|
There's lot of AI art tools floating around these days. There's also a lot of very angry people who think there's no AI art tools, just AI Art Theft Devices. I think they're kind of neat, but also an interesting ethical sort of question, so I made a thread Tools like Dall*E, Midjourney (used for the examples in this post), and Stable Diffusion use various flavors of machine learning algorithms to try to generate "art." Generally, you give them a prompt of some sort, and they'll go and produce a set of images in response. The "learning" that the machine does is based on a "training set" of images, generally numbering in the tens of thousands if not millions, sort of like googling a given term repeatedly and then trying draw something based on 50 pages of results. From the learning set, the algorithm will try to determine what constitutes a given term - what it means to have a "tree" in your drawing, or what an "oil paining" tends to look like.  Prompt: a photograph of an apple Those images can be refined in various ways, and the end result is a (hypothetically) novel piece of artwork, realistic or not, created by the AI.  Prompt: CSPAM One problem, as many see it, is that this process gives no credit towards the people who created the images in the original learning set. Even if your IP-protected picture was one in a million images used in training, does that mean the bot's owners should pay you royalties? Even the bot's creators would have no way to tell exactly which images were the "inspiration" to a given generation. Is it plagarism? Am I ethically stealing your work, or is my tool inspired by your work? You wouldn't say one artist plagiarized another because they were influenced by the latter's art. There's also, however, clearly a point where you've intentionally made the AI replicate someone else's art to the point of plagiarism.  Prompt: CSPAM, drawn by Dr. Seuss The prompts themselves play a primary role in the end product here. Some developers disallow some terms, usually involving nudity or prominent public figures. So far at least, none try to block things like artist names or style specifications. If you get specific enough in a prompt, not only can you create very interesting images, you can try to force the AI to limit the "influence" images to just those drawn by even a single person.  Prompt: Black Lotus, Lotus, Christopher Rush, Painted, Rush, Magic, MTG, Magic the Gathering, Drawing, Original, Alpha The Econ thread had a ball talking about this stuff so maybe people outside of there will also. What do you think? Can AIs make art? Is there any theft happening here? If so, at what point? Are these AIs breaking any current law, or should any new law exist (assume that laws matter for this one)? Is this the downfall of the Arteest? Is it a flash in the pan? Is it just a new tool for artists to use? or maybe just post some cool AI art to use for avatars or whatever
|
#
?
Sep 19, 2022 14:33
|
|
|
|
| # ? May 2, 2024 17:16 |
|
|
i generated a bunch of c-spam dall*e poo poo awhile back and i'm pretty sure i didn't post them because they only made me lol a little, so i guess i'll post them here      
|
|
#
?
Sep 19, 2022 14:54
|
|
|
I've been tempted to go into threads and do that, but none of them have been good enough: Stable Diffusion Prompt: "''Only Trust your Respirator' by Vitaly Alexius"  
|
|
#
?
Sep 19, 2022 15:08
|
|
|
If you have an NVIDIA ~2070 or better* you can run Stable Diffusion locally with a good UI and bells and whistles with https://github.com/AUTOMATIC1111/stable-diffusion-webui. It's pretty close to a one click install. If you are on Windows/Linux and want an actual one click install and are willing to give up some features https://github.com/cmdr2/stable-diffusion-ui is another good option. *Possible to run on AMD, CPU, or older cards but requires know-how
|
|
#
?
Sep 19, 2022 15:32
|
|
|
Buffer posted:I've been tempted to go into threads and do that, but none of them have been good enough: These are cool, and show what I mean in the op about using an artist's style Cold on a Cob posted:i generated a bunch of c-spam dall*e poo poo awhile back and i'm pretty sure i didn't post them because they only made me lol a little, so i guess i'll post them here Midjourney is getting real good at these  Jerome Powell as the joker
|
|
#
?
Sep 19, 2022 16:57
|
|
|
Skyl3lazer posted:These are cool, and show what I mean in the op about using an artist's style For sure. It's not quite his style but you can see it doing his signature, and it's definitely in the ballpark. Also you can see in your MTG card example that if what you're going for is like RPG / video game concept art it does a ludicrously good job. You might need to generate 40 or so images to get a really good one but it'll get you in the ballpark - certainly close enough to then present whatever to a real artist as a description. Like if we'd had stable diffusion when the forums were doing stuff like zyborne clock there'd be a whole world built somewhere with daguerrotypes and poo poo of Johnny Five Aces. Image to image is where it gets even.. weirder. Nothing does photos well yet, fortunately, but it's just a matter of when.
|
|
#
?
Sep 19, 2022 17:04
|
|
|
stolen from pyf:
|
|
#
?
Sep 19, 2022 18:15
|
|
|
hello, old friend
|
|
#
?
Sep 19, 2022 18:20
|
|
|
RandolphCarter posted:stolen from pyf: amazing e; dont want to ruin it TOO much Skyl3lazer has issued a correction as of 18:51 on Sep 19, 2022 |
|
#
?
Sep 19, 2022 18:22
|
|
|
RandolphCarter posted:stolen from pyf: fukin lol
|
|
#
?
Sep 19, 2022 18:51
|
|
|
RandolphCarter posted:stolen from pyf: beautiful
|
|
#
?
Sep 19, 2022 19:08
|
|
|
guys you gotta stop giving it away with the thumbnail
|
|
#
?
Sep 19, 2022 19:15
|
|
|
RandolphCarter posted:stolen from pyf: Shear Modulus has issued a correction as of 20:09 on Sep 19, 2022 |
|
#
?
Sep 19, 2022 19:17
|
|
|
TenementFunster posted:guys you gotta stop giving it away with the thumbnail
|
|
#
?
Sep 19, 2022 19:43
|
|
|
RandolphCarter posted:stolen from pyf:
|
|
#
?
Sep 19, 2022 19:48
|
|
|
TenementFunster posted:guys you gotta stop giving it away with the thumbnail
|
|
#
?
Sep 19, 2022 19:48
|
|
|
don’t really see how there’s much to give away
|
|
#
?
Sep 19, 2022 20:10
|
|
|
On my phone screen its never not small enough to be clear
|
|
#
?
Sep 19, 2022 20:10
|
|
|
Wolfy posted:don’t really see how there’s much to give away the man has a lot to give
|
|
#
?
Sep 19, 2022 20:11
|
|
|
Shear Modulus posted:the man has a lot to give no you're thinking of "the giver" this is "the receiver"
|
|
#
?
Sep 19, 2022 20:23
|
|
|
A few years ago I trained a neural net to reconstruct blurry LIDAR DEMs using public domain data from the USGS, which I think definitely qualifies as a tool.   You can create a really crude outline of a DEM, blur it, and have the AI reconstruct it using real world data. I used it to make maps for a game, and the results are way better than what I managed with an erosion simulation. There are a lot of different uses for stuff like this that I don't think are really ethically questionable, like image colorization and super resolution. It only gets sketchy once you start getting into style transfers and digital art generators.
|
|
#
?
Sep 19, 2022 21:06
|
|
|
Skyl3lazer posted:Midjourney is getting real good at these same thing, but with a picture of powell as the input (prompt: "Screenprint of The Joker as federal reserve chairman. DC Comics"):  
|
|
#
?
Sep 20, 2022 00:16
|
|
|
RandolphCarter posted:stolen from pyf: Mankind's greatest work of art... produced... by a machine 
|
|
#
?
Sep 20, 2022 00:20
|
|
|
IMO if artists don't want robots stealing their copyright or whatever they should just mint their art as NFTs.Paradoxish posted:Mankind's greatest work of art... produced... by a machine It's still not as good as the original.
|
|
#
?
Sep 20, 2022 01:28
|
|
|
https://twitter.com/Todd_Spence/status/1571910565983055872?s=20&t=WgUrzdrk8qpWP1KJa-dr5g
|
|
#
?
Sep 20, 2022 01:36
|
|
|
whats not to like here     
|
|
#
?
Sep 20, 2022 01:52
|
|
|
after like three months of this poo poo I simply do not care that a computer can now make something that looks like a magic card illustration unless you inspect any detail closely.
|
|
#
?
Sep 20, 2022 02:56
|
|
|
Serious post: it definitely seems like the training sets are just raw copyright infringement. But unless the AI reproduces images with obvious water marks (which it has for stock photo services like Getty) or the developer of the AI releases the training set you would never know if they used your images or not. Also seems like the kind of thing the courts or legislature would retroactively legalize because the cat is already out of the bag and it’s basically rich people vs artists so beyond the large stock image sites maybe negotiating some kind of nominal fee I don’t see artists winning. I will admit I have derived some enjoyment seeing people twist themselves into knots trying to argue how The Fountain is art but AI art isn’t and how death of the author doesn’t apply instead of just saying they are scared of losing their jobs and being screwed over by big tech.
|
|
#
?
Sep 20, 2022 03:29
|
|
|
Skyl3lazer posted:What do you think? Can AIs make art? Is there any theft happening here? If so, at what point? Are these AIs breaking any current law, or should any new law exist (assume that laws matter for this one)? Is this the downfall of the Arteest? Is it a flash in the pan? Is it just a new tool for artists to use? my hot take is that you can wind up with results that are cheap ripoffs of other works, and just unremarkable trash, but it ultimately depends on the person generating the images to curate the results and not select that stuff. the capacity of the medium to generate unremarkable trash or ripoffs is not a big problem in my opinion. AI art isn't going away; the tools will get better and artists will get better at using it to produce cool stuff that isn't trash. and there will be much hand wringing and pearl clutching from big babies who want to define what art is or isn't, but what else is new. speng31b has issued a correction as of 20:55 on Sep 20, 2022 |
|
#
?
Sep 20, 2022 20:52
|
|
|
What I really wish I could do is take one style of output that I like and apply it to other prompts. It does a good job of mimicking existing styles to an extent, but when I'm trying to create 10 images that all depict different characters/scenes from a story, I want them all to look like they were drawn by the same hands. I can get them all to look like "digital paintings" or "35 mm macro photography" or "pulp comic art," I can even get them all to look like one specific artist's style. But I can't take a style that the AI has created as an output and say "okay, now make this new prompt look like it was drawn by the same person who did this one."
|
|
#
?
Sep 20, 2022 21:00
|
|
|
oh nice we can use this to replace thousands of artist jobs for free.
|
|
#
?
Sep 20, 2022 21:05
|
|
|
|
|
#
?
Sep 20, 2022 21:11
|
|
|
Pepe Silvia Browne posted:What I really wish I could do is take one style of output that I like and apply it to other prompts. It does a good job of mimicking existing styles to an extent, but when I'm trying to create 10 images that all depict different characters/scenes from a story, I want them all to look like they were drawn by the same hands. I can get them all to look like "digital paintings" or "35 mm macro photography" or "pulp comic art," I can even get them all to look like one specific artist's style. But I can't take a style that the AI has created as an output and say "okay, now make this new prompt look like it was drawn by the same person who did this one." the solution is obviously a second ai that does style transfers. that way you can make it all look like a lovely picasso
|
|
#
?
Sep 21, 2022 02:10
|
|
|
Objective Action posted:If you have an NVIDIA ~2070 or better* you can run Stable Diffusion locally with a good UI and bells and whistles with https://github.com/AUTOMATIC1111/stable-diffusion-webui. It's pretty close to a one click install. Used this guide, thanks. Still fiddling with the settings, most of my attempts still come out as nightmare fuel. My friend linked me these useful previews of artists and styles it can emulate, different samplers and complexity comparisons. Artists: https://imgur.com/a/2Fi9q5G Film types: https://i.redd.it/11rot12akmj91.jpg Sampler comparison: https://i.redd.it/uy2fp799wmj91.jpg Sampler vs. steps: https://i.redd.it/ud12agb7goj91.jpg
|
|
#
?
Sep 21, 2022 09:41
|
|
|
This thing will put every political cartoonist out of a job.
|
|
#
?
Sep 22, 2022 20:52
|
|
|
Hasn't human-generated art been bad for years now anyway https://twitter.com/spectator/status/1357000445219241986
|
|
#
?
Sep 22, 2022 23:20
|
|
|
Personally the main thing I use them for is making art of D&D/PF2E characters
|
|
#
?
Sep 23, 2022 01:39
|
|
|
Midjourney is pretty great. A former coworker of mine joked she was happy to get some unexpected mail then posted a picture of a letter from a cremation service. I put in "happily being cremated" and it gave me this: When Laura Loomer was ahead in her race I put in something like "Laura Loomer wearing a crown" and it surprised me with this  Then she lost and started crying so  They can't all be zingers. Also liked this chocolate and vanilla swirl anime girl  I burned though all my free credits pretty fast. Was tempted pay for a subscription.
|
|
#
?
Sep 23, 2022 06:37
|
|
|
 Edit: Same seed, slightly different settings and resolution. 
SplitSoul has issued a correction as of 15:06 on Sep 23, 2022 |
|
#
?
Sep 23, 2022 14:44
|
|
|
|
| # ? May 2, 2024 17:16 |
|
|
Pepe Silvia Browne posted:What I really wish I could do is take one style of output that I like and apply it to other prompts. It does a good job of mimicking existing styles to an extent, but when I'm trying to create 10 images that all depict different characters/scenes from a story, I want them all to look like they were drawn by the same hands. I can get them all to look like "digital paintings" or "35 mm macro photography" or "pulp comic art," I can even get them all to look like one specific artist's style. But I can't take a style that the AI has created as an output and say "okay, now make this new prompt look like it was drawn by the same person who did this one." There are some ways of refining a result you like: Img2img is one, where you take an existing image (which can be something you've already generated with an ai) as a starting point and then have the ai make variations of it. Another is a process called textual inversion where you train the ai using a few samples of the thing you want to replicate and then use that as a word-like in your prompt. Both of these require either a lot of trial and error or several samples of an existing thing, so there is definitely room for improvement where you could just go "oh, i like that one, but zoom out and turn the subject 45 degrees clockwise"
|
|
#
?
Sep 23, 2022 15:04
|
|
