based on this, it would likely surface inside Photoshop on mobile or the Adobe Express app as a more accurate "select subject" feature.
Adobe keeps filing on on-device AI for creative tools, and that pattern matters for privacy as well as speed. Running the model locally means your photos don't have to leave your device for the selection to work. That's a meaningful difference for users cautious about cloud processing of personal images.
Adobe's 38th filing we've tracked in the AI photo editing race since May adds to earlier work like one that erases photo shadows and one that animates still photos.
Claim 1 of this patent covers any system that uses an encoder to pull out features from an image, runs a first pass to produce a rough cutout using simpler features, and then refines that cutout using more complex ones. Nothing in the claim requires a specific architecture, a specific type of image, or even a specific device. That scope is wide.
If granted as written, this claim would give Adobe a legal stake in a large category of two-pass image cutting tools running on mobile hardware, not just the particular product described in the abstract.
The real hurdle is convincing a patent examiner that this specific combination of rough-then-refined masking is a new enough idea to deserve exclusive rights. The usefulness is clear; the novelty question is where this will be decided.
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
11 drawing sheets from US 2026/0278804 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in