Edge Research
Research
We study how machines can watch live video and understand it in the moment: what's on screen, what's said, and what it means for the people watching. Our reports describe the systems behind our vision language model (VLM) in enough detail to judge them, with every number scoped to how it was measured.
A vision language model for live video
How it watches a stream, what one read returns, and the two ways it reaches a screen: straight from the stream, or through Edge ACR on a TV.
Perception
Reading live streams: frames, speech and the text in the picture, turned into a standard vocabulary.
Recognition
Edge ACR: our own content recognition on the TV, talking to the VLM, so the set can name the moment without sending the picture.
Systems
Architectures that keep facts in the perception layer, and cost tied to channels rather than screens.
Technical reports
Edge Research technical reports are not peer reviewed.Edge ACR on the TV
Our own content recognition on the television, talking to the VLM. Works on apps, HDMI and free-to-air. The picture never leaves the set.
Live enrichment: from frames to IAB categories
How a live channel becomes categories, keywords and brands seen on screen, and why the expensive model only runs when the scene is new.
The VLM above the LLM: a layered architecture for live video
Perception on top, language underneath. Why we put the vision language model first, and what the language model is allowed to do.
Matching ads to moments
Context added to the existing ad request at the break, and a brand's own ad recognised to the second on the TV.
Instructions as programs for live video
How a typed instruction, an answer shape and a validator become an app that runs over a live stream.
How we report
Numbers with their scope, or no numbers.
Every figure in our reports says what was measured, where, when and how. Development measurements are labelled as such. Arithmetic is shown as arithmetic. Where a result belongs partly to a partner, we leave it out until it is cleared.
Each report lists the published work it builds on, with links, so readers can follow the ideas back to their sources. Each is also available as a PDF.