Evolving AI Art
Вставка
- Опубліковано 9 чер 2024
- In this video, I explain how #aiart generators like #midjourney and #stablediffusion can be used to simulate evolutionary processes, and explain why this is important and valuable as an exploration algorithm. I also show very good pictures of orangutans, which is perhaps the most important part of the video. Please comment "orangutan good" if you read this, it is a sign of unwavering loyalty.
Links to my stuff:
Patreon: / emergentgarden
Discord invite: / discord
Twitter: / max_romana
The Life Engine: thelifeengine.net/
Links to cool stuff from the video:
Midjourney: www.midjourney.com/
Stable Diffusion: huggingface.co/spaces/stabili...
Pic Breeder: nbenko1.github.io/#/explore
Pic Breeder Paper: eplex.cs.ucf.edu/papers/secret...
Image Library of Babel: babelia.libraryofbabel.info/
Richard Dawkin's Biomorphs: • Biomorphs - Richard Da...
Timestamps:
(0:00) Intro + Explanation
(1:56) Midjourney Evolution
(4:46) Stable Diffusion Evolution
(6:56) Pic Breeder
(7:49) Image Space
(10:00) Exploring Image Space
(11:57) Open Ended Exploration
Music:
• Mindful Horizons (Albu...
• Stepping into the Ether
• Closed Circuits
• Mindful Horizons (Albu... - Наука та технологія
Correction: There are 256 possible pixel values (0-255, zero included), so the image space calculation is wrong. It is actually LARGER than the number shown on screen 🤯
💕 this. IMAGINE.
A GOOD JAMES WEBBEYE 👁️ TO ADD TO IT.
AMAZING 🤩 THUMBS UP !!
This newfangled "ai art" stuff is depressing as hell. It will change what being an artist is, it will throw most artists in the dustbin of history.
Art is now obsolete, worthless, just something a machine can automatically spit out. Non artists will somehow think that are artists, and true artists will be forgotten.
I hate the 21st century. I hate world the new generations are creating.
This sucks, entirely.
fuck ai art!
you forgot the point that this AI work by taking "reference" of the millions of art posted in the internet. It does NOT create by themselves, but has billions of images made by others artist saved and use other art to recreate an image. Many of the picture you showed, I already saw, made by other Artist or even Pictures.
So if you never saw this kind of art, it is maybe, because you never search in the various art community this kind of art.
Sadly like in every business there is one kind of thing (in this case art-style) that is heavy sponsored like in Music. But if you search you find it.
In few words the AI helps you with research of pictures, because you are maybe too lazy to search or because you do not know where to start.
But this software NEVER create, it just replicate.
and it gives you the feeling that you are part of the process...but you are not.
@Red Tide Someone in NY won in court because they turned the images into a product they made, comic book of raw "AI splicing". This is a birth of a new way of creating that even a child could dream up and in the future there will probably be artists specifically creating for the database. Maybe all the money from the previous version can fund those peoples lives.
As an illustrator and graphic artist i'm concerned with the progress of A.I and what it'll inevitably do to my field and to the people working in it.. but you kinda open a door to something beautifull and poetic at the same time. Thank you.
see this is why instead of art, im studying compsci, gotta be on the side of robots once the revolution happens ;D
@@ejsafara456 I'm doing the same for the same reason but i find it an interesting paradox, because we're the ones creating these, surely we will be the last to lose our jobs but at the same time technally we are writing programs that write themselves. so maybe we will be gone faster than we think.
AI don't generate pic's with story/reason behind them, so artists still be welcomed in society.
@@user-ru3ll6fx9v idk man, just hook up the ai to random word generator :x and we already have ai writing articles and such :b
@@user-ru3ll6fx9v it can duh😑
Your videos are awesome, really enjoyed the explanation of search space and evolutionary search as a means to creativity. It turns a seemingly np-hard problem into something solvable via a gradient descent through interestingness. Learned a lot, thank you!
it’s really nice having someone more knowledgeable on a subject like this give their 2 cents. i’ve been messing around with ai art for a while now ever since the back/frontdoor leak for novel ai happened and your video kind of Recontextualised all my thoughts/assumptions about this kind of technology.
I initially subbed to your channel because of your evolution simulation videos but it seems you’re really good at making videos on other topics as well. i wish you’d make more of them or they’d come in my feed more often.
Wow, I am really impressed, well done! These creatures look so different, mysterious and yet real at the same time - and those colours!
Even better, Can't wait for AI to start mass producing and distributing worldwide!
"latent image space explorer" sounds like a job from a sci-fi world
Well, we now have latent 3D NeRF explorer. I think science fiction is not keeping up
@@escher4401 I think that "It's NeRF or nothing" is pretty NEAT 😂
We are living a sci-fi world.
And we realize that's also what artists have been doing.
Mathematics asks: What is the space of imaginable universes? (i.e. patterns made of rules)
Science asks: What are are the properties of this universe? (i.e. Which universe - pattern of rules - do we live in, of all imaginable ones)
Art asks: What changes to the rules can we imagine, and what can we bring back from these alternate universes?
In principle, one can imagine that a simulation of any hypothetical reality exists in the mathematical space of possibilities. I'm not saying all reality reduces to math or can be contained in it - that philosophy ain't so simple - but a simulation, sure. Whether our reality is large enough or has the right properties to even run certain simulations is a separate problem. The point is, the simulation exists in the pattern-space of mathematics. And maybe by accessing alternate realities similar to ours we can learn things about actual reality.
So artists can be said to be accessing alternate realities (at least virtual ones) via the simulation abilities of our mind. The current AI results are the outgrowth of creating and training AI models capable of simulating real images. The actual result is that they simulate non-(immediately)-real images too. Neat.
Evolution as commonly accepted is in principle much the same - it creates organisms hypothetically fit for alternate realities, but only some of them are compatible with our reality as well. So to some degree it reaches into alternate realities to extract information about its own.
These thoughts are a work in progress, but hopefully they might make someone's day more interesting.
@@cliftut but before, only artists could travel into that world but now everyone can.
What a cool video, I really liked the part about exploration, I've been thinking a lot about life as exploring one value of the "life space"
Did not expect to be hit with the powerful philosophical analogies at the end. Bravo. Really provocative ideas, well delivered script as well!!
Bbbqvvvbbe av
V
C es. W n. No vj
Dawkins 'Blind Watchmaker' and its biomorphs blew my mind in the 80s, thanks for crediting the link from todays tech back to that. The related term 'genetic algorithims' seems to have been overtaken, but I find it still a useful one.
This was hands down the most exciting video I have ever seen. I love how maths and art combine in this video to search for beauty
One of the most inspiring videos I have seen. Thought provoking and enticing exploration. Thanks!!!!
Beautiful video and very intriguing proposition. I’ll explore it further. Thank you
This is an excellent video making a complex situation understandable. Thank you for sharing this information. It was very informative.
dude this is mindblowing content.. please keep going and never stop. it is always the art of technology. the art.
Love this video, thank you for sharing your perspectives Emergent Garden! I have been exploring AI art lately and it's been a fascinating experience.
This is a fantastic jumping off point to so many interesting ideas and insights.
Oh man, this was so great. Reminds me of being a kid and watching all those epic space documentaries.
sick, your vids are always welcome in my notifications
great explination of latent space and image space amazing stuff
you nailed it at "unlike any artwork I have previously seen" - this sums up MJ for me - constant pixel variation, the trip machine
Amazing! thanks for the upload. 2023 will be an amazing year
THIS is exactly the kind of explanation I was looking for, and trust me I've seen some. Great job, thank you!
You could automate these evolutionary trees by evaluating the outputs on a caption or a classifier network, making *it* select the next branch(es)
Don’t text-to-image models already do that?
@@Infiny92 yes, this is how they are trained in the first place. They start off in the hypothetical library of babble, generating random noise. But are then trained against the subset of human imagery. So in essence, with respect to language used in this video, the space AI models are in is the subset of babble that which pertains to human experience. And when you are following these “evolutionary trees” you are exploring the already curated space.
Now what would be really interesting is if instead of prompts and variations, you could draw or paint right along with it as it evolves.
You should be narrating documentaries. You have a profound and nice voice! I found your video's part (7:49) especially interesting and informative.
The library of babel is one of my favorite places, and until now I didn't think of one for video. But you can have one for any form of art.
Great video. One sidenote though as it may not be clear. Generative models don't have the capacity to produce every possible image in the image space. They generate cohesive images rather than noise because their output space is heavily constrained when compared to the image space you mentioned earlier. Their output space is usually still enormously big but there are far fewer images in it than points in the image space.
well yeah id say thats because almost all of image space is just white noise. but theres defos a bigger amount of images that would mean something to us that these ais can generate. which is a god thing i think, dont want to build rome too fast.
Or you take a formate and play through all possible pixel constellations, there may be a lot of noise but also everything imaginable picture like Mona Lisa and Einstein riding on the Sphinx in the morninglight of Quebec with both having pink bananas in their, em, seven front limbs so far ...
with the 3d generators and video generators that now exist, I feel like we aren't too far away from a day when something like no mans sky could literally have infinately generated alien lifeforms that don't have to be human created ahead of time.
That's already possible. The bottleneck is gameplay. You can generate an infinite variety of visuals, but they don't DO anything. Compositing and mixing behaviors is yet unsolved, as our understanding of the consciousness that drives behavior is bound to physical bodies, which computer-simulated agents aren't.
Even if we could generate behaviors, there's no guarantee they make any sense, or be entertaining to interact with. Having said that, it's only a matter of time until we can generate entire video games by prompt. How long, is uncertain.
We are starting to see this in the blender community. I believe there are already prototype AI for procedural textures, UV textures, and I believe bump maps. There might be one for mesh, but while it's easy to make an AI create models, it's incredibly difficult to get it to do so properly, so you get a lot of weird mesh that doesn't make sense or artifacts from the noise. I think we'd find more potential in using AI as a controller rather than as an artist. For example, you (a human) can make a bunch of body parts with specific parameters for size, texturing, etc (imagine like Spore but more advanced), so you wouldn't have something with no mouth and radio static as skin. Have those models as a template, and then use something like geometry nodes to morph between them and arrange them in various orientations. Then have the AI on the game side control those nodes, so it chooses which of the already pre-modeled body parts to put, where, and with what texture. This would be more on the world generation side, turning the game into the dwarf fortress loading screen for a while but then you get a procedural experience when you load that save.
As an artist, this tech is very interesting and I look forward to what humans can do with it. But as I say to people who are concerned about AI, the human element is always going to have to be there because AI is too capable - it can create an infinite number of outputs, but like with the library of babel, without a human it's wasting resources on something totally useless.
@@benji45645You're right. It needs human interaction and guidance. The fact that we have the tools now to make it happen is really exciting.
Btw, If you know of a model that generates bump maps like you mentioned from photos i'd love to see it. The best program I've found is Materialize, and that's not AI. I feel like AI could do much better.
@@hyperteleXii don't see why you wouldn't make a database with behavior information on humans, animals and evolution. Use it to train AI and simulate-test it to produce viable behavior, like an evolution simulation. You could also do this for all the aliens on the planet, so they would have complicated relations, things like eusocialism, parenting, symbiosis.. etc and things that don't happen on earth but might be succesfull in the simulation test. But that's probably not gonna be practical to do for a game with infinitely generated map.
@@ustanik9921 It's difficult to encode behavior as information. Results can be measured, but *why* did the animal do that? AI generated behavior will thusly *resemble* animal behavior, but have no logic to it = Falls apart upon inspection.
This was the best video I've watched on this subject
Amazing video, this is one of the best videos I have seen talking about AI image generators and such, keep it up, really looking forward to the next video haha
Thank you for this beautiful video.
Glad to know I'm not the onlyone who thinks of V-Rolling as evolving an image. I've done some crazy things exploring different branches.
dude this intro is really great for this topic im pretty sure its from the first alien movie where he turns on the computer to talk with mother
you scared me with this tecnology. Avator movie must hire you for their new characters.
Love your work.
Love from Afghanistan.
Thanks for the reminder to take pauses to be thankful for these new islands of creativity, and the maturing tools to seek them and foster them. Maybe I'll try to play with some good orangutans on the way.
This is such a cool experiment! You should do this again. Or more.
That explanation of image space at 9:00 really drive the point home. Well done.
I've loved your video, and I agree this idea of exploring evolving generations in the latent space, breeding, is not new but it's should be explored more.
In particular automatic breeding/evolution based on "fitness" by some aesthetics defined in adversarial network's like model would be cool to imagine!
since stable diffusion is open source, something A little like this has been made.
They call it textural inversion and it lets you take in A bunch of images that share something in common and assign A made up word/tag to describe that thing. Then you can include this new subject into any other prompt by adding this new tag to the prompt.
Most demos I see of it are people using pictures of themselves to allow adding them by name to any prompt.
I also saw one trained on pokemon that could make fictional pokemon with the features of various US presidents in the pokemon art style.
That was pretty cool.
Thanks OP that was a great film!
On Stable Diffusion, what should I do with this: "RuntimeError: CUDA error: out of memory
CUDA kernel errors might be asynchronously reported at some other API call,so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1"
It means you're out of gpu memory, which depends on your hardware. I'd just go to playgroundai.com to use SD for free.
Can you post a link to your discord channel to participate in the evolution thing or to download some of the breaded creatures?
Probably one of the most fascinating and intelligent videos I have seen for a long time!
really cool video ! (oragutan good btw)
im so addicted to MidJourney lol! V4 is amazing
Fantastic videos as usual my friend!
I tried this. Amaaaaaaaazing!
Nice images for new nightmares.
Great video as always!
Super interesting, but can we use for some generating text, or some story ?
There's already AI out there used to help writing scientific papers. So yeah!
Thats incredibilly interesting
4:21 realistic bug cat
When I was in primary school, I wrote a program in qbasic to try and make every possible icon for windows (32x32). I can't remember if I used 256 colour VGA (screen 13) or 16 colors.
Needless to say, after leaving the 386 running all day, seeing that only maybe four of the pixels had changed was what made me realise this was impossible.
All of those are good points.
This technology is getting good fast
beautiful video
I watched the whole thing and thought it had like 500k views but it only has 6.7k UA-cam needs to do something about this
Excellent video, I like Stable Diffusion because you get results right away with the right promt, then you can generate endless variations of any of the first set of images you created.
@EmergentGarden
Have you read "Darwin's Dangerous Idea" by Daniel Dennett. Some of the phrasing for ideas that you explore are very similar to how he put it in the book.
amazing video though, I'm going to play with stable diffusion right now
It's truly a strange time to exist. I only hope artists do not become obsolete but perhaps masterpieces can be created by meshing AI generated art with human instincts.
no artist can encapsulate so much knowledge and creativity
Perhaps an artist is really less creating something new and more exploring image space with the strokes they make on the canvas. the strokes they chose to make come from their gained understanding of the objects their brain was trained to understand, and they are combining those mental understandings of form and shape to create newly discovered images on the canvas. The images already exist in babble, they just haven't been discovered. Now the artist can use tools of language to pull out images from the image space. Perhaps less precise than the brush stroke, but faster and more of an iterative process. in this way the artist is not replaced (at least not directly though the barrier for entry may be lowered and the old ways of artistic profit threatened) but granted new tools in the exploratory process.
Something important to note with models like midjourney is that they _have_ been trained on the creativity of others, and while you might not ever have seen some of the art that has guided the model, everything it generates is entirely derivative. These models have none of the attributes that could be classed as creative, even with a human ultimately guiding the process through prompts. It is more akin to taking existing paintings and smooshing them around, than anything resembling original work. Their styles, all of the artistry you might find while exploring their model, it is all explicitly plagiarized from existing work. Whether you care about the ethics or not, it is hardly novel. Picbreeder is the most novel thing in this video, because it is a model trained and guided entirely by collaborative creation, all that it generates is original and purely evolved within the collective mind. (I had not heard of Picbreeder before seeing this video, but it's certainly the most interesting to me)
Amazing
Can you detail your settings in stable diffusion for this kind variations/evolutions ?
Wow...this is a....nightmarish monster breeder...cool
this is the most insane shit ive ever seen.... what an absolutely astounding explanation of image space and the possibilities we have...
Looks like we're one step closer to creating a Skynet situation. Better start stockpiling those Terminators now!😂
Какой шикарный способ создавать концепты новых существ.. причём сразу с эволюцией от эмбриона
Wow, just blown away!
being a graphic designer this is something so amazing and great! Creating abstract image in matter of mins which didn't ever exist in reality is just mind blowing!
Literally!
I was so eager to explore that I used up my trial verson before I could realise it. Thank you so much for this.
And, I can't afford pay so much for a subscription. I wish there will be more such AIs in future which will make a transformation in our perception
I believe that software engineering is actually collective software breeding with trying to apply best practices to solve common problems in the most optimal way based on the context.
And one of the my "field of interest" is evolving of software systems themselves. Where can I speak with someone about this? :D
By the way, your video is very exciting, thank you!
UPD: also, it is important to mention, that neural network implicitly adapts to suit your visual tastes, if you have control in breeding process.
Thank you for eloquently putting into words the comparison with an gallery of babel that I felt was a good way to think about AI art, but didn't have the words to explain.
What a great video! WOW! After painting for 40 years with a brush, I am now diving in that Image Space.... and it is a mindblowing trip
Brilliant!
I have been playing with stable diffusion for the past couple months. And it keeps on amazing me. Still feel like i need to learn what prompts i should give and what pictures to select for seeding the next ones. It's a whole new world.
Beautiful :)
Great video!
I was going to make some congratulatory comments about how well you have articulated some of the key intrinsic values of AI driven generative image creation but then at the end of the video I was left feeling inarticulate for a few minutes (obviously I got over that). I could feel my brain "chewing" on perception in a "new" way. It made me think that perhaps either passively experiencing image evolution (like in this video) or doing it actively yourself (like I did for a while on your discord) results in an altered mental state sort of like a meditation or "waking" lucid dreaming state. In addition to being a scientist myself (currently studying to be a data scientist and have experience coding neural network models) I am also an artist.. I find that your video has resulted in a "flow state" that I often feel when inspired during art-making. This is a good thing.
orangutan good! I just found your page. I am in this all the way friend.
thanks alot
Gracias!!
I was thinking of creating a AI generated art magazine.
Fascinating video and so very well made! Thought-provoking, eerie, beautifully scripted and a joy to look at ❤
underated vid!
mesmerizing
incredible
Any idea who owns this computer what makes these pictures?
Can someone explain how to use discord? I really dont get it. I was in a group for a game I play and dont even know how I got there lol.
if a ai can take a prompt and turn into a image, could they take one of these library of babel images and give it a description
could they organize the mess of library of babel
could they?
4:57 I nearly spit my coffee when the picture of your gpu popped up. RIP little gpu. LOL That leaf bug is amazing! If I knew more about how to use discord I'd love to come play (and if midjourney was free lol). I've been playing quite a bit with Dream by WOMBO. Glad i found your channel.
This is the sales pitch of the next generation, demo of 13 mins
As an artist, this is horrifying.
The source images, or "training sets", that are vital to this generation of AI art, did not just appear. They are not the product of some "image space" that was always there, that we are now simply discovering- they were made, by people. They had meaning, to people. It's obvious that Midjourney is sourcing from a lot of contemporary, digital illustrators and from what I've seen, they do nothing to credit or compensate those artists, while the company is making a profit.
AI art would be cool if it was JUST done with works in the public domain, or with the consent of artists, even if that might nerf these generators compared to what they are now. If people had that sort of protection, I'd be thinking "Wow, this could be great for people who make creepy surreal art". As it is, I'm like "Wow, if I made creepy surreal art, this would make me think twice about posting it".
Again, I don't hate the concept, just the way it's often done. The entire experiment of evolution through AI would be infinitely more interesting if the DNA wasn't a dirty secret.
It's also worth noting, the training sets would normally be illegal for these companies to use. To quote OpenAI, "We want to increase our ability to raise capital while still serving our mission, and no pre-existing legal structure we know of strikes the right balance." So, they found a loophole in the laws.
OpenAI is actually two companies. OpenAI LP is a for-profit company. That company is a child of OpenAI Inc, a non-profit research organization. As a non-profit research organization, OpenAI Inc gets special exemptions regarding things like copyright law for their training data, because it's considered research. They also don't have to pay taxes. OpenAI LP inherits a lot of the same exemptions and benefits, but they also get to do everything a normal for-profit company can do. Technically, OpenAI LP is "capped-profit", so after they make a certain amount they have to start giving it to OpenAI Inc. It also means investors can only make so much from their investments. So what's the cap? If you were to invest 10 thousand, you can get a maximum of 1 million back. That's fucking absurd.
What really scares me is that a lot of artists now want to increase the strictness of copyright, and it WILL only hurt them and empower corporations like Disney. Because companies like OpenAI /don't need to follow those laws./
The ML trains, like an artist do, the look to other people work and "conceptualize it", the only difference is that it does so in a more vast and efficient way.
A real artist is the concept (the idea or the prompt/evolution in these models), or the mere execution?
What if I compose it with things I drew with another software or by hand on paper? Would it be more valuable to you? Would you feel different about the art piece?
It's unethical for an artist (or an art director) to employ other artist to produce his work? This happens everyday between humans and nobody worries about it, because we're used to the attribution mechanism. Would you be cool if I grant credit to a software? Or if the software grants credits to all the possible artists he got trained upon? Would anyone care or made any difference on the way they view the subject of AI-Art?
Hits so hard
im blown away.
I also had exactly those thoughts, thanks for put it in a video
Interesting, but I like what my mind can create any day! As I’ve seen, and will see things that no one has ever seen and will ever see or know of!
This was amazing.
pretty good
What if you use the images from the image library of babel to be used in the stabled iffusion or midjourney?
what is the pupose of artists now?
Wait, why does moving around the latent space produce mutations that are interesting to us instead of just random noise appearing all around? The movement is not random then, is it? It knows where to "randomly" move, right?
In the latent space things that look alike are near each other, like a gradient
@@JorgetePanete but the same image with random pixels smeared all over also looks alike the original image, doesn't it?
When he was illustrating moving around latent space we saw the hands move around a body. Why does that happen?
Isn't this space just a multidimensional array representing every possible combination of pixels of a particular resolution?
How could then the arms moving around be in proximity of the original image? This set would have to have the concept of hands being positioned in different parts on the body imbedded in it such that such images are in closer proximity than the same image with some random pixels all around.
Random pixels appearing where his hand was is also "slightly different", but almost the same" isn't it?
Seems like he skipped over something important here, because its not adding up for me.
@@asyncasync It isn't all the posibilities, it's what the network can offer and depends on its training and structure
Fabulous