StyleGan 3 Outputs
The generation of fake or non-existant faces through the using of GANs has popularized in recent years, particularly through projects as the one hosted in the website "https://thispersondoesnotexist.com" whose generated outputs nevertheless reflects the bias of easily accesible datasets, where faces generated tend to show people with certain phenotypical characteristics, alongside a similar fashion and make-up: that of populations inhabiting the West, usually with more Nordic Hemisphere looks, hairstyles and clothing.
For the Singing Rivers project I have put together a dataset of inhabitants of the Peruvian Rainforest, and trained a pre-trained StyleGAN3 with this data, so to see how the machine learns and process these differences. Facial decorations, hair dresses, and the particularities in features are not an easy task for the GAN to learn from, even if the pre-trained model was based on faces of people.
As it will happen with all the research modules of this project, this StyleGAN3 model will continue to be trained upon new found data
It is worth noting that most images used for training, obtained by manual web-scraping, consist of pictures of indigenous activists available in public domain. This is relevant not only because of how many they are, but that, for some, their online image is the only thing left to remember them, as they have been killed for defending the forests they call home.
