What changes compared with classic vision
A classic detector recognises the classes it was trained for, very quickly. A vision-language model looks at an image and answers a free question: “is anyone without a helmet?”, “is the loading bay door open?”, “describe what happened in this video”.
Nothing needs to be trained to get started. You write the question, test it on real images and adjust the wording.
When it is useful and when it is not
It is useful when situations are varied and hard to list, when one answer every few seconds is enough and when the result is a description or an alert read by a person. Site surveillance, access control, reviewing recordings or visual inventory.
It is no use for a line producing ten parts per second. There the classic detector, trained for that defect, is faster, cheaper and more reliable.
What works best: the two together
The most effective setup combines both. A fast detector watches all the cameras and discards what is not of interest. When it sees movement or something out of the ordinary, it passes those images to the large model, which describes them and decides whether to raise an alert.
That way the large model only works when needed, and a small computer handles many cameras.
What hardware you need
The vision-language model needs memory, and there unified memory gives a lot for little power. The fast detector on many cameras needs a dedicated graphics card. So the starting point depends on the number of cameras.
| Case | Computer | Why | Price |
|---|---|---|---|
| Reviewing recordings and describing scenes | Compact AI 64 | A large model without a dedicated graphics card | €2,940 |
| 4 to 8 cameras with alerts | Compact AI 64 with RTX PRO 2000 | The card runs the detector; unified memory, the large model | Custom |
| Many cameras or high resolution | Compact with RTX PRO 4000 or 4500 | More graphics memory for more video streams | Custom |
| Dozens of cameras and recording | Server with graphics cards | Several graphics cards and drives to store the video | Custom |
Also count the powered network ports for the cameras and a separate drive for the recordings.
Images of people: process them locally
Images in which a person can be recognised are personal data. Processing them on your own machine, keeping only what is necessary for the period the rules allow, and informing with the appropriate sign, makes compliance much simpler.
Related hardware and guides
Frequently asked questions
Does it replace a trained machine vision system?
Not on a fast production line. It complements it: it is good for varied situations, for describing what is happening and for new questions without retraining. The norm is to use both.
How many cameras can one compact handle?
It depends on how often each camera is analysed. With a detector filtering first, a compact with a professional graphics card handles several cameras with alerts within seconds. We size it with your cameras and your case.
Does it work without an internet connection?
Yes. The models run on the machine itself, so it works in a warehouse, on a building site or on board, and the images never leave the installation.
Can I change what it detects without calling anyone?
Yes, that is its advantage: what the model looks for is written in natural language. Changing the question changes what it watches.
What will you use it for?
Tell us which models you want to run, how many people will use it and where it will work. We will propose the compact, with or without a professional graphics card, and its price within 12 working hours.