Get the newsletter
Qualcomm · AI / ML · Filed Dec 16, 2025 · USPTO published Nov 19, 2026 · This week's release

Qualcomm Files Patent for AI That Links Far-Apart Pieces of a Photo

Qualcomm describes a way to organize the pieces of a photo into a layered tree, so far-apart parts can still influence each other without checking every pair.

A photo broken into squares can lose the link between distant areas when lines of pixels are read one by one. Drawing from patent filing US 2026/0368548 A1.
A photo broken into squares can lose the link between distant areas when lines of pixels are read one by one.All 16 drawings ↓
Publication number
US 2026/0368548 A1(US20260368548A1)
Applicant
QUALCOMM Incorporated
Filed · Published
Dec 16, 2025 · Nov 19, 2026
Inventors
Evgenii EGOROV, Hanno ACKERMANN, Markus NAGEL, Hong CAI
Status when we published
Waiting for an examiner
Document
27 claims
US classification
706/61
Parent application
Claims priority from a provisional application 63806157 (filed May 15, 2025)

What Qualcomm's tree-shaped AI patent actually does

Some AI models handle a photo by flattening it into one long line of pieces and reading them in order, like words in a sentence. The catch: two pieces that influence each other can land far apart in that line, and the connection gets lost.

Qualcomm's filing adds a tree on top of that line. Small groups of pieces feed into a group node, groups of groups feed into a bigger node, all the way up to one top node. Information travels up the tree and then back down, so the first piece in the line can still affect the last one.

The other way to catch those links is to compare every piece with every other piece, which gets expensive fast. The tree aims to get long-range context without that cost.

From the filing · CLAIM 1
… effectively matrix multiply an inverse of the tree matrix with the sequence of tensors to generate state variables according to a state space model …

Translation: It combines hierarchical data structures with math models to track relationships across the image.

How the tree carries context up and back down

The system starts by scanning an input, such as an image or a 3D point cloud (a cloud of measured points in space), into a sequence of tensors (bundles of numbers describing each piece). Flattening like this loses relationships between pieces that end up far apart in the sequence.

To restore them, the filing builds a tree with at least two levels. The bottom level holds the pieces. Higher levels hold either virtual nodes (stand-in nodes that summarize a group) or existing pieces promoted to parent roles. Each node carries four things:

  • A self-term (how it relates to itself)
  • A child-to-parent term
  • A parent-to-child term
  • Its own input features

These terms form a tree matrix, a grid of numbers that is mostly empty (sparse). The system multiplies the inverse of that grid with the sequence to get the state variables (the model's running summary of what it has seen). The inverse is dense, so every piece can influence every other, but it is applied in two passes: leaves to root, then root back to leaves.

The result can feed another machine learning model, and the system then acts, for example by detecting objects, classifying them, or recognizing faces.

Why cheaper long-range context matters for phones

If your phone tags objects in your photos or recognizes faces, models like this sit behind it. The filing names those tasks, plus generating text replies, and describes running on devices such as a smartphone with a dedicated AI chip (an NPU).

The appeal is cost. The filing says comparing every piece with every other piece grows expensive quickly, and it describes a coarse version of the tree that works on groups of pieces to cut computing and memory demands. The patent gives no speed or accuracy numbers, so how much it saves is an open question. Qualcomm makes phone chips, so efficiency gains like this are a natural interest.

This is the 12th Qualcomm filing we've tracked in our AI image and video coverage since May, adding to earlier work like one on shrinking AI generators for phones and one on giving photo editors time to think.

Source: US 2026/0368548 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 Qualcomm patent news breaks.

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 sheets

16 drawing sheets from US 2026/0368548 A1 · click any drawing to enlarge

Why flattening an image into a line misses distant connections, drawing sheet 1 from US 2026/0368548 A1
Sheet 1 Why flattening an image into a line misses distant connections
Math behind standard sequential AI tracking, drawing sheet 2 from US 2026/0368548 A1
Sheet 2 Math behind standard sequential AI tracking
Charting only direct family connections to save memory, drawing sheet 3 from US 2026/0368548 A1
Sheet 3 Charting only direct family connections to save memory
How the AI connects every part of an image, drawing sheet 4 from US 2026/0368548 A1
Sheet 4 How the AI connects every part of an image
Gathering image tiles into larger blocks to preserve context, drawing sheet 5 from US 2026/0368548 A1
Sheet 5 Gathering image tiles into larger blocks to preserve context
Alternative column and row groupings for image pieces, drawing sheet 6 from US 2026/0368548 A1
Sheet 6 Alternative column and row groupings for image pieces
Building a family tree over flattened image chunks, drawing sheet 7 from US 2026/0368548 A1
Sheet 7 Building a family tree over flattened image chunks
Linking a full sixteen-patch photo through shared parent groups, drawing sheet 8 from US 2026/0368548 A1
Sheet 8 Linking a full sixteen-patch photo through shared parent groups
Show the other 8 sheets
Building a tiered hierarchy directly from image elements, drawing sheet 9 from US 2026/0368548 A1
Sheet 9 Building a tiered hierarchy directly from image elements
Mapping out connections between distant parts of an image, drawing sheet 10 from US 2026/0368548 A1
Sheet 10 Mapping out connections between distant parts of an image
Mathematical steps for solving connections across branches, drawing sheet 11 from US 2026/0368548 A1
Sheet 11 Mathematical steps for solving connections across branches
Gathering details upward from pixels to the main image, drawing sheet 12 from US 2026/0368548 A1
Sheet 12 Gathering details upward from pixels to the main image
Passing context back down to individual image pieces, drawing sheet 13 from US 2026/0368548 A1
Sheet 13 Passing context back down to individual image pieces
Tuning the network during training by working backward, drawing sheet 14 from US 2026/0368548 A1
Sheet 14 Tuning the network during training by working backward
How the system scans an image and takes action, drawing sheet 15 from US 2026/0368548 A1
Sheet 15 How the system scans an image and takes action
Device chips and memory that run the vision AI, drawing sheet 16 from US 2026/0368548 A1
Sheet 16 Device chips and memory that run the vision AI
Patent filing page

More in AI image & video

All 226 AI image & video patents →
Reader comments

No comments yet

Start the discussion

Real name or a handle, either is fine. Comments are read by a person before they appear, so allow a little time. Keep it about the filing.