The main reason to come was to deliver a workshop titled Collective Vision of Synthetic Reality. Our goal was to educate attendees about various machine learning and artificial intelligence models and let them speculate about a variety of future use-cases when these models are already standard parts of our lives. Lenka even designed a set of trigger cards to help attendees get up to speed quickly. These were a huge success, and you’ll be able to read about them, and the whole workshop in Lenka’s write-up soon.
Because the event was a zine fair, organizers asked us whether we’d be able to come up with some merch we’d present at the table. We revived some of our old StyleGAN experiments, where we generated random vectors and fed them into the GAN. As you’d have guessed, this produces a realistically looking face, similar to ones imagined at thispersondoesnotexist.com. But what happens when we change this code
import numpy as np sigma = 1.0 shape = (1, 512) z = np.random.randn(*shape) * sigma images = Gs.run(z, None, **Gs_kwargs) return images
import numpy as np sigma = 1.0 shape = (1, 18, 512) z = np.random.randn(*shape) * sigma images = Gs.components.synthesis.run(z, **Gs_kwargs) return images
Suddenly, you’ll end up with a beautiful set of creepy, uncanny, or abstract results, depending on how much random is your vector input. This can be controlled by altering the
sigma value - standard deviation.
Using this process, we generated 5000 images, and Lenka cherry-picked 40 of them (she saw faces everywhere after this :D). We turned these into 20x20 cm prints and brought them to the fair. Yay!
If you are fascinated by these results as much as we are, you should try the website we launched: StrangeAttractions.xyz - it will give you a new art piece every time you click on the image.
Later, we even made an Instagram filter called “Strange Attractions,” which you can find and install from Lenka’s profile. The outputs from this filter are also astonishing!
Finally, we were also experimenting with StyleGAN2, but this did not provide as colorful and eye-popping visuals as the original StyleGAN.