› Pixel Engine

A picture is a grid of numbers

An image model never sees your cat. It sees one number per square: how bright it is. Robot eyes shows those numbers, then slides a small 3×3 grid of weights, called a kernel, over them. Each square's new value is its neighbors times the weights, added up and divided. Some kernels light up edges, some blur and some sharpen. The first layer of a real image model is many little kernels like these, except the model learns its weights from examples instead of you picking them.

Augment makes the 8 flips and turns of your sprite, a little brighter or darker each time. Training on copies like these is how image models learn that a turned cat is still a cat. Then take it with you: a PNG, NumPy files to open in Python, a palette for GIMP or Aseprite, or a brick mosaic that opens in Brick Builder with a real parts list.

Nothing you make leaves this device. There is no camera and no photo upload: everything starts as a drawing. More engines →