How AI image generators make a picture from your words
An AI picture starts as a square of random dots
You type a few words into an AI image generator. A few seconds later, a picture is on your screen. No camera took it, and no artist drew it. Nobody cut it out of a photo. Where was that picture one minute before?
-
1
You type "a fox reading a newspaper in the snow" into an image generator, an app that makes pictures from words, and press Generate.
-
2
A few seconds later, there it is. No camera took this picture, and no artist drew it.
-
3
The app makes nothing itself. It passes your sentence, called the prompt, to the program that makes the picture: a model. The one shown here is Stable Diffusion 1.5.
-
4
The model does not look your fox up on the web, and it does not cut up photos to glue them together. It starts from a square of random dots, called noise.
-
5
Then it guesses which dots are noise and takes a little of them away. It looks again, and takes a little more. Pass after pass, the dots that stay gather into shapes, and the shapes into your fox.
-
6
Yet that first square hid no fox at all: lift a corner, and there is nothing under the dots. So how did the model know which ones to take away?
-
Go further
What your app does, in the words of the field, and why this model is the one shown.