Inspired by a number of #StyleGAN takes, and in particular, thispersondoesnotexist.com from @lucidrains, I spent the last couple days teaching myself enough to put together ThisAirBNBDoesNotExist.com. These listings are entirely invented; no aspect of them exists in the world.
This page uses a combination of models: StyleGAN models for profile pictures and bedroom photos, plus a handful of text networks trained with tf.keras to generate place names, host names, titles, and descriptions (working from Tensorflow example code colab.research.google.com/gi…).

Feb 17, 2019 · 6:24 PM UTC

All training was done in Google's Colab (research.google.com/colabora…), which gave free access to GPUs + TPUs for training and generating data from models.
Each of the models does independent predictions, so there's plenty of cases where bits don't agree -- eg, the description will claim to be a one-bedroom, while the listing is fixed to four bedrooms, or faces + names won't line up well.
But overall, the experience is pretty compelling -- and I've had a bunch of fun learning more about how to play with and manipulate some of these models, thanks to Colab, and the excellent work of the #StyleGAN crowd.
I'll be doing a longer writeup over the next couple days/weeks of some of the things I learned in this experience, but in the meantime, enjoy the somewhat creepy nature of completely fabricated AirBNB listings. thisairbnbdoesnotexist.com/
This tweet is unavailable
I found the StyleGAN repository description on this to be pretty straightforward! Note that you'll probably need a *lot* of images to train a decent model. github.com/NVlabs/stylegan#p…