Offline computer vision

Label images and train models on your own machine.

AnyLearning is a desktop app for building computer vision models. Your images, your annotations, your trained weights — none of it leaves the machine it was created on.

Object detection

Drag to draw a box, then pick a class

03 objects
personpersoncup

Drag on the image to draw · click a box to select · Del to remove it

This is running in your browser, with nothing sent anywhere. The desktop app works the same way — it just also trains the model.

How it works

Three stages, in order

You cannot train without labelled data, and you cannot ship a model without a finished run. The app tracks where each project actually stands rather than where you happen to be looking.

  1. 01

    Label

    Draw boxes, polygons or whole-image classes. Segment Anything can pre-label for you, so you correct instead of starting from scratch.

  2. 02

    Train

    Pick a model size, set epochs, start the run. Training happens in a separate process on your own hardware — GPU if you have one, CPU if you don't.

  3. 03

    Ship

    Every finished run registers a model you can try on a new image and export to ONNX for whatever runs it next.

Training

Watch a run, without starting one

A real run writes progress to the project database from a separate process, and the app polls it. This is a simulation of exactly that, so you can see the shape of it before you install anything.

Training session

nanodet-plus-m · 320

finished

Epoch

20/20

Loss
0.0905
Accuracy
94.0%
Elapsed
12.0s
Loss0.09
0.000.501.0011020
Accuracy0.94
0.300.651.0011020

[epoch 20/20] loss 0.0905 acc 0.9403 lr 0.0010

Runs in a separate process on your machine. Nothing leaves it.

Privacy

Offline is the feature

Plenty of tools will train a model for you if you upload your data first. That is exactly the step some work cannot take — patient scans, factory-floor footage, anything under NDA, anything a data-protection officer has an opinion about.

No upload step

Images are read from disk and stay there. There is no bucket, no ingest, no sync.

No account

Nothing to sign into. The licence is checked against the machine, not a session.

No running cost

Training uses hardware you already own, so a longer run costs time rather than money.

Tasks

Five kinds of project

Each one comes with its own labelling tools, model variants and training defaults, so the choice you make when you create a project is the only configuration most people need.

  • Object detection

    Boxes around things that matter

  • Image segmentation

    Outline a region precisely

  • Instance segmentation

    Separate every occurrence

  • Image classification

    One label per image

  • Handpose classification

    Gestures from hand landmarks

The app

See it in action

Pricing

One purchase, no subscription

A free tier for labelling, and a paid licence when you want to train. Licences cover a fixed number of machines and do not expire.