Paper here: https://arxiv.org/pdf/1903.07291.pdf

Page 14 and beyond show a number of examples vs previous methods.

IIUC, their contribution is a masking normalization layer (the 'SPADE' layer [see pg 4]) that allows the network to preserve the structure of the input segmentation/mask, leading to results quite a bit nicer than previous ones.

All future media should be watermarked with immutable metadata linking it back to it's original source to verify authenticity and combat fake news.

GAN generally works very well for landscapes, which is what is being demonstrated, not so much for objects esp humans. It’s still impressive though.

Aren’t GANs behind things like thispersondoesnotexist.com?

Yes. But that's a network finely tuned on facial images, which is more complex than landscapes. GANs can create all types of images, it's just that landscapes happen to be the type of media currently created with the least complexity due to the more abstract nature of the details involved.

Can't we just be honest about the current state of the art and say "near photorealistic"?

Unless you're half-blind you can tell there's something off about these pictures.

How did this not get more attention??