loab, explained

ever dreamed this woman?

If you’ve been on Twitter in the last day or so, you’ve probably run into this thread from user @supercomposite. In short, supercomposite has discovered, or created, a remarkably unsettling image of a woman in AI image generation tool (they declined to specify which one, but it isn’t DALL-E and I don’t believe it’s midjourney). For unclear reasons, if you combine her image with any other, you get Beksinski-esque horrors like this:

The thread was not written to go viral, and as such is full of machine learning terms that go unexplained by supercomposite, which can make it hard to understand. I’m not an expert in ML, but I (think I) know enough to decipher what’s going on here. I’ll try to explain as best I can, so everyone can understand Loab and her future.

(Because she very much has a future. As with any demon, Loab draws power from her name. Having named her and publicized her image online, supercomposite has ensured that future AI image generators will include her in their datasets, and will be far easier to summon).

To start with, supercomposite used a negative prompt weight of Marlon Brando. If you’ve ever used DALL-E or a similar tool, you know what a prompt is — you type in “oil painting of a wizard made out of green goo,” for example, and the AI tries to generate an image that matches your input by hunting through its dataset to find pictures and data that match all the terms it was given (wizard, goo, the color green) and then crafting them into a single image.

A negative prompt asks the AI to generate an image which is as statistically unlike the image or prompt you feed it as possible.* Supercomposite asked their tool to generate the statistical opposite of Marlon Brando, and got this logo for DIGITA PNTICS. They then wondered if the process was reversible — if you ask for the opposite of DIGITA PNTICS, do you get Marlon Brando again?

You do not. You get Loab.

Loab’s name, by the way, comes from some text that appeared in the background of one of her initial images; you can see it here in the top right.

The inner workings of deep learning tools approximate black boxes — it’s essentially impossible to know why the tool decided that DIGITA PNTICS was the opposite of Marlon Brando, nor why Loab was the opposite of DIGITA PNTICS. Nothing about this is exceptionally odd. What is weird is that the AI associates her with nightmares. Supercomposite discovered this when they combined a pretty image of an ice tunnel or something with the above Loab and got things like this:

Supercomposite adds that there were images which were even worse than this, but declined to share them with me. Or anyone.

Loab, supercomposite notes, is a gestalt — a combination of data points, or “a collection of traits that are equally far away from something.” Deep learning systems group similar datapoints together: if you give one pictures of eggs, balls, bullets, and dogs, it will group the first three together as more similar to one another than any is to the fourth. You can visualize this spatially, and I’ve done so here:

In this very bad approximation of what the AI might be “thinking,” you can see that I’ve clumped related datapoints together: the DIGITA PNTICS logo is on a building, it looks like a skyline, etc. The Loab set is opposite it, as unlike those things as the computer can create. For some reason (maybe her rosacea?), supercomposite’s tool is grouping “Loab” with some really horrific datapoints, which manifest when the image of her is combined with another (at which point it pulls from its latent space — that map of data points — to add things like screaming child heads). Loab herself, as I said, is a gestalt: a combination of datapoints like “woman,” “red pronounced cheeks,” “squared jaw,” et cetera. The AI is quite consistent in recreating her: she’s clearly recognizable when mashed up with Kirby or bees, and can even spontaneously re-appear further down the image lineage (i.e., after multiple recombinations have erased her from the image). According to supercomposite, the AI is really good at reproducing Loab, doing so “more easily than most celebrities.”

There’s no way to access Loab through text prompts (yet). The AI doesn’t know her name is “Loab” (yet). Supercomposite just happened, by pure chance, to stumble on a very disturbing cluster of datapoints in their AI’s dataset that reliably generate Silent Hill-looking monstrosities, for reasons absolutely no one understands.

Whatever the reason, she has a name now, and since machine learning datasets grow in part by pulling images + associated text off of the internet, it’s virtually guaranteed that future iterations of image generation tools will let you summon her directly.

*Negative prompts can be weighted from 0 to -1, where -0.1 would be 10% dissimilar, -0.2% would be 20% dissimilar, and so on until you get to -1 which is 100% dissimilar. Supercomposite asked for Brando -1, so the precise opposite of Marlin Brando.