Edge Video AI

Edge Research / Our model

A vision language model for live video.

Our vision language model (VLM) watches a live stream, reads what's on screen and what's said, and turns each moment into signals: IAB categories, scene keywords and brands seen on screen.

Diagram: the VLM on top watches the stream, follows every moment and knows the ad break; a language model underneath helps it reason
Perception on top, language underneath. Frame from Edge's explainer film.

What it does

Watch. Read. Signal.

Watch

It follows the stream continuously: the picture, what is said, and the text and logos on screen. It knows when the ad break starts.

Read

Familiar scenes are recognised from a local vocabulary. The model is asked only when a scene is new, and every answer must fit a fixed shape.

Signal

Each moment becomes IAB categories, scene keywords and brands seen on screen, in shapes that ad systems and software already understand.

Two ways to the screen

On a stream, no TV software at all. On a TV, Edge ACR.

Stream path · no ACR

The model reads the live stream itself

Where we can see the stream, the model reads it directly at the source, once per channel. Nothing runs on the viewer's screen and no automatic content recognition (ACR) is needed. The signals go to the systems that already serve the stream.

Report 02: Live enrichment

TV path · ACR

Edge ACR on the TV

Edge ACR: our own content recognition on the TV, talking to the VLM. Works on apps, HDMI and free-to-air. The picture never leaves the set. One read is 72 bytes, sent about every five seconds while a listed programme is on.

Report 01: Edge ACR

Design rules

Facts come from the picture.

The VLM owns the facts

What is on screen is decided by the vision language model looking at it. The language model underneath helps reason and build; it does not get to invent the scene.

Fixed shapes, checked answers

Answers must match a strict shape and an allowed vocabulary, or they are dropped.

Empty beats wrong

If there is nothing reliable to say, the answer is empty and the systems downstream carry on as before.

Cost follows channels

Each channel is read once at the source, so more viewers mean more lookups, not more model calls.

Where it runs

Any live stream. And on TVs, any input.

HDMI (streaming boxes, consoles, cable and satellite), free-to-air (antenna TV) and apps (the TV's built-in apps)
On a TV, recognition works from the rendered picture, so the input does not matter.

A short public answer

What a vision language model for live video is.

A shorter public-facing answer, written for this question, is at A vision language model for live video.