You are signed in. Close

You are signed out. Close

Thank you. Your subscription has started. Close

Thank you. You will get an email when a new comic comes out. Close

We could not confirm your payment. If you paid, use the link "Sign in" at the bottom of the page. Close

Something went wrong on our side. Please try again in a moment. Close

Previous episode Seed: why the same prompt gives a different AI picture every time

How AI image generators make a picture from your words

Does an AI image generator keep the pictures it learned from?

Can an AI image generator give back a picture from its training data, almost the same? Researchers tested this. First, they checked how much the model's file can hold. Then they tried many prompts and seeds.

  1. 1 Before training, a program went through the web and listed pictures that came with a short text describing them. Stable Diffusion 1.5 trained on a collection of 2.3 billion such pairs, called its training data.a cup of teared fox in grasssale posterthe training data

    Before training, a program went through the web and listed pictures that came with a short text describing them. Stable Diffusion 1.5 trained on a collection of 2.3 billion such pairs, called its training data.

  2. 2 Training ran its exercises on those pictures, round after round, and Miva, a researcher, wonders what stayed behind. If the model kept them, the right prompt could hand one back, almost as it was.the training datain trainingMivaIs it allstill in there?

    Training ran its exercises on those pictures, round after round, and Miva, a researcher, wonders what stayed behind. If the model kept them, the right prompt could hand one back, almost as it was.

The rest of this episode is for subscribers

Subscribe to read every premium episode. You can cancel when you want.

Paid yearly: $99 instead of $119.88

Next episode Why AI pictures get writing and hands wrong

Share this comic

More comics