Getting Started with Pixeltable: Versioning AI Datasets in Minutes

Pixeltable promises a Python-native dataframe API over versioned multimodal data. In this tutorial, you will build a small image dataset, run a vision model over it, and see how incremental computation and data versioning actually work — with only a few lines of code.

Step 1: Install and Connect

Install the Python client and connect to a Pixeltable instance. The client talks to the backend over a simple API, so your data and compute stay on the server:

pip install pixeltable

import pixeltable as pxt
pxt.connect('http://localhost:8080')

Step 2: Create a Versioned Table

Tables hold mixed column types — URLs, text, and structured values — and every mutation creates a new version:

t = pxt.create_table('image_catalog', {
    'image': pxt.Image,
    'caption': pxt.String,
})
t.insert(image='https://example.com/photo.jpg',
         caption='A city skyline at dusk')

Step 3: Run Inference in the Dataflow

Attach a vision model and add a computed column. Pixeltable tracks dependencies and only recomputes rows whose inputs changed:

t.add_computed_column(
    embedding=t.image.embed(embedding_model))
t.add_computed_column(
    label=t.image.classify(classifier))

Step 4: Replay and Branch

Wrap-Up

In a few minutes you had a versioned, queryable multimodal dataset with model inference built in. That combination of dataframe ergonomics, automatic versioning, and incremental compute makes Pixeltable a compelling backend for serious AI data work.