Just fooling around with a GAN to generate some NFTs. I only had about 200 source images that I scraped from a few different projects. It’s pretty obvious that this output is a little overfit. When observing the outputs of the GAN over many generations, it’s really interesting to see it create images like this or even better, then deteriorate into noise. Hyperparameter optimization required.
My GPU took 45 minutes to generate these beautiful people from nothing. Well, not nothing. It was a Generative Adversarial Network. What’s that you ask? It’s an architecture for a neural network. This architecture was developed by NVidia, where they showed how, with just a bit more hardware, you can generate photo-realistic people that never existed.
Pretty neat. I’ll fool around with this a bit to see if I can optimize it, but I’m also looking forward to testing it on other media.
Huge shout out to Jeff Heaton for the YouTube tutorial! I’m looking forward to trying this out on other data sets.
