PixelCNN — Pixel Convolutional Neural Network

Fact

A convolution model that makes images one pixel at a time, in order.

In Plain Words

PixelCNN is like making pixel art with a hall monitor watching. Start top-left, and each new square may only peek at finished squares.

It can make small images and score how likely an image is. You meet it in early image-generation research.

Related Concepts

• Autoregressive Model
• Convolutional Neural Network
• Generative Model

Autoregressive Model
PixelCNN is an Autoregressive Model because it predicts one next pixel at a time.

CNN
PixelCNN uses masked convolutions in a CNN to limit what it can see.

Generative Model
PixelCNN is a Generative Model that learns image patterns and samples new images.