|
https://twitter.com/mudevol/status/1576360644822958080
|
#
?
Oct 2, 2022 22:35
|
|
|
|
| # ? Jun 9, 2024 15:28 |
|
|
And they say AI can't make art.
|
|
#
?
Oct 2, 2022 22:52
|
|
|
it's a fun toy. cool tech. lots of valid concerns. cat out of bag etc. etc. but "these pictures suck! they suck!!!!" is a weak leg to stand on like look at where this tech has gone since Deep Dream in 2015, and look at the progress just in the last six months, and look at imagen and parti, and think about where that's going to be in 1 year, 5 years, 15 years. there are plenty of good reasons to poo poo on AI art, and reasons to like it, but "heh, look at those hands and eyes" is as dumb a take as the handful of people going "it's objectively very good! ready for primetime, right now!" you will be chasing the suck of the gaps forever.
|
|
#
?
Oct 3, 2022 02:10
|
|
|
Mola Yam posted:there are plenty of good reasons to poo poo on AI art, and reasons to like it, but "heh, look at those hands and eyes" is as dumb a take as the handful of people going "it's objectively very good! ready for primetime, right now!" This is also not a great take because a lot of this tech doesn't exist on a straight line from "bad" to "good." It's incorrect to assume that anything like this will just keep steadily marching toward some line that's finally acceptable instead of dead-ending with goopy eyes and wonky hands. People can only evaluate what's in front of them right now and it's not any more correct to just assume that it will keep improving until it beats out human artists.
|
|
#
?
Oct 3, 2022 06:13
|
|
|
sure but just in the last couple of months we've seen three? text-to-video models (all of which sort of suck rear end, but so did deep dream), as well as at least one text-to-3d mesh model and the (heavily controlled/curated) images coming out of parti/imagen, with 10x more parameters, look really fucken good compared to the public models   like even the 20B versions are kinda hosed up if you look closely at them, and i agree that this will likely all plateau at some point and we'll have an AI art winter, but it seems silly to say that dall-e, midjourney and stablediffusion are as good as it gets this generation
|
|
#
?
Oct 3, 2022 06:32
|
|
|
BUT i agree that anyone calling themselves an "artist" because they use one of these AI tools is kind of huffing their own farts although i think there are far fewer of those people than the artist community imagines - most people are just dicking around with tech toys and going "huh cool" or "lol that's hosed up"
|
|
#
?
Oct 3, 2022 06:35
|
|
|
Paradoxish posted:It's incorrect to assume that anything like this will just keep steadily marching toward some line that's finally acceptable instead of dead-ending with goopy eyes and wonky hands. I mean, the lovely free ones like Craiyon and Stable Diffusion are pretty bad examples of what AI can do. When you look at DALL E 2 and Midjourney they definitely make stuff sometimes that could be passed off as made by humans no problem. Midjourney also has features to enhance/spend more time an existing render to make it even more believable Skyl3lazer posted:Personally the main thing I use them for is making art of D&D/PF2E characters Like that bottom left one looks pretty legit, and it's probably just a one pass render. I could definitely see that as a character portrait in a game and would just assume they had a digital artist make it.
|
|
#
?
Oct 3, 2022 06:39
|
|
|
Paradoxish posted:This is also not a great take because a lot of this tech doesn't exist on a straight line from "bad" to "good." It's incorrect to assume that anything like this will just keep steadily marching toward some line that's finally acceptable instead of dead-ending with goopy eyes and wonky hands. it'll probably improve since it has attracted the attention of the nerd death laser but it is very funny to see artist freakout over it, because it'll definitely wipe out the market for mediocrity.
|
|
#
?
Oct 3, 2022 06:41
|
|
|
ai art got that the problem where it looks like this guy
|
|
#
?
Oct 3, 2022 06:43
|
|
|
suffice to say theres going to be a huge boom in comic books/manga/etc because AI is just tailor made for passable enough to tell a story with that any yahoo can use.
|
|
#
?
Oct 3, 2022 06:44
|
|
|
Mola Yam posted:it's a fun toy. cool tech. lots of valid concerns. cat out of bag etc. etc. They do suck, they suck poo poo. And they're based on having them look at a zillion pieces of existing art. until it steals without stealing. without the seed of prior art it's nothing. and to as a rebuttal say that about artists, human artists, who need only the seed of their actual sense data, the first time every time, hunter gatherers straight to masterpieces no waiting, makes one a polo shirt. the shirt's fused to your body. the e-waifu will never make art.
|
|
#
?
Oct 3, 2022 06:49
|
|
|
Yeah that's also my argument about synthetic fertilizers. On real human poo poo for this grower!
|
|
#
?
Oct 3, 2022 06:57
|
|
|
the white hand posted:They do suck, they suck poo poo. And they're based on having them look at a zillion pieces of existing art. until it steals without stealing. without the seed of prior art it's nothing. basically it'll crush hitler. hitler would have been told "i can do this with computer"
|
|
#
?
Oct 3, 2022 06:57
|
|
|
WrasslorMonkey posted:Yeah that's also my argument about synthetic fertilizers. On real human poo poo for this grower! It works great on your posts.
|
|
#
?
Oct 3, 2022 07:01
|
|
|
the white hand posted:It works great on your posts. I guess I'll start pouring your posts on mine then!
|
|
#
?
Oct 3, 2022 07:02
|
|
|
when will spotify get AI generated music where you throw your fav poo poo into a mixer and get 5 outputs. that poo poo would own.
|
|
#
?
Oct 3, 2022 07:05
|
|
|
Concept artists get the same advice the old weavers got when the power loom became a thing "flee with all alacrity from the trade"
|
|
#
?
Oct 3, 2022 09:55
|
|
|
Edit before posting: If you don't wanna hear any of that do you think they can use this to make drawing tools for artists? I could see it being able to help with speeding up drawing individual elements a lot easier than building a work from scratch. Like this would probably sooner let a good artist draw faster than it would let a bad artist become good. They'd still need to put a lot of money up front to develop each tool indiviy, and they'd have to redo it for every change to an element and for every style. I somehow doubt the ability to recreate the human artistic spirit in a machine/program when they can barely stand to nurture it in actual humans. I could imagine a world in which nuclear fusion power was developed in the 70s and where we worked to build everything efficiently and with an eye for long term research and development where we use programs similar to this but designed to pluck out interesting data from the algorithms to try to see what can be understood about various artistic concepts, to give actual artists interesting ideas to focus on. Maybe it could be used with the help of artists to make real art that a human might almost never make themselves, but could be appreciated by a much wider audience. Given my understanding of how complicated systems work I don't think it's gonna outright replace artists in a world that can be fooled by Bitcoin in any significant portion (not that I'm saying anyone here is, this is a much more interesting idea than that was on its face, and has some merits I think).
|
|
#
?
Oct 3, 2022 11:24
|
|
|
the white hand posted:They do suck, they suck poo poo. And they're based on having them look at a zillion pieces of existing art. until it steals without stealing. without the seed of prior art it's nothing. and to as a rebuttal say that about artists, human artists, who need only the seed of their actual sense data, This is still a dumb and untrue point. Learning models are the sensory inputs for these programs, you could easily imagine it being hooked up to a camera taking a shitload of pictures of whatever subjects
|
|
#
?
Oct 3, 2022 15:07
|
|
|
is art, art? what doth life?
|
|
#
?
Oct 3, 2022 15:43
|
|
|
Tungsten posted:is art, art? this mf spittin
|
|
#
?
Oct 3, 2022 15:44
|
|
|
Skyl3lazer posted:This is still a dumb and untrue point. Learning models are the sensory inputs for these programs, you could easily imagine it being hooked up to a camera taking a shitload of pictures of whatever subjects Which would then allow it to regurgitate images that resemble the ones captured by it's camera. If you have humans add labeling to what it captures you could use that for weighted prompts. Again, it's really overselling what these do or underselling what the human brain does with the information you're bringing in with you're eyes.
|
|
#
?
Oct 3, 2022 16:22
|
|
|
Skyl3lazer posted:This is still a dumb and untrue point. Learning models are the sensory inputs for these programs, you could easily imagine it being hooked up to a camera taking a shitload of pictures of whatever subjects wow, it's leonardo now 
|
|
#
?
Oct 3, 2022 16:40
|
|
|
Gumball Gumption posted:Which would then allow it to regurgitate images that resemble the ones captured by it's camera. If you have humans add labeling to what it captures you could use that for weighted prompts. Again, it's really overselling what these do or underselling what the human brain does with the information you're bringing in with you're eyes. Just because your brain does a thing (see thing + assign definition to) automatically, and a computer is given that data instead, does not make what you are doing unique. I'm not pretending these programs have a soul or some poo poo, just that it's really dumb and unhelpful to try to claim they're incapable of making pretty pictures just because people do it too.
|
|
#
?
Oct 3, 2022 17:23
|
|
|
it's completely unique, it's what assigns these products meaning
|
|
#
?
Oct 3, 2022 17:25
|
|
|
Skyl3lazer posted:Just because your brain does a thing (see thing + assign definition to) automatically, and a computer is given that data instead, does not make what you are doing unique. I'm not pretending these programs have a soul or some poo poo, just that it's really dumb and unhelpful to try to claim they're incapable of making pretty pictures just because people do it too. We're nowhere close to developing AI that functions similar to the human brain, especially considering we don't fully understand how it functions. Much like the art that's produced the machines function in a way that produces outcomes that are loose copies of what we can produce but the process is not the same and that difference in the process is crucial.
|
|
#
?
Oct 3, 2022 17:32
|
|
|
Skyl3lazer posted:Learning models are the sensory inputs for these programs, you could easily imagine it being hooked up to a camera taking a shitload of pictures of whatever subjects no you couldn't.
|
|
#
?
Oct 3, 2022 17:34
|
|
|
Gumball Gumption posted:We're nowhere close to developing AI that functions similar to the human brain, especially considering we don't fully understand how it functions. Much like the art that's produced the machines function in a way that produces outcomes that are loose copies of what we can produce but the process is not the same and that difference in the process is crucial. we have a very good understanding of how the brain functions, we just can't square what we understand with our assumptions that brains are the cause of behavior (because they're not, they're overgrown behavioral autopilots.)
|
|
#
?
Oct 3, 2022 17:36
|
|
|
Zodium posted:no you couldn't. well I guess I can and one shouldn't depend on the brain function of other posters
|
|
#
?
Oct 3, 2022 18:00
|
|
|
Skyl3lazer posted:well I guess I can and one shouldn't depend on the brain function of other posters that's not how it works though
|
|
#
?
Oct 3, 2022 18:40
|
|
|
to train a neural network that knows what things like cows are then you need to have a selection of pictures of cows that are all annotated to let the AI know that it is looking at cows because it's a dumb computer program that only understands pixels. then you need to have another set of pictures of cows, separate from the pictures of cows it's learning about, so that it can be graded on how well it's doing at making cows in comparison to cows that it has never seen before. then you gotta do this for everything
|
|
#
?
Oct 3, 2022 18:52
|
|
|
turn off the TV posted:to train a neural network that knows what things like cows are then you need to have a selection of pictures of cows that are all annotated to let the AI know that it is looking at cows because it's a dumb computer program that only understands pixels. then you need to have another set of pictures of cows, separate from the pictures of cows it's learning about, so that it can be graded on how well it's doing at making cows in comparison to cows that it has never seen before. yep, I am in the process of reading up on the myriad ways to achieve this, because we have some thousands of pictures one of our kids drew 2020-present that my we have been bulk-scanning into image files, which more or less break down into a handful of motifs that they've gotten heavily into for a while. ("Monsters", "families", etc). I have professional experience with very different kinds of ML apps but which are still baked around the general principle of training a model based on an input set and metadata. For sure the shittiest part of this project is going to be manually adding metadata to several thousand images, and so I am taking some time to understand what I'm doing before heading down a bad path. The end results may be entirely unremarkable but overall this idea is a lot more interesting to me than endless midjourney prompts, which were fun for a while, and then so was stable diffusion for a while, and then it's just more of the same. Running my own models that have only been trained on the specific kinds of weird poo poo our kid drew would be highly amusing to us even if no one else thought so, just like most of the poo poo I do for fun. barge posted:it's fun to see a drawing and know who made it just by their style. I htink people excited about this stuff just don't realize that, people might actually enjoying seeing the wonky eyed wizard you drew on printer paper for your DnD campaign while not one single person is ever going to enjoy seeing the digital paint wizard you had dall-e print out. I haven't played a paper RPG in years so I don't know if I'd give a poo poo about this kind of content, I mostly played RPGs as an excuse to get lightly intoxicated in a group setting and pay mild attention to the game. Some people really seem to like it, though, so I am sure Hasbro has their eye of sauron on that and is trying to think of how they can monetize that concept  The best Midjourney/StableDiffusion stuff is on par with some of the more middling Magic the Gathering card art. I can't believe Hasbro hasn't noticed that, too, buyt they already get away with underpaying and treating artists so badly it may not be worth their while, also I think the MtG playerbase might flip out in a way D&D people aren't because of the differences in how the poo poo is monitized. Won't be at all surprised to see "generative fantasy art" make its commercial debut in a Hasbro/WoTC thing, though.
|
|
#
?
Oct 3, 2022 20:19
|
|
|
When Sol LeWitt created his series of wall drawings, he would do so by creating an instruction set and having others follow those instructions:Wall Drawing #1180 posted:Wall Drawing #382 posted:On a black wall, a white circle within which are white vertical parallel lines, and a white parallelogram within which are white horizontal parallel lines. The vertical lines within the circle do not enter the parallelogram, and the horizontal lines within the parallelogram do not enter the circle. Wall Drawing 869C posted:From the top of a 48-inch (122 cm) square, draw a not straight horizontal line. The line is black. The second line is drawn beneath the first line, as close as possible, imitating the first line. The next line is drawn beneath the second line. Continue copying, until the bottom of the square is reached His work raised several questions that I think inform the discussion of applied machine learning in art today. Was Sol LeWitt the artist who created these works? Were the people who interpreted his instructions and physically drew the art the artists? Were they both? Are the instructions art? Or just finished product? Or both? Would a human following LeWitt's instructions create art but a fuzzy algorithm not? What if the human drawn and machine drawn versions were visually the same?
|
|
#
?
Oct 3, 2022 22:58
|
|
|
Mola Yam posted:it's a fun toy. cool tech. lots of valid concerns. cat out of bag etc. etc. The hands and eyes are special cases because our brains are very good at parsing them. If anything, those will be the last subjects to look passable. It’s not impossible for AI art to get there, it would just take longer. It might not be a coincidence that both those examples you posted are wearing glasses/visors. I would compare it to self driving car tech. Now your car has a light that tells you when it’s safe to switch lanes, but that doesn’t mean it can do all the object detection required to drive itself. It isn’t impossible for AI to improve enough to drive the car, but it’s been 5 years away for a while now because you can get pretty close without the cigar.
|
|
#
?
Oct 3, 2022 23:40
|
|
|
   
|
|
#
?
Oct 4, 2022 00:19
|
|
|
Skyl3lazer posted:Personally the main thing I use them for is making art of D&D/PF2E characters Hmmm. Time to reinstall bg2.
|
|
#
?
Oct 4, 2022 00:38
|
|
|
Trabisnikof posted:When Sol LeWitt created his series of wall drawings, he would do so by creating an instruction set and having others follow those instructions: These are interesting thanks
|
|
#
?
Oct 4, 2022 00:57
|
|
|
Trabisnikof posted:When Sol LeWitt created his series of wall drawings, he would do so by creating an instruction set and having others follow those instructions: sol lewitt is one of my favorites. I think he said something like “the idea becomes a machine that makes the art” and he was quite happy leaving his instructions unexecuted if he couldn’t wrangle a team of graduate students to help him out. so there ya go, the end result is the art but the ideas are your own (or lewitts in this example)
|
|
#
?
Oct 4, 2022 00:59
|
|
|
Doesn't that leave us at the prompt is the art, the bot is the tool, the sweaty goon at the keyboard is the artiste?
|
|
#
?
Oct 4, 2022 01:08
|
|
|
|
| # ? Jun 9, 2024 15:28 |
|
|
Slavvy posted:Doesn't that leave us at the prompt is the art, the bot is the tool, the sweaty goon at the keyboard is the artiste?  My Art: 
|
|
#
?
Oct 4, 2022 01:13
|
|