Author Here: Just wanted to share some technical details. The code can be found here https://github.com/valentinvieriu/stylegan2
The main code and modification were put together by https://github.com/pbaylies/stylegan2.
I've started with the pretrained ffhq model, and I've trained it using come hand picked 1000 images of cubist Wikiart paintings.
I think you have an opportunity here --- I found myself wondering how much it would cost to get a nice-quality print on canvas or metal. Something I could hang in my office. They look that good. Are they (or could they be) generated in sufficiently high resolution for a large format print?
Hey this is an awesome project with really cool results. I think it can be grouped with "found art" conceptually. Several of the generated pieces are really quite interesting, though at some point the styles seem a bit repetitive, though I imagine that's simply due to training on cubism.
Great idea. Looks very convincing and aesthetic to me, but I am not an art expert. But I do know a lot about music and I am not yet convinced of the current offers of AI in this regard. So it would be very interesting to know how art experts react to these paintings.
Really great project. I'm interested in how you've implemented the back-end? How did you productionize the GAN model? Or did you just pregenerate 1000 images? (ie. are the images generated on-demand?)
Very cool. Some of the results reminded me of Max Ernst’s frottage works and futurist compositions. If you are interested in exploring this with other curated collections, reach out to me. I have a MOMA collection of about 10k images.
This is awesome! I would hook this up with Shopify and one of those printing and shipping services and see where that leads you.
I'm a Shopify expert, already did that for customers, if you need advice let me know
I would love to see TensorBoard learning curves from your experiment: score/real and score/fake. I've tried fine-tuning FFHQ-model before and training was not progressing as I expected it to.
Not trying to denigrate the amazing technical skill needed to make this thing, but from the point of view of someone who makes art, this is a pastiche image maker. I have yet to see AI generate anything that's not a combination of existing images and techniques.
Probably wading into murky philosophical waters here but I reckon art needs to be more than a nice 'arty' looking image to be considered any good. If you are going to make Art that is worth the capital 'A' on the front, then it has to mean something.
However I suspect the days of actually making Art are in the past. We're on to something else now. Not necessarily better or worse, just different. Art is an old idea, not entirely relevant any more.
This is indistinguishable from something that would be hanging in an art museum, and anyone pretending otherwise only believes so because they know a priori that it is generated by ML.
Honestly I would consider buying something like this, though I'd never waste my money on something so pretentious from a human. This, however, is an achievement, and there's beauty in that.
Edit: in fact, if love to see two tests: a blind test with a series of humans, and a classifier trained to differentiate between machine and human generated art. I have no doubt that the humans would not score much better than chance - I bet only an artificial discriminator would be able to tell, and then only because of subtle differences in pixel distributions that humans won't perceive. Throw in a couple different GANs trained on different distributions and even a machine will have trouble telling the difference!
"This is indistinguishable from something that would be hanging in an art museum."
Haha, the exact opposite of my scroll through. I thought to myself: "These look like a clone stamp sampled and then vomited up all the lovely modernist paintings I've seen at art museums."
This is indistinguishable from something that would be hanging in an art museum, and anyone pretending otherwise only believes so because they know a priori that it is generated by ML.
I certainly couldn't distinguish an individual images from something I might see in a museum. At the same time, scrolling the images felt very different than scrolling through a series of images by a given artist or images from a given show.
This is what an image search for cubists and wikiart looks like. Also feels very different.
Which is just to say this sequence feels different. I couldn't tell if it's better or worse. It does feel bleak, opaque, oppressive, washed-out. All that is something some art aims for. But it could just be the randomness of the pictures, like they're well-spoken words in a language I don't know.
Whilst these all the samples seem to have decent form, use of colour and space, one can quickly see derivate concepts and images from more famous original works and an excess of 3D forms. There's also a challenge in understanding the meaning of each painting. Also, a lot of penguin type shapes for some reason.
One of the tests you describe - a classifier trained to differentiate between machine and human generated art - is exactly what the ML GAN model is doing.
The generative adversarial network (GAN) has a generative part and a discriminative part which compete with each other. The generative model creates the images that you see in the gallery. The discriminative model tries to predicts whether an image is real or was generated. When it starts failing often, we know that the generative model is getting better (of course, the details are much more complicated).
I don't it's fair to call it pretentious if from a human but beautiful if from a computer. You even said it yourself
> ... anyone pretending otherwise only believes so because they know a priori that it is generated by ML.
To me, a large part of an artworks beauty comes from its ability to make me feel something. These works make me feel nothing at all. So I cannot call them beautiful.
The further down the page I scrolled, the more I got a kind of despair. To realize that the space was infinite, and I would never reach its end. Each individual image is pretty cool. As a collection, they're sad.
What's funny is that I then tabbed back to HN, where I was deep down the comments tree, and momentarily I forgot that real humans make these comments. I thought "I shall never reach the end of the HN comments. It's all AI - generating a pastiche of plausible opinions and sentences-that-appear-sentient." Same despair. Quite a relief, then, to get to the end of the page.
Interesting, so would you say that the infinite aspect helped out the experience or damaged it? I've had doubts first when I've introduced it ( more from caching perspectives and data transfer costs ). But then I've wanted to showcase this amazing feature of those networks of generating an infinite amount of unique items, and also offer an unique experience to every user, the possibility of seeing unique art and bookmarking / printing it if considered
I get the appeal from a "it's fascinating we can make computers do this" stance.
I don't get the appeal as trying to appreciate this as actual art (or music in the case of ai-generated tunes).
Art for me is about the connection I can make with the person who made it. I can't make a connection with an algorithm. I view it and I am like "OK this is interesting" but that's about it. There's no desire to try and understand it because it was programmatically generated.
Given that this is the product of human written code trained on art generated by other humans, you are seeing something that’s an amalgam of what humans have done.
You can’t take away the human and get what you have here.
What about connecting with the person who wrote the algorithm? Imagine a time in which music was only heard when performed by the right group of people. Without that the appreciation was mostly for a cultural memory or written record of the music. There are more ways to connect with art and it's maker than playing back the audio or opening the image.
Is this a portent of truth's final days? Is absurdism the only final, knowable truth, in a world where only fact carries meaning, as humanity has been deconstructed and rendered noumenal? The voices of a million potential future artists just cried out in agony and were silenced forever. The Tesla CEO has it right. AI is the end of humanity -- but (probably) not in a SkyNet sort of way, but one that is much worse because it renders the term "hope" itself meaningless.
To my untrained eye, many of these images are as good as anything else I've seen from humans. What qualities make me say that? Your images create a visceral response. They evoke a lot of interesting thoughts and feelings. If anything they remind me of 20th century modern art. Very impressive stuff. It'd be amazing to make these bigger, ensure texture is there, print them large like an oil painting.
I have been studying art non-professionally for the past 5 odd years now and I would be hard pressed to tell you if any one of the pictures your software generated was created by a human artist or not.
It reminds me of those time slice type photos, where the bottom is when the little 'tree' on the right is young and each layer is a new year or three along the way, in which the environment gets a bit older and gnarlier over time.
Site appears to be down at the moment. Some shameless self-promotion since there appears to be some interest in this, I recently trained a StyleGAN model on Fauvist artists:
http://www.sanbase.com/ - these were created long before AI became a household name. Actually no AI involved at all. All just pure math. Think it is called "generative art" or something to that tune
generative art is of a completely different tune...
even if for this style they might not be that different results, the neural net(s) here (the "AI") can have details and textures simple math cannot, such as I saw a horse head in one, some other things in others...
but it's still very cool, even if the site may or may not be slightly deteriorating (top banner's broke for me, wouldn't be surprised if some other stuff was too)...
They look great. If you make something where people can pick say paintings/styles they like, and then you generate custom for them until they choose one (and presumably then un-watermark or high res it), you may have a business here.
Ok. What’s the point of that, original poster or canvas print? In a way training AI on existing art and generating some more amounts to vulgarizing abstract art or art in general and doesn’t add much to it, it is a gimmick. It certainly looks attractive but does it have any real value?
Reminds me of Debris by a NullSoft employee. It would download random pictures using a search engine, and it would craft it together. The end result is unique, and allowed you to look into the details. You could even edit the search terms or the result, add cameos (some vague nudity, for example). What I found really funny though, is that the end results often had actual funny easter eggs. I guess if you add up all of this randomness, and you watch closely, the chances of something funny being included increase. It just might take a while till you find it.
I love cubism and this plain sucks. You see clearly that is not human work, there is no subject, there is no meaning, there aren't emotions...is garbage.
[+] [-] valentinvieriu|6 years ago|reply
The frontend is Vuejs and it's using the amazing https://github.com/Akryum/vue-virtual-scroller for the very fast infinite scroll option. The app runs on Cloudflare Workers, using https://github.com/l5x/vue-ssr-cloudflare-workers-template
Hope you guys enjoy the experience and would love to get some constructive feedback
[+] [-] rfeague|6 years ago|reply
[+] [-] adzm|6 years ago|reply
[+] [-] Rochus|6 years ago|reply
[+] [-] willvarfar|6 years ago|reply
Is it that that's the style of art you like?
Or that you found to work best?
It would be really interesting to see the kind of images generated when its trained with other styles.
[+] [-] kernelsanderz|6 years ago|reply
[+] [-] Jupe|6 years ago|reply
Who decides what gets shared? Are you posting everything, or do you curate the results of the algorithm?
[+] [-] ricg|6 years ago|reply
Is it a random number or another image or something else?
Let's say you wanted to re-render the exact same image given the same state of your pre-trained model, what information would you use here?
[+] [-] yantrams|6 years ago|reply
[+] [-] joantune|6 years ago|reply
[+] [-] dimitry12|6 years ago|reply
[+] [-] iandanforth|6 years ago|reply
https://github.com/reiinakano/neural-painters-pytorch
[+] [-] gwd|6 years ago|reply
[+] [-] kengor|6 years ago|reply
[+] [-] allovernow|6 years ago|reply
Honestly I would consider buying something like this, though I'd never waste my money on something so pretentious from a human. This, however, is an achievement, and there's beauty in that.
Edit: in fact, if love to see two tests: a blind test with a series of humans, and a classifier trained to differentiate between machine and human generated art. I have no doubt that the humans would not score much better than chance - I bet only an artificial discriminator would be able to tell, and then only because of subtle differences in pixel distributions that humans won't perceive. Throw in a couple different GANs trained on different distributions and even a machine will have trouble telling the difference!
[+] [-] justanothersys|6 years ago|reply
Haha, the exact opposite of my scroll through. I thought to myself: "These look like a clone stamp sampled and then vomited up all the lovely modernist paintings I've seen at art museums."
[+] [-] joe_the_user|6 years ago|reply
I certainly couldn't distinguish an individual images from something I might see in a museum. At the same time, scrolling the images felt very different than scrolling through a series of images by a given artist or images from a given show.
This is what an image search for cubists and wikiart looks like. Also feels very different.
https://www.google.com/search?tbm=isch&sxsrf=ACYBGNRYdV4hOBO...
Which is just to say this sequence feels different. I couldn't tell if it's better or worse. It does feel bleak, opaque, oppressive, washed-out. All that is something some art aims for. But it could just be the randomness of the pictures, like they're well-spoken words in a language I don't know.
[+] [-] staplers|6 years ago|reply
[+] [-] flarg|6 years ago|reply
[+] [-] tomerv|6 years ago|reply
The generative adversarial network (GAN) has a generative part and a discriminative part which compete with each other. The generative model creates the images that you see in the gallery. The discriminative model tries to predicts whether an image is real or was generated. When it starts failing often, we know that the generative model is getting better (of course, the details are much more complicated).
[+] [-] null0pointer|6 years ago|reply
> ... anyone pretending otherwise only believes so because they know a priori that it is generated by ML.
To me, a large part of an artworks beauty comes from its ability to make me feel something. These works make me feel nothing at all. So I cannot call them beautiful.
[+] [-] nathell|6 years ago|reply
[+] [-] TheGallopedHigh|6 years ago|reply
[+] [-] mattsahr|6 years ago|reply
What's funny is that I then tabbed back to HN, where I was deep down the comments tree, and momentarily I forgot that real humans make these comments. I thought "I shall never reach the end of the HN comments. It's all AI - generating a pastiche of plausible opinions and sentences-that-appear-sentient." Same despair. Quite a relief, then, to get to the end of the page.
[+] [-] valentinvieriu|6 years ago|reply
[+] [-] shironineja|6 years ago|reply
[+] [-] anigbrowl|6 years ago|reply
[+] [-] kangnkodos|6 years ago|reply
[+] [-] alharith|6 years ago|reply
I don't get the appeal as trying to appreciate this as actual art (or music in the case of ai-generated tunes).
Art for me is about the connection I can make with the person who made it. I can't make a connection with an algorithm. I view it and I am like "OK this is interesting" but that's about it. There's no desire to try and understand it because it was programmatically generated.
Does anyone else feel this way?
[+] [-] ringzero|6 years ago|reply
Tech is more of a, “why is a painting of a triangle worth so much!?” crowd (see: other comments).
[+] [-] jcims|6 years ago|reply
You can’t take away the human and get what you have here.
[+] [-] qmmmur|6 years ago|reply
[+] [-] dwohnitmok|6 years ago|reply
[+] [-] Rochus|6 years ago|reply
[+] [-] zsz|6 years ago|reply
[+] [-] dsco|6 years ago|reply
[+] [-] dvh|6 years ago|reply
[+] [-] starchild_3001|6 years ago|reply
[+] [-] alainchabat|6 years ago|reply
There was a recent Show HN project about generate quotes having some great results having sense : https://machineswisdom.com
[+] [-] flyGuyOnTheSly|6 years ago|reply
I have been studying art non-professionally for the past 5 odd years now and I would be hard pressed to tell you if any one of the pictures your software generated was created by a human artist or not.
Nice work!
[+] [-] jcims|6 years ago|reply
https://vcloud42.com/file/art42-cdn/cubism/seed_0000086582.j...
It reminds me of those time slice type photos, where the bottom is when the little 'tree' on the right is young and each layer is a new year or three along the way, in which the environment gets a bit older and gnarlier over time.
Absolutely wonderful work.
[+] [-] gdubs|6 years ago|reply
http://gregorywieber.com/art/a-walk-through-latent-space-mak...
[+] [-] FpUser|6 years ago|reply
[+] [-] jan6|6 years ago|reply
but it's still very cool, even if the site may or may not be slightly deteriorating (top banner's broke for me, wouldn't be surprised if some other stuff was too)...
[+] [-] valentinvieriu|6 years ago|reply
[+] [-] davidajackson|6 years ago|reply
[+] [-] tacheiordache|6 years ago|reply
[+] [-] Fnoord|6 years ago|reply
[+] [-] ecmascript|6 years ago|reply
I do not appreciate modern art anyway, but I love this since it will for sure upset some artists that think they're talented.
[+] [-] lchiodi|6 years ago|reply