A scary AI-generated woman lurks in the depths of latent space

Share

However, is this AI model actually haunted, or is Loab just a chance collection of images that just so happens to appear under certain peculiar technical circumstances? Unless you believe ghosts can inhabit data structures, it must be the latter, but it’s more than just a disturbing visual-it’s a sign that what passes for a brain in an AI is deeper and creepier than we could otherwise have anticipated.

We are all frequently astounded by AI’s skills to write and create, but who realized it could evoke such fear? An AI researcher makes the startling revelation that the “latent space” comprising a deep learning model’s memory is haunted by at least one frightening figure –a bloody-faced lady now known as “Loab.”

A musician and artist who goes by the handle Supercomposite on Twitter found Loab, or perhaps she was summoned? Although there is no shortage of strange scary AI things on the network, she detailed the Loab phenomena in a thread that garnered a lot of attention, indicating it struck a nerve.

Supercomposite was working with “negative prompts” as part of a proprietary AI text-to-image model that was not DALL-E or Stable Diffusion.

The model often responds to a prompt by working toward developing an image that matches it. If you just have one prompt, it has a “weight” of1, which indicates that it is the only goal the model is aiming towards.

You can also divide prompts, saying something like “hot air balloon::0.5, thunderstorm::0.5” and it will work toward both of those things equally – this isn’t actually necessary, because the language part of the model will accept “hot air balloon in a thunderstorm” and you might even get better results.

The fascinating part is that you may also have negative cues, which makes the model actively try to avoid that idea.

This process is far less predictable because no one is aware of how the data is actually set up in latent space, which is sometimes referred to as the “mind” or “memory” of the AI.

The latent space is almost like navigating a map of various AI concepts. A prompt, according to Supercomposite, “is like an arrow that directs you where to go and how far to travel in this concept map.

The latent space is almost like navigating a map of various AI concepts. A prompt, according to Supercomposite, “is like an arrow that directs you where to go and how far to travel in this concept map.

Therefore, if you ask the AI for an image of “a face,” you’ll land somewhere in the middle of the region that contains all the images of faces and receive an image of a rather ordinary-looking face, she explained. A more precise prompt will place you among the smiling faces, the ones in profile, etc. However, when presented with a negatively weighted prompt, you take the opposite action: you flee from the idea as quickly as you can.

What, however, is the opposite of “face”? The feet, perhaps? Is it in your head’s back? an object without a face, like a pencil? While we can debate about it, in a machine learning model, it was agreed throughout the training process that whatever visual and verbal notions were encoded into its memory, they can be traversed consistently — even if they are somewhat arbitrary.

With the instruction “Brando::-1,” Supercomposite was experimenting with the idea of navigating the latent space by having the model output whatever it believed to be the exact opposite of “Brando.” It generated a strange skyline logo with the text “DIGITA PNTICS,” which was illegible but slightly readable.

Crazy, huh? But once more, we wouldn’t necessarily understand how the concepts are organized in the model. Supercomposite wondered if she could undo the process out of curiosity. She entered “DIGITA PNITICS skyline logo::-1” as the prompt. If this image was the polar opposite of “Brando,” might it also be the inverse, and it would find its way to, perhaps, Marlon Brando?

Instead, she received the following:

Image Credits: Supercomposite

She kept submitting this negative cue, and the model kept producing this figure, with bloodied, slashed, or unhealthily red cheeks and an eerie, otherworldly appearance. The AI model’s best prediction for the most remote notion from a logo incorporating meaningless words is this woman, whom Supercomposite named “Loab” for the text that appears in the top-right image there.

What occurred? Supercomposite continued her metaphor from before by explaining how the model may think when given a negative prompt for a certain logo.

Horrors don’t usually result from negative impulses, much less so consistently. Anyone who has experimented with these picture models will tell you that even for relatively simple prompts, it can be rather challenging to obtain consistent results.

If you search for “a robot standing in a field” four or forty times, you might receive as many various interpretations of the idea, some of which are hardly distinguishable from robots or fields. But Loab frequently responds negatively to this particular cue, to the point that it sounds like a spell from an antiquated urban tale.

The kind where you’re instructed to say “Bloody Mary” three times while standing in a dark bathroom and staring in the mirror. Or even earlier folk directions for finding the gateway to the underworld or a witch’s residence: Close your eyes and take a hundred steps back from a dead tree while holding a sprig of holly.

Although the phrase “DIGITA PNITICS skyline logo::-1” isn’t quite as catchy, it is at least sufficiently arcane for magic words. It also has the advantage of working. Only on this specific model, of course; every AI platform’s latent space is unique, though who knows if Loab might also be waiting to be called in DALL-E or Stable Diffusion.

In fact, the spell is so potent that Loab appears to corrupt even split prompts and combinations of other images.

According to Supercomposite, “Some AIs can use other images as prompts; they can essentially understand the image and turn it into a directed arrow on the map just way they treat word prompts.” I combined Loab’s image with one or more other photographs as a prompt, and she usually always appears in the final image.

“Some AIs can use other images as prompts; they can essentially understand the image and turn it into a directed arrow on the map just way they treat word prompts.”

One portion may occasionally be treated more loosely in combination or more complex instructions. Those that do, however, seem to not just veer toward the weird and terrifying, but also incorporate Loab in a very unmistakable way. Loab is front and center, dictating the composition with her wounded face, indifferent expression, and long dark hair, whether she is juxtaposed with bees, video game characters, film styles, or abstractions.

Any prompt or visual that is so persistent and haunts other prompts in the manner she does is uncommon. Supercomposite made some guesses as to why this might occur.

Although it’s oversimplified, latent space really is like a map and the prompts are like directions for navigating it. The system draws whatever ends up being nearby where it’s asked to go, whether it’s familiar ground like “still life by a Dutch master” or a synthesis of enigmatic or unconnected concepts: “robots battle aliens in a cubist etching by Dore.”

The layout of that map has something to do with why Loab exists, just speculatively. It’s probable because business logos and gruesome, frightful visuals are conceptually extremely dissimilar, as Supercomposite noted.

As a result, negative prompts may operate as a method for the AI to explore the edge of its mind map, skimming the notions it deems too bizarre to store among more mundane concepts like joyful faces, gorgeous landscapes, or frolicking pets.

Image Credits: Supercomposite

Unsettlingly, neither the structure nor the rationale behind latent spaces are fully understood. Many studies have been done on the subject, and there are some signs that they are structured somewhat similarly to how our own minds are, which makes sense given that they were essentially modeled after them. However, they also have completely original structures that connect across enormous conceptual distances.

It should be noted that they are undoubtedly being made on the spot, and Supercomposite informed me that there is no proof the digital cryptid is based on any specific artist or piece of art. It is not as if there are a number of photographs, especially of Loab hiding somewhere. Latent space is inert for this reason. Similar to how languages were grouped based on similarities in the prior Google visualization, these visuals were created from a collection of bizarre and awful notions that all happened to be stored in the same location in the model’s memory.

What hidden place or unconscious associations allowed Loab to emerge, complete and coherent? The model’s route to her location cannot yet be determined since the latent space of a trained model is so large and intricate.

The only way we can return to the location is by using the magic words, which we must utter while moving backward through the area with our eyes closed until we arrive at the witch’s hut, which cannot be reached by regular means. Loab isn’t a ghost, but she is an anomaly, and ironically, she could be one of an almost endless number of anomalies waiting to be summoned from the deepest, unlit recesses of any AI model’s latent space.

It might not be paranormal. yet it most certainly isn’t natural.

Read more

Recommended For You