That would be some sort of reinforcement learning, where the reward is the number of likes/views/… .
So one might create thousands of ai accounts which generate prompts based on existing ones. Those are then use to generate images which are posted to the ai accounts. The best prompts are added to the prompt generating dataset (at least for some time). Now the next evolution of the prompt generator can be trained, … .
That’s a reasonable self-improving (at least for the prompt) approach, which is sort of close to what I got from your description.
I don’t think this makes too much sense, but there are other ways in my opinion which would also allow it to come up with sort of “original ideas”.
You are a bit skeptical here. I am convinced that those systems can be made better, such that they don’t just produce remixes without evolution, but they will actually come up with new ideas which will be appealing to humans.
But even with that, I don’t see why art should be dead. It will mean a massive change, but human made art will still not disappear.
How can AI generate new ideas if all we are feeding it is premade art? You can never get more information out than you feed in. AI will only be equal to or worse than the art we feed it. If we depended on AI during the renaissance do you think we could have gotten the art of Picasso?
The only way AI can produce new styles is by pure randomness, which is like once in never.
As I mentioned, it will take a huge hit. The newer generation will find no meaning in art. We are already seeing professional artists giving up, how can we expect children to create just for the sake of it? They will find no interest in our current knowledge of what makes good art- appeal, color theory, anatomy, composition etc
Instead of pure randomness, you can have a guided process where images are generated (for instance trained on aesthetically pleasing images), but only the ones are kept which are pleasing according to another neural network which rates them. Among those, the unique ones are kept and added to the “aesthetically pleasing” training dataset. When this kind of process is repeated, you slowly get more unique training examples that differ from the training dataset.
There is already a thread where I described that in more detail:
Why should newer generations find no meaning in art? They may as well appreciate it even more if it was human made…
Humans spend years studying what makes good art- which frankly is completely arbitrary. This information cannot be pinpointed and made into a to-do list, you can only gain this knowledge through practice. Unless you can somehow give that information to a computer, I don’t think there is any chance of a computer developing new styles
What I mean is they will not be interested in the fundamentals of art. They may appreciate art, but will not understand why something looks good.
As said before, unless you tell an algorithm what actually makes good art rather than providing examples, I highly doubt it will create any new styles. In the other thread, you asked why is consciousness required. A very simple answer is that a human understands what another human would like. You can ask other humans if something looks good, but a style is so complex that this process will take an extremely long time.
As I said, you can’t teach art principles to an algorithm. You can’t write this information down. If you give an alien all the art books and all the artworks created to date, he will still not be able to guess a style that a human would like
I have no reason to believe that all humans would loose the interest in creating art, just because computers can do it. To create art, those people would clearly need to learn the fundamentals of it.
Training neural networks with examples is the only known way that works as of now. It has been shown many times that this is how they can learn underlying concepts. It is not perfect or even quite naive, but it works surprisingly well.
It has been demonstrated many times that this underlying understanding in combination with a few other systems can be enough to come up with new ideas or new knowledge.
A very prominent example for this would be AlphaGo or AlphaFold (2).
I would argue that recommender systems that are used in YouTube/Facebook/Amazon/… are a good counterexample for that. Apparently, learning based systems are better at understanding what humans want in those cases. Sure, they are not perfect or even quite horrible in many cases. But statistically, they make better suggestions (according to the metrics defined by those companies) than the algorithms that have been hand designed by humans who according to you should understand better what humans want. (I hope I am not misrepresenting what you meant here!)
Sure you can - and no “writing down” is needed anymore, not for deep learning. Chess were something you couldn’t teach a machine. Then Go was unsolvable problem. Then protein folding. Now art.
Oh, opting out is enabled by default now? Now that every AI generation program is using the Laion 5B database already? Wow, that’s so meaningful!! that’s my problem with the ArtStation NoAI thing too- it doesn’t matter anymore, the database is already made
Many deep learning datasets which are used in practice don’t come with the actual images, but with links to the sources. Everyone who downloads those datasets now, doesn’t get no-ai tagged images by default (when using this package). It is not unusual for those datasets to be updated and improved, making redownloads necessary.
The datasets are usually provided in this kind of format:
It’s the weirdest argument. Art has always been democratized. Literally anyone can pick up a pencil. Or if you can’t afford a pencil, you can make art with dirt and a stick
You wouldn’t be able to explain that to a kid trying to make “pretty pictures” with AI. I for sure will not be able to explain to a 9 year old what art meant before the AI takeover.
In this case the algorithm was getting instant feedback. In art we are bound not by the computation power, but how fast humans can rate them
Again in these systems too the feedback is pretty much instant and from thousands of users at once. AI companies could hire a bunch of people, but they do not need as they do not care about creativity and originality.
A kid with the goal to get a “pretty picture” has to goal to get exactly that. If that kid is more interested, it might start to dig deeper and get a better understanding over time.
It may just as well happen that more kids get interested in art, because the initial barrier is way lower and they can experiment quicker and discover their interest like that, which could have been missed otherwise.
That’s why my suggestion has a neural network which rates the images, based on existing user feedback, allowing it to judge images immediately. Of course, it would be necessary to still check the results from time to time, to make sure it doesn’t go off on a tangent.
I had a thought but I am not a programmer so it may be nonsense. Do you think that it would be possible to develop a new file format for 3D that would make the 3D data unusable for machine learning? Maybe that would be a better solution than a ‘please I’m asking nicely’ tag on websites that hold your data. I am pretty sure that all cgtraders and so on will be doing with 3D models what artstation and deviantart are doing with the 2D data soon (if that’s not already happening).
We barely have working 3D formats for use right now. The most common is FBX but everything else seems to fight for the second place while being at best 6th place.
