Item — HN Reader

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.

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

12 comments

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?

Working in an office decorated by a bot is the most demoralizing experience I could imagine. I would probably go into an existential rage and start breaking things until someone more physically powerful came and subdued me.

It can be worse you could also work on SAP, like the author (No harsh feeling/flame war ; thank you to the author for this artsy experience ; but I couldn't help but notice this, I believe, relevant context).

To me these pictures are very nice looking from afar, but once you try to look in the details it makes you feel stupid that you don't get it, until you realize that it is senseless, there is no coherence, no point, no soul and it's normal that there is nothing to get from the picture as it's not in the algorithm.

The technique is perfect. But the art side reflection and emotion must still come from the eye of the beholder. It is deeply moving art in the sense it inspire strong negative emotions, the feeling that there are some stronger forces coming to crush you.

One such reflection that these collection should inspire is : "Is this the direction we want to take ?". Infinitely many garbage art stealing attention away from human artists.

Don't get me wrong, I like generated art but it's necessary to situate it in its context. That what makes it interesting. I even believe you can have machine explore thing and discover interesting thing on their own without it being formulaic, but we are not there yet.

Your SAP remarks are funny and somehow true (maybe), but I'm maybe one of the lucky guys. I work for an interesting Labs section inside SAP (https://cxlabs.sap.com/) and we are lucky enough to do research on some interesting topics.

I don't fully agree with you on the purpose of art. Art can create different meanings for every watcher and they could be quite far away from what was the intended purpose ( the message ) of the artist. I would say that main function of art is to ask questions, and also to give you a way to experience life in a way you would otherwise not be able to. In this context I think AI generated art can bring something new, a new layer of reality, a new set of questions, that might not been surfaced because we use brains to create art and brains to interpret it. It could be a new type of input that would push humanity further.

I think that postmodern view on art has fallen out of popularity. It’s silly to abstract away the artist and kinda pretend there may as well not be one, and that the full meaning may be purely in the sensory matter... because this leads to an “everything is art” perspective: everything in the field of my senses at any point is art. Which is preposterous for “art” to tell me that my senses are meaningful because I can ask questions about them and learn something. Thank you, art, for intervening to add a completely vacuous annotation on my experience.

These pictures are interesting though. To me the thing that popped out, without having read background on the project, was how this robot artist is copying modernists like Picasso and possibly Pollock and DeKooning. And doing it along a few themes, with high degree of repetition; or tweaking some small things in each variation. The limitations are glaring and I’m wondering will it evolve and how.

> To me these pictures are very nice looking from afar, but once you try to look in the details it makes you feel stupid that you don't get it, until you realize that it is senseless, there is no coherence, no point, no soul and it's normal that there is nothing to get from the picture as it's not in the algorithm.

My train of thought exactly. Even second-rate/beginner/hobbyist paintings, like those on Binned Art for example, are absolutely crushing these in every dimension.

Is your problem that you know its bot-generated, or that you think there's some flaw in the pictures?

What if you didn't know that the painting was the work of a bot?

What if it was just some abstract painting that someone took enough fancy to to pay to have put on a canvas?

Personally, I couldn't really tell them from the human-generated wikiart pictures that seeded the algorithm. Those human-generated pieces didn't speak to me either.

I'd love to see the algorithm work on some other art style seeds.

I'd say it depends. For me art is as much about the artist behind a specific piece as it is about the actual piece. I agree with the person you responded to. I want human feelings in my art, no matter how abstract, for it to truly speak to me!

Playing devil's advocate: it sounds like you're deluding yourself ;)

Perhaps the Vogels https://en.wikipedia.org/wiki/Herbert_and_Dorothy_Vogel knew the artist? The way they amassed their collection is truly inspiring.

Picaso's wife infamously knew he was having an affair with his muse from glancing at a painting. But did every other Picaso fan who saw that same painting? Did truly know anything about the artist behind the picture?

So the chances of everyone else who appreciates an artist actually truly knowing anything about the actual artist? Complete rubbish, say I!

Take Turner, for example: turns out, the location of a lot of his pictures have been wrongly attributed! How everyone can claim to 'know' the artist and yet not realise that the picture is of Portsmouth or a lake in Scotland instead of Venice or somewhere else seems like they didn't really know the artist after all?

So there's a picture on the wall by Satoshi Nakamoto. You feel that genuine, because it's signed? Then when you get told that its painted by a bot, you feel cheated? Because you felt you knew something about the artist by looking at the art?

Having bot-made art in the office just gets me asking all the wrong questions like "What is this? Why is it here? Why am I here? What am I doing with my precious finite time on earth? Why did they automate the creation of art? Why didn't they automate my job?"

Ironically a human artist who could evoke a response like that would be doing better than most.

Those seem like difficult questions, and maybe we don’t usually want to grapple with them, but they don’t sound like the wrong ones.

One might argue that there is plenty of human feeling in this bot-art. After all, it's trained on art with human feelings. Seems like it might be a very effective way to propagate it. It's a sort-of dispassionate third party who finds the thread common to all of the cubist pieces in the training sets.

Such as a security bot?

Sure, once those things are common place it'll be dystopian; but for now it's just a fun conversation starter.

This sounds like a plotline of a Ligottian short story.

Tbh, with direct printing to (large) canvas being reasonably priced now, this could be an easy sell

Especially if they can be generated at 300dpi+

Well the resolution is good enough.I've gave it a try and looks quite good on a 50cmx50cm canvas. You need to upscale it once or twice with a Super-Resolution AI algorithm

So meta

I've added the option to print now. Feel free to give it a try

In the spirit of the effort, might as well complete the value stream and just hang a big display. Very cool project tho'

You can use a super-resolution tool like https://letsenhance.io/ to upscale the image which might make it better for printing.

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.

I'm not an art expert or an artist, but I spent a lot of time in art galleries and museums forty years ago looking closely at and thinking about abstract art. Some of the pieces here strike me as remarkably good and interesting to look at. If they had been implemented in oils rather than pixels, they would not seem out of place on museum walls.

That would be the next step to close the circle https://twitter.com/sougwen/status/1218672595350032385/photo...

The challenge is that the Latent Space ( the space of all the art pieces of this model ) is so HUGE ( technically is infinite ) that there could be masterpieces sitting there undiscovered. I was aiming to create an interface to allow for crowd sourcing this process of finding those masterpieces. Hope the ui and and easy to pick some favorite process will help

Or... train AI to pick out the masterpieces...and loop.

Well people have to create the training dataset so we're on that path, should the author choose to continue on it.

Yep. Please click on the favorite button :) so I can keep the favorites

Why cubist?

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.

I have other styles too, and I'm planning on exchanging the AI on a regular basis. This was the one that I've personally liked the most. It was also the one that required the least training. I only have a 2080ti, and those models are quite hard to train. If you have recognisable shapes ore faces, you can quickly create some very disgusting art pieces that will make people sick. At this moment abstract seems to be the best solution.

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?)

So the backend is simple, i've chosen to pregenerate 1 mil images at the beginning so it gives you some idea of diversity. I"ve tried to generate them realtime, but my server would have not survived the hackernews popularity today. The gan model is created by me, see in the comments above. I'm planning on bringing up new images on a daily timeframe, keep the one that people choose as favorites and delete the ones that were not popular

Fascinating stuff.

Who decides what gets shared? Are you posting everything, or do you curate the results of the algorithm?

What is the input to produce one of these images (aside from the pre-trained model)?

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?

Yes there is an initial seed that it's needed, and it's baked in the name of the file. So if in the future I'll find a way to improve the quality, you can get the same result by passing in the seed

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.

Would really appreciate it! Please connect with me on twitter @valentinvieriu

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.

Do neural painter next!

https://github.com/reiinakano/neural-painters-pytorch

Interesting idea. I will give it a try to see what is the output. I was looking for something in this direction to make the brush feel more appealing. Thank you

What I think would be interesting would be to add a human discriminator test: Can you tell genuine cubist art from computer-generated cubist art?

What would the point of that be may I ask? Something to do with a Turing like test? Lets not forget that the would be cubist model is actually fed tons of cubist images and all it does is regurgitating that back. Art done by people is somewhat similar but it does take place in a context that which AI could not yet understand.

Well, I'm personally curious. When I looked through this thread several hours ago, there seemed to be two very distinct camps: "This is really cool you should see if you can sell these", and "This is just random stuff mashed together; there's clearly no artistic intent". The first group, it seems, sees no "artistic intent" in real cubist art, and so sees no difference; the second claims to be able to see real "artistic intent" in real art, and to note its absence here.

I have to admit I'm rather in the first camp. I have spent time in art galleries, but not a huge amount; I have opinions about art, but 90% of what I hear said about paintings seems to me to be nonsense.

I know a lot about music; and so far I've always been able to distinguish 100% of computer-generated musical snippets from genuine classical snippets. And specifically, the key difference is exactly what is claimed by the second group -- that computer-generated melodies are just piles of similar musical ideas mashed together, without any apparent intent or direction.

One of the claims of skeptics of art is that there really is nothing there, and that all the "art critics" are just doing a wetware version of GPT-2. If someone trained in art can really distinguish a cubist painting with artistic intent from one generated with this neural net, then it's 1) proof that there really is something there, and 2) it's maybe worth learning about.