Training @NvidiaAI's #StyleGAN on Google Earth sattelite imagery, will take a few more days of training to get good samples, but already looks promising! Unfortunately don't have enough compute budget to run the model at full 1024 resolution...

Feb 16, 2019 · 5:15 PM UTC

Real samples vs Generated samples after 24 hours of training. Gonna wait another day to generate latent space interpolations, let's not spoil all the fun at once ;-)
Final generator samples after training #StyleGAN on satellite imagery for 3 days using 2xV100 GPUs at 256x256 resolution. Absolutely love the variety & esthetics of the results!
Next, here are some mandatory slerp-interpolations in latent 'w' space, using some code from @phillip_isola gist.github.com/phillipi/d29… GIF loops between 3 latent midpoints and finally crosses the mean of the distribution since loopable GIFs are dope 😋
And finally, @gwern, here are some rather artsy-looking high-resolution samples generated with ESRGAN. The super-resolution GAN adds some funny artefacts which makes the result kinda look like oil paintings. (didn't use these for the interpolations since they introduced jitter)
What budget do you need and what GPU on AWS ?
I'm using 2x V100 GPUs with 12 cpu-cores and 100Gb SSD memory, a little over 100$/day of training. Training the full 1024-resolution model costs roughly 2k$ when using cloud compute..
Love this. I've been experimenting with some sky images. Do you mind sharing your data source?
Hey there, any chance you've open-sourced this model? Would love to play around with it for a music-to-latent-image-space art project I'm working on.