If you follow the news on deep learning, you may have encountered the applications of Nvidia’s StyleGAN that generates photo-realistic fake faces. Since the typical dataset that StyleGAN is trained on consist of amateur-produced portraits from Flickr, I was curious to see what StyleGAN would learn from social media posts. I crawled over 1M social media posts from major fashion brands and set StyleGAN to work on the subset where a single face is detected. Given the limited GPU computational resources available to lowly marketing professors like myself, it will take quite some time for the model to train sufficiently. Here, I will keep track of the training progress where 1 frame = ~ half an hour of training on my home workstation.
Video: Updated 6-16-22 (Started 5-31-22, trained on ~10,000k images (and their rotations/translations) so far)
How it started
Initial fakes, Frechet Inception Distance 50k (FID50k)=315
After 512k Examples, FID50k=77.6
After 5120k examples, FID50k=55.3
After 6720k examples, FID50k=49.2
After 8016k examples, FID50k=45.7
Dr. Yang Wang 