Pictures generated by artificial intelligence machines he is Twitter account Created 2 weeks ago and you have already Tweeted or Retweeted Hundreds of photos created by artificial intelligence. But we are not faced with the typical face generator, which aims to produce realistic images and It is indistinguishable from reality, No.

Instead, this account submits Pictures that fit a short description (alternatively, one or a few keywords) provided by the user. Drawing on what you’ve learned during your training with thousands of real photos, The algorithm produces from scratch an image that, in concept, corresponds to the description.

In many cases, descriptions are contributions from account followers who, Placing a public order, those responsible for it choose and publish it. In other cases, they limit themselves to retweeting photos created by other users using the same artificial intelligence.

While in some cases the result is ridiculous or incomprehensible, in most of the pictures It is wonderful – and even hypnotic – to appreciate the result and how it relates to the concepts presented: It turns out that there is a disturbing gap between the human imagination and the “artificial imagination”, even though the latter is inspired by the former.

These robots don’t dream of electric sheep…

Let’s see some examples:

In addition, in some cases they also publish Videos that allow us to appreciate the “development process” of the image. In this case, it shows how a certain base image is provided, which the algorithm will modify to match the given description. As in the following example…

• Video “Re-Wave Basin” … created from a photo of a hand:

Recycled swimming pool

I grew up with a picture of my hand pic.twitter.com/5t1zsRNIr7

– Images generated by artificial intelligence machines (images_ai) 25 June 2021

What is the technology behind these nightmares?

The key to powering this AI is what we call GANs Ohhostile generation networksWe give this name to pairs of neural networks that – simplifying the mechanism a little – compete with each other: one of them (which we call ‘generative’) builds one image and the other (file “discriminatory”) responds with a degree of similarity to others (that of training, contributed by humans) with similar descriptions.

Then, the first refines the image for the second to judge the degree of similarity to a higher degree. And so, over and over again until it is impossible for the former to improve the resulting image.

In this case, the technology used to create these images has been dubbed big sleep, the result of the combination of GAN (commonly known as Biggan) With CLIP (Pre-training for Divergent Picture Language), an artificial intelligence responsible for converting texts into images, developed by the creators The famous and powerful GPT-3.