You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
It's been a year since we last used GANs. You may have been lead to believe that GANs are -the- way to get realistic results but believe me, there's actually better ways IMHO. I can't tell you exactly what we're doing now but I can tell you this: 2/
It was a good thing I got past my ego (after some time) and listened to @jeremyphoward and @fastdotai when they said they that they are getting "better than GAN" results using perceptual loss in super resolution :)
3/
I'm trying to straddle the line between maintaining a competitive advantage and giving back to the community that served us so much. It's tough to make the call sometimes. 4/
What kind of perceptual loss did you end up using?
That's the part where I have to refrain from details (don't want to give it all away). I just want to at least point people in what I consider a very productive direction, which is to take the idea seriously that you can achieve great results without GANs!
If you use perceptual loss in GANs, it still doesn't get better results than the method you are using now?
That's what I was originally doing actually- perceptual loss along with GAN loss. GANs tend to go haywire with glitches and introduce undesirable constraints. It's a net negative even when paired with the latest training we're doing now (I've tried).
Are we talking about normalizing flows? What approach do you think is the best at the moment?
For people looking to dive in here, check out "Perceptual Losses for Real-Time Style Transfer and Super-Resolution" from @jcjohnss @AlexAlahi @drfeifei: https://arxiv.org/abs/1603.08155
GAN generates details in its way. I feel like the result with GAN is better but quantitatively worse. I usually consider perceptual loss firstly rather than GAN. Since plausible colors can be different from the ground-truth, how did you evaluate your results?
So we have two metrics: FID (super slow) and another one that's very fast that isn't quite as good but gives us fast feedback across resolutions to let us know how good the generalization on that is. Unfortunately neither of these are 100%, so visual inspection is a must. Lots.
Is VQ-VAE-2 on the right track?
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
