Frequently asked questions

Software that teaches a computer to recognise things in images, and that runs on your own PC instead of in the cloud. In industry this is also called computer vision or machine vision, and software that runs on your own hardware is often called on-premise. Devimation Studio is local AI vision software: you label your own images, train a model and test it, all on your own Windows PC with an NVIDIA GPU.

Devimation Studio trains dedicated detection models instead: Faster R-CNN, Mask R-CNN and Keypoint R-CNN. A VLM describes images and answers questions about them in words, and usually needs a lot of computing power. A detection model trained on your own images is built for one job: finding exactly your parts and defects, quickly and the same way every time, on a normal NVIDIA GPU.

They stay on your own PC. Labeling, training and testing all happen there, and your images, labels and trained models are never uploaded or sent to us. The app only goes online to check your licence and, once per model size, to download the base model that training starts from.

That is exactly what Devimation Studio is built for. Sign in once, and from then on you can label and train without an internet connection. The app only asks you to sign in again when your licence expires or when it has been offline for more than a few days.

Cloud platforms run on someone else's servers, so your images have to be uploaded first, and training often costs extra per run or per image. Devimation Studio runs entirely on your own computer: nothing is uploaded, nothing leaves your company, and you can train as often as you like for the same licence price.

It is a good fit if you want to keep your images on your own PC. Devimation Studio does labeling and training entirely locally, without uploading anything.

You don't. The app tells you what to do at every step, and built-in tours explain what the scores mean and whether your model is good enough. There is no code to write.

Anything you can see in an image and point out: parts, products, defects such as scratches or dents, missing components. The model can draw a box around each object (Faster R-CNN), trace its exact outline (Mask R-CNN), or place points on fixed spots of it (Keypoint R-CNN). Typical uses are visual quality inspection, defect detection and counting parts.

A few dozen examples per class is a reasonable start. Images that look like the real situation, with different angles, lighting and backgrounds, help more than many near-identical photos. After training, the app shows a score and a list of the model's biggest mistakes, so you can see where more images would help.

You can bring them along: add a folder with labels in YOLO or COCO format, or import your labels from Label Studio.

Labeling is built in, so you do not need a separate annotation tool. You can draw boxes, trace exact outlines (polygons) or place keypoints, and your labels stay on your own PC, next to your images.

Yes. SAM 2 is built in: click an object once and the app draws its outline, which you can adjust. Auto-label lets a trained model suggest labels for new images, which you check. All of it runs on your own PC and nothing is uploaded.

A Windows 10 (version 2004 or newer) or Windows 11 PC, 64-bit, with an NVIDIA GPU that has at least 4 GB of VRAM and a driver that supports CUDA 12.1. At least 16 GB of RAM and about 10 GB of free disk space. See What you need for the full list.

For training you need an NVIDIA GPU. Without one the app still runs: you can build datasets and label, including one-click outlines with SAM 2, but training, testing, comparing and Auto-label need the NVIDIA GPU. The app runs on Windows only, so a Mac is not supported.

The installer brings everything the app needs, so there is nothing else to install or set up. When a new version is out, the app updates itself.

English, Dutch, German, French and Spanish.

Devimation Studio is licensed per month; the plans are on the pricing page. Right now it is in a pilot with a small group of companies, and licences go on sale after the launch.

If you need the model connected to your own program or production line, send a backend request and we will see whether we can build it for you.

Email us at info@devimation.nl and we will gladly help.