To avoid misunderstandings: Borne Sulinowo - town in Poland which during the Cold War housed large Soviet/Russian military base and therefore was erased from maps [0].
First, you have to collect a few thousand images of the same thing (maybe more or less depending on how complex your thing is or how good the results should be). Then, you train a generative adversarial neural network on those images to generate new images. https://github.com/NVlabs/stylegan2-ada-pytorch works quite well. https://github.com/NVlabs/stylegan3 is supposedly even better, but I did not try it yet.
I would suggest getting a colab pro account. If you only use it for a month that's only $10 but it's plenty for training lots of these models (maybe 1-3 days per dataset depending on resolution and dataset diversity). https://colab.research.google.com/github/dvschultz/ml-art-co...
As with all these "this X doesn't exist" sites, the images look ok at first sight, but when you start to look closer you notice that some things aren't quite right. In this case, it's mostly roads that don't seem to connect to anything...
As someone spending time looking at satellite and orthographic photos, i agree, this looks weird. City grid looks like cloth texture, roads and rivers make no sense. In any case, it manages to slightly impress me.
It's like it can't decide between a city on a grid or not, and instead ends up with a weird fusion of both - and like you said - looks like a woven fabric instead.
How does one get an orthographic photo that is not artificially made? I know what an orthographic camera does in 3D graphics, but how can you accomplish that in a real life?
they are "artificially made" by postprocessing aerial images. (You can actually do cameras that create orthographic images, but that's more for industrial vision etc since the frame covered obviously can't be larger than the lens, which doesn't really work for aerial photography, so if people talk about "ortophotos" in that sense they mean processed)
The low resolution to me undermines the entire thing. On some of these images I couldn't even tell if there was even a city or just a barren landscape with a handful of warehouses
I wonder if you could improve this GAN by taking a cropped part of each generated image and trying to discriminate against that part of the image as well.
Because the problem seems to be in the details. The GAN is great at the global picture but the details are off. And this problem benefits from the fact that a picture of a section of a city should also be recognized as a picture of a city.
Back in the day, ~1995'ish I was a tech supporting Windows NT4 (amongst many other things) and I had a customer with a beefy dual processor server living on premises in a closet. For some unknown reason NT would slow down and overall performance was a bit rubbish. It turns out that the customer had turned on the pipes screensaver and it basically soaked up all of the CPU time. Fortunately my customer was more agreeable to disabling this compared to this similar story from those days of yore:
Unfortunately the Dall-E demo (openai.com/blog/dall-e) doesn't have useful deep links nor let you choose free-form prompts, but if you modify one of the existing text prompts to "a snail made of faucet", it generates some pretty realistic pipes--the plumbing variety--that I'm pretty sure do not exist.
Fascinating how the replies are about the plumbing pipe variety, while the link above is clearly to the smoking type, which was what I meant. I guess you can use this one to generate that though.
Unfortunately, the joke misses (or misses me) a bit.
But it does seem like there is a critical distinction between 'this is not a pipe' and 'this pipe does not exist'.
The original targets the distinction between the image and the 'reality', the difference between the map and the territory and our linguistic/perceptual confusions as to such.
In this case however, this 'pipe' does exist. It can't be said to exists in terms of reference to an object outside of the image, but the image itself is as extant as any image. We could argue about issues of permanence---what happens when I turn my device off and the data representing such in memory and on the screen evaporates?, etc---but likewise the universe (at least according to current scientific understanding) is likely to 'evaporate', so in that regard the 'real' world is just as permanent as the image on a screen that is only a temporary electrical phenomena interacting with human perception.
Maybe I've missed something or am overthinking it.
(And yes, it's just a joke website, but if a joke can't lead one to spiraling into existential despair at the nature of existence, what's the point?)
I think you’re overthinking it :). The original meant "this [depiction of a pipe] is not a pipe", and the website is saying "this pipe [depicted here] does not exist".
A picture of an imaginary pipe means the pipe it represents does not exist.
At the same time, a picture of an imaginary pipe is not itself a pipe.
They are two separate statements that aren't connected to each other - it's just that they are both satisfied independently by this website, so it's a happy and fun coincidence.
This is Not A Pipe That Exists, which returns some AI convolution of many pipes with some heavy weighting toward the Magritte pipe would be ideal for me.
It's ok, but it never takes any bold risks putting in any distinctive "this landmark does not exist" buildings. It would never generate a gaudy pile of lights like Time SQ nor an oversized oyster shell like the Sydney opera house. I'd be fairly impressed if it generated distinctive features that feel like deja vue.
This is cool! There has been a bunch of these lately, is there some more or less comprehensive list of *doesnotexist websites somewhere? Would be fun to browse through them all for inspiration.
https://thisrentaldoesnotexist.com/ was hilarious. The bedroom of one looked like a squalid Soviet dorm-room, while the kitchen of the same rental looked like it was from a $100m megamansion.
This one looks like a graphical Markov chain generator.
At this level of detail it's hard to tell if the generated city is implausible, because cities are shaped by many factors that aren't visible on the map (history, soil, location within larger region, economic and social policies, etc.) So a clump of buildings near a river avoiding mountains is a good guess every time.
Y'all need to stop making stuff like this because when we go into the dark ages or humanity is wiped out, future people or aliens will be trying to find things that don't exist.
You mean, things like humanity? On a more serious note, I think it's kind of cool that archaeology has sometimes worked in kind of the opposite way; everyone assumed the Trojan War was just made up, and then they found Troy.
I've started to think that the plausibility of recovering a lot of the digital data we're creating in the case of a nuclear winter might be really low. Imagine walking a wasteland of abandoned computers in a fallout-style land. The vast majority would be behind passwords and even when it wouldn't, the internet wouldn't be accessible. Raid serverfarms? They'll be torn apart for metal or even if not, behind even stronger encryption than regular computers. Feels like our accumulated knowledge is a lot more fragile than we acknowledge.
Off-topic: @dang, was the title automatically stripped of a leading "This"?
Off-topic^2: @dang, If you're reading this, what is the proper way of notifying you of minor stuff like this (if at all)? And how exactly do you see these comments? Do you just read ALL comments, or is there a system that detects mentions or stuff like that?
Sometimes the HN algo will strip what it thinks is non-essential leading words from submissions. When it does that, the contributor can edit the title to replace it ... and the algo won't argue.
I thought they were going to play on the theme "this X does not exist", then slam your emotional cords by showing you images of Ukrainian cities before and after shelling. Fortunately, I was wrong.
93 comments
[ 3.5 ms ] story [ 162 ms ] thread[0] https://en.wikipedia.org/wiki/Borne_Sulinowo
https://en.wikipedia.org/wiki/Generative_adversarial_network
My follow-up question is whether anything will work with an M1 yet. I'm guessing nothing from NVlabs. :-)
https://pytorch.org/docs/stable/backends.html
Whether or not it would be prohibitively slow is another question, however.
That's presumably the same training the AI had :p
Because the problem seems to be in the details. The GAN is great at the global picture but the details are off. And this problem benefits from the fact that a picture of a section of a city should also be recognized as a picture of a city.
An oval mouse mat with a beautiful illustration of a mouse sitting on a mat, with "Ceci n'est pas un mousemat" as the caption.
https://www.pinterest.co.uk/pin/157766793165878650
Apologies for the pinterest link; not many photos of this. And even this one is (like mine was) discoloured from decades in the light.
C'est la vie.
https://thedailywtf.com/articles/A-Fat-Pipe
Examples: https://cdn.openai.com/dall-e/v2/samples/animal_concept_tran... https://cdn.openai.com/dall-e/v2/samples/animal_concept_tran... https://cdn.openai.com/dall-e/v2/samples/animal_concept_tran...
[1] https://en.wikipedia.org/wiki/The_Treachery_of_Images
But it does seem like there is a critical distinction between 'this is not a pipe' and 'this pipe does not exist'.
The original targets the distinction between the image and the 'reality', the difference between the map and the territory and our linguistic/perceptual confusions as to such.
In this case however, this 'pipe' does exist. It can't be said to exists in terms of reference to an object outside of the image, but the image itself is as extant as any image. We could argue about issues of permanence---what happens when I turn my device off and the data representing such in memory and on the screen evaporates?, etc---but likewise the universe (at least according to current scientific understanding) is likely to 'evaporate', so in that regard the 'real' world is just as permanent as the image on a screen that is only a temporary electrical phenomena interacting with human perception.
Maybe I've missed something or am overthinking it.
(And yes, it's just a joke website, but if a joke can't lead one to spiraling into existential despair at the nature of existence, what's the point?)
That doesn't exist. :-)
Edit: TBH I'm not sure why they put that picture of a pipe there, just a bit of decoration I suppose. :-D
https://en.wikipedia.org/wiki/The_Treachery_of_Images
My artist girlfriend shared this with me once, and I'm happy to be able to pay it forward. haha
A picture of an imaginary pipe means the pipe it represents does not exist.
At the same time, a picture of an imaginary pipe is not itself a pipe.
They are two separate statements that aren't connected to each other - it's just that they are both satisfied independently by this website, so it's a happy and fun coincidence.
https://m.xkcd.com/2585/
At this level of detail it's hard to tell if the generated city is implausible, because cities are shaped by many factors that aren't visible on the map (history, soil, location within larger region, economic and social policies, etc.) So a clump of buildings near a river avoiding mountains is a good guess every time.
Deniable cities, so to speak.
Some shapes that seemed like roads but had small river deltas fanning out.
Off-topic^2: @dang, If you're reading this, what is the proper way of notifying you of minor stuff like this (if at all)? And how exactly do you see these comments? Do you just read ALL comments, or is there a system that detects mentions or stuff like that?