There is definitely a recognizable style of NN-generated art. It looks "piecewise-consistent". After a few years of seeing these I find them really boring and unpleasant to look at.
I think AI is currently able to generate images I find pleasant to look at, but they have no impact on me beyond that momentary enjoyment. It is still decoration and not yet art, according to my personal definitions.
When I can tell an AI to make an image about the Spanish Civil War and it produces Guernica, then I will be impressed in the way human artists impress me. And then we will have a powerful new tool to communicate. I would like my own personal Guernica each day to help me learn about something happening in the world.
But how much of that is context? Is the absence of deeper impact because you're aware that a soulless NN produced these, or is it something intrinsic in them?
I'm not at all an art expert, but I definitely looked at several of them and could easily picture a critic describing the significance of the relationship between the shapes and colours.
I feel the same way about many human-produced works, regardless of critics' views, so the fact that I'm aware these are produced by NN is probably not the only explanation.
Yeah, it sounds like in that case your criticism may be more about abstract art as a style than it is about whether a given piece was produced by a human or a machine.
This doesn't sound like it would be unrealistic. I think this can be partially solved by two systems: one that paints and one that finds the best painting to match a current event (in the form of a topic). The latter system could be trained on a set of paintings and corresponding topics or meanings, and then uses this training to match newly generated paintings to topics/meanings.
There are definitely people working on this sort of thing. I recently read this conf paper on generating modified flags based on a topics (e.g. Cyprus + peace):
Correct. This is actually quite doable with word embedding or image classification of a topic, semantic analysis, and then providing those as input to a network like the one I used for 1SecondPainting. I love the future.
Judging by the URL of the images themselves these seem to be pre-generated, which makes me think there might have been some human selection process to filter out the badly generated ones.
Still some really cool looking art in there, though. I'd be perfectly happy hanging some of these up in my apartment.
I would say that the part you quoted suggests it's not generated on the spot.
However,the sentence in the huge font right before the one you quoted could be reasonably interpreted to suggest it's generated on the spot. And as you mention it just looks like a gallery of images that was already generated.
It's still cool, but I'm a bit irked at the misleading framing. Gwern's website showed us amazing text generation from GPT-3, that had all been generated beforehand. We didn't have to be tricked into thinking it was generated on the spot to appreciate it, so I don't see why that's needed here.
The fact that the link says “Try Now” instead of “next image” or something like that is implying that you are using the AI generation when you click the button.
The fine-print at the bottom explains it: "On a Tesla T4, it takes on average 0.173 seconds per novel generation."
I don't think you can expect someone to pay for T4 instance 24/7 for some hobby project.
>I don't think you can expect someone to pay for T4 instance 24/7 for some hobby project.
That's a needlessly uncharitable interpretation of what is happening here. The site is giving the misleading impression that these images are generated on the spot, and then walking it back in the fine print.
It's a valid criticism, and lecturing people about the economics is not a charitable engagement with that criticism.
That seems backwards. I think GP's point that the critics on HN are themselves being uncharitable towards the site (or maybe just missed the explanation) by acting like it's a nefarious bait and switch, rather than considering the possibility that the author of the site is just trying to save on computing costs.
He or she's not lecturing anyone about economics. In fact, if we want to talk about charity, I think that accusation is about the least charitable thing I've seen so far on this thread!
I'm not seeing how either of those interpretations are accurate. The page title and body text say one thing, the fine print walks it back. The balance of emphasis definitely puts forward the impression that images are being newly generated.
And the commenter most definitely was criticizing a very strawmanny 'expectation' that they pay 24/7 to serve an image generation app, which is projection of a view that wasn't expressed by anybody and is certainly not the most charitable reading of what people mean when they say the message creates a misleading impression.
For one example, a more reasonable 'expectation' would be that the language be changed to say it's a gallery of images already generated. Which is different from the uncharitable and unreasonable assertion that everyone is expecting them to pay to maintain a server.
The title literally says "generate abstract paintings in one click" and it is not doing that. I clicked on it expecting it to generate abstract paintings in one click, and now my disappointment is uncharitable?
First of all, my post was quite clearly interpreting a prior comment, not making a claim of my own about whether anyone should be disappointed.
But since we're here: the gnashing of teeth on this thread about having been "mislead" does seem to me to be a bit out of proportion. This person really does seem to have developed a cool toy that uses AI to generate convincing abstract art, but many here aren't saying about it because they object to some of the wording on the landing page!
It's their own fault, of course, for writing the page the way that they did. But still...
I think people are just making a normal, perfectly correct and reasonable observation that the description was misleading, because it really was.
But the hallmark of many internet comment threads is to try and get additional mileage out the conversation by subjecting said reasonable observations to the ritual exercise of switcheroos, contrarianism, idiosyncratic distinctions and unusual interpretations. Which leads to the original wisdom being repeated, which makes it seem like it's being blown out of proportion.
But I think the simpler explanation is just that it's a correct observation and that it's not that complicated.
Fair enough. FWIW, when I originally weighed in, this conversation about the description being misleading was way up at the top of, and seemingly dominating, the entire thread. But I'm glad to see a bit more discussion about the actual work up top now.
No. They could've just said something like "get a new abstract painting in one click" and used phrases like "never before seen". But as-is, their claim is just not true.
Still no, since you’re not _generating_ the image at the point of clicking.
That matters for two reasons, first because words and semantics are important, clear and correct communication is important. Second, the technology required to generate new paintings (cheaply and quickly) on demand is different from the technology required to generate a very large (infinite) number of paintings up front. They’re both interesting but for different reasons, and one shouldn’t be misrepresented as the other.
This is correct. I agree with many of the commentators below - I could have been clearer. I simply didn't want to pay for a 24/7 GPU to generate on-the-spot (a little short on funds at the moment), instead opting to replace each series of 10,000 paintings once per month. Glad I could provide enjoyment to as many people as I did in the meanwhile!
That is a good insight which is not so obvious. The author should have written a white-paper for more credibility.
But I would not be surprised if the actual generation of a new painting takes less than a second as generative networks can be run quite fast to generate new image. I assume the author had the challenge of creating this new "generated" content on-the-fly in browser from the trained model and hence just loads from already generated thousands of images for convenience of the user-interaction. But yes it is a little misguided approach for Karma.
Close! I used anime to test the basic functionality/UI of the site. Transfer learning was in fact done on the FFHQ512 dataset from Nvidia (I figured the complexity of its features was a little higher, which would lend itself well to abstract textures).
Actually, I'm pretty sure those are generated by a neural net. There are clear artifacts on the images that give it away. Maybe the author experimented with different source material before going for abstract paintings.
I got one which was just an anime face (not similar to, an actual professional quality animation) which leads me to believe it's passing through images it trained on (or making extremely minor changes to them).
Congratulations on the project!
This looks very similar to my project: https://art42.net
I've used the higher resolution model 1024, and I also chosen to generate the picture in advance. It's expensive to generate them realtime.
This isn't live/dynamically generated, but if it was, what would be a good way of architecting it in a basic implementation? Say you keep the WordPress site, do you then just send a request to some endpoint that is served by python which keeps the model in memory for quick responses?
When I was doing AI lyrics, I just ensured that the same seed generates the same text. So server just pregenerated several thousand texts into a queue, which you drain from instantly upon a click to get random.
The generated texts are then cached, and if not visited for a while - expire. But thanks to the ability to regenerate the same text from seed are still accessible from URL.
Very cool! Wonder how the copyright works on this sorta stuff. Since you fed your AI with actual images of other people art... but a art student would study other people’s art too. So not sure if AI would be seen as a copy or a derivative work. Seems like if you were designing a game and needed random art to fill frames on a wall could be cool to use these. But I feel like the legal part of doing that might be a little murky and unclear and probably even varies by country since still a very new technology.
Definitely something I considered while feeding it a tremendous number of famous paintings (Rothko, Pollock, etc). Considering the nature of random seed generation/traversing through latent space, I don't see a strong legal case to be made for infringement. The law has done crazier things, though, so I'll keep an eye on it.
This appears to reuse NVIDIA's StyleGAN network, just like artbreeder.io which was previously on HN. If I remember correctly, NVIDIA's terms don't allow non-research use.
As a digital abstract artist who also does generative art, I find making the tiny images sort of funny. How about making a 30x30 inch painting? If it takes 1 second to make a tiny 250 pixel image how long would it take your system to make a 9000x9000 image?
Certainly a good point. Unfortunately, the output size is constrained by the training data (which in this case was ~800x800px). I think it would be kind of a pain to find 14,000 9000x9000 images, though, so I don't know if this will ever be able to fulfill this requirement. ;)
I'm curious to know from the people commenting on this thread that they would buy one of these generated paintings whether they would still be interested if the paintings were artworks made by people. I'm guessing that there are two distinct markets here.
2 is a very small number. Peoples personalities and needs vary wildly and are ever changing. And there will be some artist sitting somewhere catering to some unimaginable combo. Thats how we end up with a gazillion Youtube videos.
On the otherhand look at Netflix (or Tiktok though I am less familiar with it) where there is more sophisticated feedback loop deciding what content get produced. They are generating Art, so much Art that personally speaking, I can barely remember who the creator was or what the creators story is. And I have no problem paying.
Yet obviously there are a whole bunch of people very very different from me, part of huge fan clubs and movements built around connection with the creators and their narratives, much like sports fans, political junkies and religions people.
So who knows, we might get a cult or two or seven hundred following different Algos soon.
The number of those cults (and markets) possible is dependent on the personalities and needs distribution of the population. Zuckerberg probably can generate the data. But it changes like the weather everyday. With love, with war, with age, with depression etc etc etc. Where Art is concerned its definitely more than 2 though.
If i could get large enough resolution to print something on the order of 20-30" and it still look good i'd spend an hour mining through to find good stuff. Most of them aren't interesting but there are a few that are very pleasing to look at. Plus the backstory is interesting.
Why is this downvoted? Chained and piped AIs will probably be the next step towards building ever more powerful AI systems. There is even a framework to chain Machine Learning outputs and use other systems to improve: https://singularitynet.io/
Absolutely! I was planning on providing this, but wanted to keep the model free of commercial services. There are open-source alternatives out there, though, which I may add support for in the future.
Neat. I made my own abstract painting generator a while back, and is admittedly much more low-tech and produces mostly similar results. It doesn’t pull from a set of pre-existing images either ;)
That really depends on how you create them. If it's a print then no one would care and it would likely fade out after a few decades. If it's painted on canvas then you've just made yourself a painting so might as well sign your own name :)
It made me think about couple years old trend of style transfer using GAN: "make your selfie look like Van Gogh". Generating abstract painting may be reduced to transfering average style of several abstract painters onto random seed image. But then it should be a trivial task, even possible to do client-side, no GPU.
Gorgeous! Would absolutely pay for most of these. Best thing though: this is the first software I've ever seen that advertises "in one click" yet really requires only ONE CLICK.
I would like to know how people are judging these painting. I have no knowledge or understanding of abstract paintings. To me the only criteria seems to be does it look "organic", but organic to me is just "not completely random".
As Marshall McLuhan said: "Art is anything you can get away with." (and Andy Warhol later quoted him).
Art is cultural phenomenon and modern art is even more so. You have to study painting and modern art to be able to judge it in context.
In the quality of modern painting is almost completely path dependent and not absolute in any way. Some of those cheap decorative paintings you see on hotel wall would have been masterpiece for artist who sees the style first time in 1920's. Today art critic would not look twice at them because they are just lazily copying style.
Modern art scene is just like any scene. Just because it's usually done by adults and rich people are attracted to it does not mean it's any more or less valuable, or more or less sophisticated than some other scene.
The novelty factor (1920 style looks cheap now because it is copied) is the gamication engine of this scene.
You have to be told this is good by an accepted thought leader who was accepted by previously thought leaders. Meanwhile everyone stands around and pretends to say something but are really trying to copy others.
