Making Art With GANs

Final Project Progress

For my project I’m exploring gender in GAN generated images, using the model Disco Diffusion. Inititally I wanted to investigate how bias and mysogyny in the data set used to train the model might manifest by feeding the model prompts with gendered language.

My first experiment I gave the model almost no details, using the prompt:

“Very detailed group of women, 4k ultra HD photograph”

Right away I am validated in my desire to pursue this topic because the images I got were quite pornographic. I was running this model on a train and had to hide my tab because the images coming up were undulating amorphous blobs of naked body parts.

All I can say is YIKES.

The “4k ultra HD photograph” part, is part of the default prompt in DD, so I left it in. Prompting the model with words like “detailed” and “HD” do result in a clearer image. Generally those words shouldn’t effect the content beyond that but I believe in this case it made the results more pornographic.

After much experimentation with this topic I started to find it kind of depressing and not really interesting. My hypothesis was proven in the first five minutes- breaking news: the internet hates women and views them as objects! Shocking. I decided to move on to try to reclaim these tools to create pictures of my version of femininity, using it to tell my stories and working around the gender bias that became so apparent to me.

I revisited my journals from undergrad and experimented with different ways I could use machine learning models to depict my stories of becoming who I am today, and how to control a medium that has a mind of it’s own.