---
title: "Pixeltable Core Concepts"
date: "2024-07-31"
author: "Pixeltable Team"
tags:
  - Core Concepts
  - Pixeltable 101
  - Declarative AI
  - Incremental Compute
  - Data Engineering
description: "Learn the fundamental principles of Pixeltable. Discover how computed columns and a declarative, incremental engine simplify multimodal AI data workflows."
url: "https://pixeltable.com/blog/pixeltable-core-concepts"
---

# Pixeltable Core Concepts

Pixeltable is designed to simplify working with multimodal data for AI applications. Its core principle is to provide a [declarative](/blog/declarative-vs-imperative-ai-pipelines) and incremental data processing framework. Instead of manually building complex pipelines to handle data loading, transformation, model inference, and storage, you declare these steps as part of a table's schema. That schema is a [multimodal data table](/blog/what-is-a-multimodal-data-table): structured columns plus typed media, with computed columns and indexes on the same object. Pixeltable's engine then handles the orchestration, execution, and persistence automatically.

## The Core Engine: Tables and Computed Columns

 The fundamental concept in Pixeltable is the **computed column**. This feature drives most of the framework's power and automation. Here's how it works:

 - **Tables as the Foundation**: Everything starts with a `Table`. A table holds your data, which can be simple types (like strings and numbers) or complex multimodal types (`Image`, `Video`, `Audio`, `Document`).

 - **Declarative Processing with Computed Columns**: You can add a column to a table whose values are not inserted directly but are *computed* from other columns. This is done by providing a function, which can be any Python function ([`@pxt.udf`](/blog/python-udfs-pixeltable)), a built-in operation, or a call to an AI model (like OpenAI or Hugging Face).

 - **Automatic, Incremental Execution**: Once a computed column is defined, the Pixeltable engine automatically:
 

 **Backfills**: Runs the computation for all existing rows in the table.

 - **Updates on Insert**: Runs the computation for any new data that is inserted.

 - **Re-computes on Change**: If the data in a source column changes, or if the function definition for the computed column is updated, Pixeltable intelligently re-computes only the affected values. This [incremental nature](/blog/incremental-embedding-indexes) saves significant time and computational resources.

 

 
 - **Persistent and Versioned**: All results are stored persistently. Pixeltable also tracks the history of changes, allowing you to revert to previous states if needed.

## Key Benefits

 - **Simplicity**: Focus on *what* you want to compute, not *how* to execute and orchestrate it. Read more on how this compares to traditional pipelines in our post on [AI Functions vs. Pipelines](/blog/ai-functions-vs-pipelines).

 - **Efficiency**: The incremental computation engine avoids redundant work, saving time and money.

 - **Reproducibility**: The entire data processing workflow is captured in the table's versioned schema.

## Learn More

 Understanding these core concepts is the first step to mastering Pixeltable. To see how these principles apply in practice, check out these related posts:

 - [What Is a Multimodal Data Table?](/blog/what-is-a-multimodal-data-table) — definition of the category

 - [Your First Pixeltable Project: Build a Smart Image Organizer in 10 Minutes](/blog/your-first-pixeltable-project) - Perfect hands-on tutorial for beginners

 - [The Case for a Unified Multimodal AI Infrastructure](/blog/unified-multimodal-ai-infrastructure-pixeltable)

 - [Building Multimodal AI Apps with Just a Few Lines of Python](/blog/building-multimodal-apps)

 - [Announcing Pixeltable: The Multimodal Data Store for AI](/blog/pixeltable-launch)