Qualcomm · Filed Feb 4, 2025 · Published Aug 6, 2026 · verified — real USPTO data

Qualcomm Patents an AI That Copies Motion Style From One Video and Applies It to Another

Qualcomm has filed a patent for a system that can lift the way something moves in one video and apply that motion to a completely different subject. Think of it like copying a dancer's choreography onto a cartoon character, automatically.

Qualcomm Patent: Custom Motion Video Generation via AI — figure from US 2026/0230685 A1
Figure from the official USPTO publication.
See all 11 drawings from this filing ↓
Publication number US 2026/0230685 A1
Applicant QUALCOMM Incorporated
Filing date Feb 4, 2025
Publication date Aug 6, 2026
Inventors Sunghyun PARK, Seokeon CHOI, Sungrack YUN
CPC classification 725/116
Grant likelihood Medium
Examiner ALAM, MUSHFIKH I (Art Unit 2426)
Status Non Final Action Mailed (Jun 9, 2026)
Document 20 claims

How Qualcomm's motion-transfer AI actually works

Imagine you have a short clip of a person dancing, and you want your own photo to move the same way, without hiring an animator or using a complicated editing suite. That's the core idea behind this Qualcomm patent.

The system takes two inputs: a reference video that shows the motion you want to copy, and a reference image of the subject you want to animate. Two AI models work in parallel, one studying the motion in the video, the other focusing on the appearance of your chosen image. A third shared layer merges what both models learned.

The result is an output video where your subject moves the way the reference video does, but looks like the original image. The whole process is designed to run on a single device, which fits squarely into Qualcomm's push to put AI directly on phones and chips rather than in the cloud.

How two AI models share attention to blend motion and appearance

The patent describes a two-model pipeline for what researchers call motion-customized video generation.

  • Model one (the motion reader) takes a reference video that has been processed through a technique called noise inversion (a way of encoding a real video back into the compressed mathematical form that AI video models understand). It also takes a reference image. Together, they produce a set of features capturing how things move in the clip.
  • Model two (the appearance model) takes the same reference image, but instead of the real video, it starts from pure random noise, known as Gaussian noise. This is the standard starting point for AI image and video generators. It produces a parallel set of features focused on what the subject looks like.
  • Shared attention layers sit between the two models. Attention, in AI terms, is a mechanism that lets a model decide which parts of its input to pay most attention to (similar to how your eyes focus on a moving object and blur the background). By sharing these layers, motion information from model one can influence what model two generates.

The output is a video where the subject from the reference image moves according to the motion pattern extracted from the reference video. The patent positions this as a general framework that could apply to faces, bodies, objects, or other content types.

We find one patent like this every day. Get the best of each week in your inbox, free →

What this means for on-device AI video generation

Qualcomm's core business is making chips for phones and other mobile devices, and the company has been investing heavily in running AI models locally on those chips rather than sending data to remote servers. A patent like this fits that strategy: if video generation can be made efficient enough to run on a Snapdragon processor, it becomes a feature differentiator for Android handsets.

For you as a user, on-device motion transfer would mean being able to animate a photo of yourself or a character without uploading anything to a third-party server. That has real privacy implications. It also suggests Qualcomm is positioning its silicon as a platform for creative AI tools, competing on more than just raw processing speed.

Editorial take

This is a real technical approach to a problem that AI video tools have struggled with: keeping a subject looking like themselves while faithfully copying motion from another source. The shared-attention architecture is the kind of engineering detail that separates polished output from the wobbly, identity-shifting results you see in many current tools. Whether Qualcomm ships this as a chip-level feature or licenses the technique, it signals the company is serious about generative video as a hardware differentiator, not just a cloud-side party trick.

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 2026/0230685 A1 · click any drawing to enlarge

Patent filing page

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.