I am so confused by two completions of a human girl here. How is this possibly close in image-space to all the other images, especially given this prompt?
That’s an unusually realistic face, and a distinct hairstyle. I suspect that’s a real person and if it is, knowing who might shed some light on how the prompt could possibly be tapping into her—that she shows up twice (and it’s obviously the same girl twice given the hair style and clothing are the same) suggests there is some sort of real connection, like she’s an animator famous for cartoon monsters or something.
This also replicated when I asked someone else to generate new images for the same prompt (one image out of the 10 was again in this very different style and displaying approximately the same person).
Very strange. I did some searching in Google Images & Yandex for the cropped face and for ‘furry black monsters’, and asides from being impressed just how many more women Yandex turns up who do in fact look a lot like the sample, didn’t find anything obviously relevant.
Interesting. Both the 2 images of her have a white house wall to the left with same lighting, same hair color and bottom colored hair, same shirt color, and same skin color.
Maybe the words ‘wearing’ and ‘outfits’ and ‘black’ and ‘alley’ and even ‘dolly’ and ‘photo’ may have triggered it to give us an alley—but one that has a clothing fashionist in it lol.
It still seems to be choosing a single source though mostly.
This seems like a major case study for interpretability.
What you’d really want is to be able to ask the network “In what ways is this woman similar to the prompt?” and have it output a causality chain or something.
I am so confused by two completions of a human girl here. How is this possibly close in image-space to all the other images, especially given this prompt?
That’s an unusually realistic face, and a distinct hairstyle. I suspect that’s a real person and if it is, knowing who might shed some light on how the prompt could possibly be tapping into her—that she shows up twice (and it’s obviously the same girl twice given the hair style and clothing are the same) suggests there is some sort of real connection, like she’s an animator famous for cartoon monsters or something.
This also replicated when I asked someone else to generate new images for the same prompt (one image out of the 10 was again in this very different style and displaying approximately the same person).
Very strange. I did some searching in Google Images & Yandex for the cropped face and for ‘furry black monsters’, and asides from being impressed just how many more women Yandex turns up who do in fact look a lot like the sample, didn’t find anything obviously relevant.
Interesting. Both the 2 images of her have a white house wall to the left with same lighting, same hair color and bottom colored hair, same shirt color, and same skin color.
Maybe the words ‘wearing’ and ‘outfits’ and ‘black’ and ‘alley’ and even ‘dolly’ and ‘photo’ may have triggered it to give us an alley—but one that has a clothing fashionist in it lol.
It still seems to be choosing a single source though mostly.
This seems like a major case study for interpretability.
What you’d really want is to be able to ask the network “In what ways is this woman similar to the prompt?” and have it output a causality chain or something.