Are my photos or webcam frames uploaded?

No. The image model is downloaded once and runs in your browser. Samples, training and predictions all stay on your device; projects are kept in this browser's storage.

How many samples do I need?

Ten to thirty per class is usually enough to start. Add at least four per class so some can be held out and the confusion matrix can show how the model does on samples it did not train on.

What does the exported model contain?

A JSON file with your labels, the recipe for the image model it sits on (file, SHA-256, resizing and normalisation), and the weights of the small classifier you trained. Usage code for JavaScript and Python comes with it.

Which feature model should I pick?

Accurate (CLIP ViT-B/32, 89 MB) learns new classes from fewer samples. Light (ResNet-50, 26 MB) downloads faster on a phone. You can switch at any time; samples are re-read with the other model.

Train your own image classifier, right here

Name a few classes, show the webcam some examples or drop in photos, and train. It learns on your device in seconds, then tells you live what it sees. Nothing is uploaded.

The example trains on 64 bundled shapes and downloads no model.

1 Classes

2 to 10 classes

2 Train

Image model (runs in your browser) Accurate — CLIP ViT-B/32 · 89 MB
Learns from fewer samples. Light — ResNet-50 · 26 MB
Faster download on a phone.

Classifier Linear — fast, hard to overfit Small neural net — one hidden layer of 64

Advanced settings

Epochs Learning rate Weight decay (L2) Held out for testing Balance classes with unequal sample counts After testing, retrain on every sample

3 Try it

Train first, then point the camera at something.

Confidence

No trained model yet.

Export

Train a model to get its usage code.

How it works

Each sample becomes a vector. An image model that runs in your browser turns every 224 × 224 sample into a list of numbers describing what is in it.
A small classifier learns your classes. Only that last layer is trained, on your device, so a few dozen samples per class are enough.
It is tested on samples it never saw. A share of every class is held back; the confusion matrix shows which classes it mixes up.
Everything stays here. Projects are saved in this browser. Use "Save as file" to back one up or move it to another device.