My intuition is that in such a setup, the discriminator can only be as good as your generator. This work provides a nice general-purpose generator, but I'm not sure as to how far you can take this (especially since vision nets don't really operate like human vision)
It seems like the major cost here is that you need to train a high quality StyleGAN for your dataset -- which is very data hungry - at least thousands of images for high quality results. (The famous face generator used millions of faces)
Yes, but at least it's unsupervised.
I think I'd seen StyleGAN2 using like hundreds of examples! And I'm pretty sure Flickr-Faces-High-Quality only uses 70k examples.
I think that the idea is that you use transfer learning to bootstrap from a general gan to a gan for your domain - and then use that to generate odd data.
Last I checked, transfer learning for StyleGAN2 still produced somewhat wonky results if the datasets were not extremely close. Have there been improvements in this field?
I don't know - but I guess that must be in the paper. I had intended to read this properly after I looked at it when the story came out, but I haven't had time, sorry.
But that’s not bad compared to the cost of pixel-wise labeling to get training data for segmentation.
Looks a lot like this [1] paper from a few weeks ago (which has actual code).
This looks like it advocates hooking the output of one network up to the input of another, which seems like a very roundabout way of copying the weights from one network to another.
This is the best role the Shanara lead actor got after the series.