Breaking StyleGAN
Art made between 2019 and 2021 by training StyleGAN2-ADA on custom datasets, perturbing its layers and searching its latent space with CLIP.
Crazy Hand
I trained StyleGAN2-ADA (opens in a new tab) on only two photographs of my hand. With sound, an Elektron Analog Rytm drum machine plays the music and moves the hand through settings such as the truncation, which sets how far the picture may stray from the average hand, and the interpolation between two latent vectors. In the silent loops, random draws move it.
Monolith
StyleGAN2-ADA (opens in a new tab), trained on surrealist paintings by André Masson, Victor Brauner and Leonora Carrington, generated these stones, which move with the music.
Jesus and the latent space
"Jesus" CLIP (opens in a new tab) searches the latent space of StyleGAN2-ADA trained on faces for the face that best matches the text "Jesus". Each round, the best latent vector so far is mutated into hundreds of similar ones, and CLIP keeps the image closest to the text.
The same search for "Satan", "Personification of God" and "Robot" shows what CLIP learned to picture from images on the web, biases included, which I discussed in a thread on Twitter (opens in a new tab).
"Satan", over four rounds 
"Personification of God" 
"Robot" Ouroboros
These snakes come from StyleGAN2-ADA (opens in a new tab) trained on photographs of many species. The video walks its latent space in a loop that ends where it began, like the ouroboros eating its own tail.
Hair
Random changes to some layers of a model trained on faces erased the faces and kept only the hair.
