Qualcomm Patents Compression That Keeps Virtual Faces Sharp During Live Video Calls

By Patentlyze Team · Updated Jul 21, 2026

When you're on a video call as a 3D avatar, your face has to move in real time across a live connection. Qualcomm's new patent tackles the surprisingly tricky problem of sending those facial expressions efficiently without blowing up your bandwidth.

What Qualcomm's AR avatar compression actually does

Imagine you're on a video call, but instead of your actual face, both sides see 3D animated avatars. Your avatar has to mirror your expressions, live, as you talk. The problem is that a 3D face is made up of thousands of tiny points that all have to shift position every time you raise an eyebrow or smile.

Sending all those point positions constantly over a network connection takes a lot of data. Qualcomm's patent describes a system that compresses those facial movements before sending them, much like how a ZIP file shrinks a document by cutting out redundancy, so only the changes from a base face shape get transmitted rather than the entire face from scratch each time.

On the receiving end, your device unpacks those compressed changes and rebuilds the facial expression locally. The goal is to make AR avatar calls feel smooth and responsive even on connections that aren't particularly fast.

How blendshapes get encoded and sent over a network

The technical piece at the center of this patent is something called a blendshape. In 3D animation, a blendshape is a snapshot of a face in a specific expression, like a smile or a raised eyebrow, defined by how far each point on the face mesh has moved from its neutral position. Video games and animated films use blendshapes constantly.

Qualcomm's system starts with a base mesh, a neutral 3D face shape that both sender and receiver already have. Instead of sending the full face for each expression, it only sends the differences between the base mesh and the new expression. Those differences are then:

• Quantized (rounded to compact numerical values to reduce file size)
• Normalized (scaled to a standard range so they compress predictably)
• Paired with indices that tell the receiver which specific face points actually moved (so unchanged points don't need to be sent at all)

A transformation matrix (a compact mathematical description of how the whole face or a region of it has rotated or scaled) is also encoded and sent alongside those difference values. The receiving device decodes all of this and reconstructs the full facial expression locally, without needing the full geometry every frame.

What this means for real-time AR video calling

AR video calling, where participants appear as 3D avatars rather than flat video feeds, is a direction several companies are actively building toward. The technical barrier is real: lifelike 3D faces require far more data than a standard video stream, and that data has to arrive fast enough to stay in sync with a live conversation.

This kind of compression work is the unglamorous plumbing that makes those experiences actually usable. If Qualcomm builds this approach into its chips or codec standards, it could influence how AR communication works across a wide range of devices, from phones to headsets, particularly on mobile networks where bandwidth is limited and variable.

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