The neural network created a creepy image of a Loab woman that scared users
We have long been accustomed to the fact that artificial intelligence is used for creativity. By correctly specifying the description of the neural network image, you can get an accurate and beautiful image of any object. But few people know that a neural network can generate images that are as far from the task as possible. A Twitter user with the nickname Supercomposite used an unusual algorithm to get a creepy realistic image of a woman.
When creating images, Supercomposite used the Negative Weight Prompts method, that is, "requests with a negative coefficient". Using this algorithm, the neural network began to generate images completely opposite to the task.
Most often, the neural network gives unexpected, but quite harmless pictures to "negative requests". But in this case, she showed a portrait of a scary woman, which she signed as Loab. It all started with the fact that the user entered "Brando:-1. Artificial intelligence in response to this showed a picture with an urban landscape and the phrase "DIGITA PNTICS".
Then Supercomposite entered the line "DIGITA PNTICS skyline logo:-1", hoping to get the opposite result, possibly an image of Marlon Brando. But in response, the neural network created a portrait of an elderly depressed woman with an unhealthy blush.
The Tweeter did not give up and continued to experiment. He asked the system to generate an image at the request of "an over-compressed glass tunnel surrounded by angels and in the style of Wes Anderson." To the result obtained, Supercomposite added a previously created portrait of Loab. In the task for the neural network, he indicated that it was not necessary to use a text description. It was assumed that the new image would be generated only on the basis of images.
After that, the neural network began to produce one image scarier than the other. By posting them on Twitter, the author even set a bar that the content can shock the user.
Having played enough with the woman and the tunnel, Supercomposite began to cross Loab with a variety of images. These were pictures, frames from films, descriptions of events and people. Sometimes the neural network gave out something neutral, but in most cases variations of the creepy Loab appeared on the screen.
The surprised experimenter described it this way:
It is interesting, but some faces do not look, at first glance, like Loab. But if you look closely, parts of her face still show up. Supercomposite writes that the woman's face drove him into a frenzy. There was no way he could get rid of it in the images.
Under the creepy post, a heated discussion of an unusual phenomenon began. But even those who specialize in working with neural networks could not explain anything. Any images created on the basis of Loab, no matter how many other components were added, were inevitably terrible.
It is incredibly difficult to understand how neural networks generate images. They are trained on billions of pictures and their logic is incomprehensible. Experts advise those who want to get predictable results to avoid long descriptions. Then there is a chance to see something nice, for example, naked beauties.