Holy crap, this is some mind-melting stuff. Google’s “Deep Dream” software takes images, identifies certain features and builds on them, often arriving at results that are completely different from the source image. Some fine still examples here.
The really trippy shit happens when it processes video. For example, here’s an unmodified clip from Fear and Loathing in Las Vegas
And here’s what that same clip looks like after Deep Dream has had its way with it. :eek:
I saw the stills a couple of weeks ago. Video is interesting, too.
This is the neural-network thing, isn’t it? You train it to recognise, say, a puppy, and then turn it loose on non-puppy input and see what it does…
I’m convinced something like this explains people seeing faces in things.
Aye, that’s what a dog-centric acid trip might look like alright.
I wonder what it would do for an already trippy movie; like what would Koyaanisqatsi look like after Deep Dreaming it?
Hmmmmm, I might have to try that out…
Sure. What’s miraculous about the human brain is that we aren’t seeing things that aren’t there, all the time.
VICE just put up a story featuring porn pics run through Deep Dream… it’s pretty twisted, totally NSFW and contains at least one image that could disrupt your own deep dreams… forever
Note: Edited to hide the actual link. Some images are NSFW- ITD
Not sure if you’re talking about the face-full-of-dicks image, but that one made me LMAO.
Of course it would be a Machine Elf that comes to show us freaky trippy images.
Is there a mislabeled Nightmare Fuel switch somewhere? That donut almost made me sick.
I downloaded the software last night and I’m going to try setting it up on my computer. I want to see what happens when I turn it loose on my sketchbooks.
Edit: Link.
Did this algorithm get frontloaded with the notion that it should see/find animal faces? I mean, it isn’t drawing flowers or skyscrapers, it’s inventing faces, most of them of the earthly but nonhuman design.
This guy gave it photos of the Republican candidates to chew on.
Google’s AI had nightmares about the GOP presidential candidates. This is what it saw. Scroll down for the pictures.
These pictures look like a team effort by Salvador Dali, Picasso, and Hieronymous Bosch.
Add H.R. Giger to that list, and I’d say you’ve got it.
Dang, Bobby Jindal is cute as a button! The Puppyslug Nebula was great, too. Reminded me of the three wolves/moon T-shirt.
The doghead pizza was a little unsettling, though. :eek:
I dread the day when we will have to explain to a sentient AI that Grimm is, in fact, not a documentary.
Maybe I missed it upthread, but why are there so many “suggestions” of dog faces in the images?
The “neural network” is “trained” by showing it a vastly large number of known pictures, and letting the AI software figure out for itself what their common salient features are. Apparently, the version that they release to the public had been “trained” with a large number of dog pictures, although there were other pictures in the training set too. That’s why you also see some birds, fish, and other strange stuff showing up.
Somewhere on-line I saw some other pictures, produced by a network that had been trained mostly with pictures of buildings. Then, given a picture of a forest, it turned all the trees into pagodas. I’ll come back with a cite if I can find it again.
ETA: Okay, here’s something for starters:
Why Google’s nightmare AI is putting demon puppies everywhere
I’m still looking for the one I saw earlier that showed a whole forest of trees turned into pagodas.
In the meantime, here are some other bits and pieces:
Inceptionism: Going Deeper into Neural Networks, June 17, 2015. Blog post from Google Research discussing a bit about what is going on in training their networks. Shows one example of one tree turned into a pagoda. Also, lots of clouds turned into strange fish, and other examples.
What is #DeepDream and why is everyone getting so weird with it?, June 7, 2015, from Kill Screen Daily. Also some discussion of how it works and examples.
DeepDream reddit page. Everyone is posting links to their pictures here, along with some discussions.
What are deepdream images? How do I make my own? Can I do audio/video? Why are there dogs everywhere?! Discussion of the techniques, instructions how you can make your own, why there are so many dogs.
Here’s the one I was looking for. It was trained on landscapes. This doesn’t show what the source picture was, but the result was a bunch of pagodas:
https://twitter.com/Synapticism/status/618490397958868992/photo/1