How did she appear?
Yesterday I was experimenting with the new Protogen v3.4 model for Stable Diffusion - an AI architechture for image generation.
I was playing with the option to give the model an input image along with a text prompt. Curiously I just used a black image as input with the simple prompt:
a photo of a woman
.
Nothing fancy. No styles applied. No negative prompt. Just these 5 words. I chose 768x768 as the output size, as it is the highest setting that runs on my GPU. The first image I generated is the image you see in the header above: A woman indeed. But something seems off about her. So I ran the model again with randomly chosen seeds, and again, and again. I realized one thing:
It was the same woman over and over again. First I thought it's a coincidence, but there is something about the black input image that evokes this creature.
She does not look creepy in all of them. In fact it is just a slighty asian looking, middle aged woman with wrinkles. What creeps me out is the fact that she, and almost exclusively she, appears in almost all images I generated, even though the seeds were random. I had so many questions. Here are some of many examples:
What is this woman? Is she a result of overfitting the neural network? If so, what is the base for this? Does this woman exist?
Only a few weeks back I read an AI horror story about an emergent woman named Loab. I have to admit that this woman I found in the dark image is less creepy, but the phenomenon bears some similarity.
Notive how sometimes the woman is older or younger:
The search for the woman
I tried to use Google Image Search and Google Lens with several generated images to find images online that resemble any similarity with the woman. However they yielded no reasult. If this woman had been a well known person, I wouldn't be surprised by her recurring appearance, but it seems like this woman is entirely made up by the AI.
Only the more younger looking appearances gave me some association: Barriss Offee, a character from Star Wars: The Clone Wars. Because both have black dots in their faces.
But the woman out of darkness has totally different dots that look more like age spots. The weird skin is another similarity to Loab, who is admittedly far more horrible to look at.
Variations: What about girl, lady, wife?
I also tried the following prompts to see if they yield different results: a photo of a girl
, a photo of a lady
and a photo of a wife
.
The last lady looks like she knows something, she doesn't want to tell us.
Is there a woman out of the white?
I also tried to use a plain white image as an input as opposed to the black input. Be there was usually just some hairon the side and a woman did not show up. Only one time I managed to get a snippet of a woman:
Questions remain
I still have a lot of things I want to try out:
What about different colored input images? red, blue, green? If I put in
photo of a man
do I also get a similar photo over and over again? What is the influence of step count and image size? What about other stable diffusion models?
After all I got used to seeing this woman, she is not super creepy to look at after all, even though her lip expression and black eyes are a bit weird. Please let me know what you think about this effect and if you experienced something similar before.
About Tadeo Hepperle
Hey, I am Tadeo, a Software Developer with a passion for Rust, Dart and Deep Learning. I love to explore how to make use of generative AI, especially for games.