Yeah I can’t really describe it but this guy lives in his own little world. He’s a friendly bloke but just does not give the slightest shit about anything that doesn’t affect him personally.
Then there’s my fucking cunt of a housemate who lkves on the floor above me that’s the loudest mother fucker on the face of the earth and fucking stomps around until 1am every night and opens and slams close every fucking draw and door in his room, and re arrnafes his furniture every night, then gets up at 6.30 in the morning so he can stomp around some more, hack up his lungs as loudly as humanly possible and slam every door in the fucking house before leaving for work exactly when my alarm goes off, specifically so I can’t have a single good night sleep ever.
You’re talking about genetic algorithms which is not the method of training used by AI in stable diffusion.
I understand that the topic of AI provokes a big emotional reaction in you, but I suggest you don’t actively participate in tbe discussion until you get a better grasp of the subject at hand, else you just inadvertently end up spreading misinformation.
Nah I agree with him. I missed it a but at first because I was only paying half attention but the whole “did you just assume the meeps pronouns” bit and the ending non binary stuff was very forced and clumsy and just distracting.
And I say that as the the stereotypical pink haired non binary gay feminist. Inclusion should be normalising making it a big deal and metaphorically winking down the camera does as much harm as good.
Yeah no shit if you train your program wrong you’re going to get bad results. But that’s all overfitting is, a training failure. So instead of your bot being able to evaluate your image based on how it fits the tags, it evaluates it based on how it fits the overffitted data. So while it still doesn’t actually remember the overfilled data you still end up converging on it.
Though I’ll give you that you can make the argument that if you train your program on a single image, then it is just remembering that one imagine and essentially copying it. But that’s literally no different to hardcoding a program that produces that one image.
Yes, in that example bot 2 would have been trained on the mona Lisa, but bot 1 still has no idea what the mkna Lisa looks
Picture it like this. You are an artist that has never seen the Mona Lisa, and I tell you to draw it. Because you don’t know what it is you give me something completely random. Then I give you a score of 2/100 and ask you to draw it again. You give me something slightly different so I score it 4/100 so now you have a better idea of what to do so the next drawing gets 10/100, then 20/100 and we keep going like this until you get to 90/100 and you’ve pretty much drawn the mona lisa. You’ve still never seen the actual mona lisa, and I havnt told you what it looks like either.
why aren’t we automating all the hard labor and leaving humans to do the art?
Now this argument is one that tally boils my piss.
We fucking are??? Have you ever been to modern factory in your life? The majority of production is automated, farming is highly automated, customer service jobs are being automated, delivery jobs. So much time effort and money is going into automating every job. But its just easy to get a computer to do something entirely digital than it is to get it you build you a house. Robotics is 1000× more complicated than AI.
disgusted by the potential if AI is removing what fulfillment we are able to have as people.
Human made art is still legal. Its not going anywhere and if you’re good you will still be able to make money off it. So nothing is stopping people from making art still. It’s like knitting after the invention of the automatic loom. You can still hand knit things, and have fun doing it and maybe even make money from selling then, but it just means we don’t kneed thousands of people working in textile Mills.
I mean if I don’t know how AI works then I should probably hand back my masters degree.
It’s not a semantic argument, it’s a fact.
The simple terms explanation of how stable diffusion works is you have 2 bots. The first bot, the one that actually “genrates” the image, the only thing it knows how to do is to take a blurry image and make it less blurry.
The second bot is the one that’s trained on the data and all it can do is tell you how much a given image matches a key word.
So you generate a completely random grid of pixels. Tell bot 1 to unblur it, then bot 2 evaluates how much it resembles the prompts, then bot 1 takes that evaluation and uses it to unblur the image a bit better, bot 2 evaluates again, and so on and so on until you have an image that passes evaluation. So bot 1 has never even been trained on other peopes data.
So with that said, saying something like.
attempting to reproduce the signatures of artists they trained on
Is just nonsense it’s not attempting to reproduce anything like that. At best you can say it’s recognising the pattern of “lots of art has signatures in the bottom right corner” and so generations that put a little squiggle in tbe bottom right corner match the evaluators patterns better.
And unless you specifically set up a system to produce slightly modified versions of famous art pieces and then actively try to get slight modify versions of famous art peice then you aren’t going to get any. It’s like saying all digital art is derivative because you literally have a copy paste function you can use, while ignoring the far more common use of it to create new art pieces.