Photos from Crude Sketches: Nvidia's GauGAN Explained Visually
adamdking.com
adamdking.com
That said, I'm not sure why one wouldn't get a similar result training on the EC2 or GCE instances that have 8 V100s. Or even training with fewer GPUs but accumulating gradients to get the same batch size.
I should feel happier about it, but I can't stop feeling a bit odd that a sketch can go to photorealistic north of the bad so well now: I expected 5-10 more years for this.
If this was intruded background in a movie at HQ, I'm not sure I could always tell the difference although its equally possible the renders would be unsustainable under motion, not enough real-world change to look "real" and it re-enters uncanny valley
the other day i opened one of those links and it was GTP-2. besides all the insane implications of GTP-2, what bothers me is that i am no longer able to assume that any internet comment is written by a human, no matter how convincing. there are still comments that GTP-2 could not write but anyone who points that out is pretty short sighted because it wont be long before there are vanishingly few comments that could not have been generated. i kind of liked knowing that a person was typing out (almost) all those comments.
one of the biggest realizations ive had recently is that technology does not cut equally in both directions. everyone in my generation has thought of technology as a neutral entity: for every benefit of a given technology, one can point out a corresponding disadvantage. on the surface it seems like the scale dips neither for the societal disadvantages nor for the societal benefits. this is a very fundamental belief. and its wrong. its funny how people put so much faith in such fuzzy logic.
the implications of that realization are difficult to swallow. it means that with every new technology introduced into the world, there is the potential for it to harm peoples quality of life. or improve it. but there is no regulation of technology so its a crap shoot. weve been rolling the dice for a long time and we didnt even know it. and i think weve been winning. but i think that high level automation is not going to be a win for us.
besides all of that, there is absolutely no debate that these advancements in ai are to our generation what personal computers were to the baby boomer generation. without close attention, we will fall behind and our kids will have fluency in the new world of automation while we cling to very old and outdated patterns. in other words, it makes me feel very old.
Maybe that's just because technological innovation is slowing down.
Societal changes lag technological ones by at least 5-10 years. So the changes we're feeling now were largely the result of technological changes in the early 2010s. But I do think technology today is slowing down. Individual processor speed certainly has, which has far reaching implications. Cloud computing and GPUs have given general purpose processes another step in "perceived" performance, but those are pretty much one-trick ponies.
If processor individual performance doesn't increase, the economies of scale that a large data center gives you eventually has diminishing marginal returns, and you're again limited by individual processor speed. GPUs similarly give a speed-up for applications that can optimize for them, but eventually they will run into the same performance walls that general purpose chips run into.
Other technologies, like machine learning and much of genetics heavily rely on exponential improvements in the underlying hardware.
If the death of moore's law is really happening, it will have far reaching implications in all computational based industries.
The future golden age of indie creators could be fueled by this.
Nicely done.
If the massively online education people had that kind of quality, maybe people would actually finish the courses.
Specifically this is the "texture-by-numbers" application. Ex: https://www.mrl.nyu.edu/projects/image-analogies/potomac.htm...
Every single fancypants application of neural nets in graphics today is a retread of one of the applications of the Image Analogies algorithm.
"we first label a photograph or painting by hand to indicate its component textures. We then give a new labeling, from which the analogies algorithm produces a new photograph"
DL approach generalizes over many images and can derive some some "idea" of how given class should look like (ie. a tree)
The nvidia version seems to inpaint new details into the user segmented areas, like a collage of sorts.