DALL-E 2, Midjourney, Stable Diffusion. The web has given itself entirely to these algorithms, and in recent weeks we've done nothing but generate millions of images using artificial intelligence. Over time, we've come to accept many of their errors and quirks, and a Swedish artist known as Supercomposite discovered Loab, a terrifying woman who for some strange reason keeps appearing among her creations. Is Loab the first cryptid of artificial intelligence?
In recent days we've learned about the main limitations of these image generation algorithms. Personally, I can say that hands and eyes are recurring weak points, and this is easy to verify by doing a quick search on portals like Lexica. If you search for “hands” with its engine, you'll notice that no decent human hands in the gallery; all you'll see is a plague of hands with six fingers, double thumbs, and other aberrations.
Speaking of aberrations, a Swedish artist known as Supercomposite recently shared the results of her experiments with artificial intelligence, and one of the most disturbing is Loab. Who is this woman? Supercomposite describes her as a “persistent presence” that emerges with the application of “negative weights” in the prompts.
Loab: Hidden Monster or Algorithmic Quirk?
What does a negative weight in the prompt mean? Suppose you want to reinforce the idea of the ocean in your description. For that, you can use the expression ocean::1, but we've also seen many prompts with exclamation marks. However, a negative weight turns the previous expression into ocean::-1. When requesting the generation of a logo with a weight of -1, Supercomposite found four variants of the same devastated woman covered with rosacea.
The crossbreeding of images, that is, using an image as influence or mixing material for the generation of others, only made things worse. For some reason, the algorithm seems completely tied to the idea of Loab, even when its main characteristics are eliminated (for example, red cheeks). Supercomposite also explains that the woman disappears from some image generations, only to return with force later.
In summary, Loab not only “invites” us to explore the potential of prompts with negative weights in image generators, but also forces us to ask ourselves what other kind of digital horrors are hidden in training models....
Source: Supercomposite on Twitter
