So wrote an application / version of text to image using bigGAN. Pretty trippy "interpretations" when you use any token related to "space" in the input. Here's one after just a few hundred iterations..

< 10% of the generated output is "good" from my (non-artist) POV. There's definitely two angles...1) the GAN I'm using as a source (bigGAN) needs more training samples and 2) there's always a human component to art (curation)

Space is probably so fascinating to us because few of us can extrapolate or conceptualize reduced (or lack of) gravity, no O2, how space flight or object motion works. Trees are also cool. So let's see what the algo's interpretation of "a space tree" is...

Frustrating limitation of #AI generated #art is resolution of training samples used for the GAN. Question becomes do you up res on training or can you approximate the up res via a separate downstream model. Power / training time vs. perceived "accuracy" of the end product..

The Milky Way and the beach at Fire Island — with a massive light-pollution dome from NYC and Long Island.

#failure is such a weird concept. When purely professional (vs. personal hardship, etc), it's probably the greatest "proof" of being able to take chances and survive. Can choose to view it as evidence that you have been & will be ok...and try, or let it become crippling fear.

"Code-free" #website builders sit in such an uncanny valley for me. They speed up 80% of the tasks, but the last 20% of customization you need to do takes 80% longer as you figure out how to hack around their limitations

And New York mobile.nytimes.com/2017/12/28/nyr…