Engineer & Data Scientist

OLIVIA FRYT

THIS FASHION DOES NOT EXIST

Jan 2023-Ongoing
Early results (12,000 KIMGs) from the trained StyleGAN2 ADA model.
Latent space representation of a randomly generated dress from the model trained to 12,000 KIMGs.

This Person Does Not Exist was launched in 2019 by engineer Phillip Wang in 2019, and showcases a trained StyleGAN2 (by NVIDIA) model that renders hyperrealistic images of people that, well, do not actually exist. The model was trained on images of real people, and it learned how to recreate those images. GANs have piqued my interest for years, starting in 2018, with an attempted CAPTCHA-cracker to enable my security research team and me to automate the login and scraping of a criminally-run marketplace selling stolen credit card data. This art project has a bit more whimsy than that application!

Fashion is art, and I'm especially drawn to dresses and gowns that make a statement and arguably can be considered art pieces. For this project I hand-selected a number of my favorite designers and collected almost 8,000 images for training - a relatively small training dataset for most GANs. I preprocessed the images by making the backgrounds a consistent white color, ensured the same amount of padding was used between the dress and the top/bottom of the image bounds, resized the images a height of 1028, and padding the width accordingly to also a value of 1028. I then began training NVlab's StyleGAN2 ADA model. I've actually ended up training on multiple platforms, primarily GCP and Google Colab. I've had to make some custom edits to NVIDIA's code to account for stopping/restarting training from a checkpoint, as well as relevant image augmentations for my images.

I'm considering this an ongoing project as I try not to immediately break the bank on computing resources. Early results seem promising, and I'm looking forward not only to continuing to train the model to completion, but also fine-tuning or training on top of this model with additional, more focused (perhaps seasonal) datasets. As a closing remark, I'd like to add that I do not believe this model will every be able to match the limits of human creativity. Rather, I think it can be used as a tool to highlight such - a medium by which artists, creators, designers, and appreciators can look back on previous work and progress from a new perspective. Not only does this model create beautiful art in and of itself, but also I believe it has the propensity to provide inspiration to others for their creative endevors.