Concern about DALL-E 2 and other Image generators

[JizzyJazz]

How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.

You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

[HellBladePhoenix]

I've had that thought since Deep fakes where a thing. Just like Dall-e 2, im worried that it'll used for nefarious things. However as long the Dall-e 2 developer put on limitations on the ai, such as purposely make it unable to construct or cartoonified celebs, then i don't think its something to worry about imo.

But if what i said isnt true then it'll happen much sooner then 5 years

[Aistis]

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

Still quite a while. They might look awesome and they might cause some harm and mayhem, but we're still magnitudes away from reaching such a point where there'd be no point for human artists to exist.

They base everything on collected data (read: stolen art from the internet) without reading much into emotion or storytelling. Sure, by pure randomness you could generate something like that, but good look generating at least one issue of a comic of about 20 pages with an entertaining story with quality drawings, with the ability to generate a great issue after issue. It won't.

[Zymbot]

They're funny, tho.

[Cerbskies]

AI will weed out the weak

[Deity-Donkus]

At 8/25/22 03:38 PM, Cerbskies wrote: AI will weed out the weak

Whos Al?

[WVDB]

Simple: watermark.

[guyfromKillerRAT]

Oh, no. I asked it to generate polaroids of oil paintings of rats with guitars.

Shit looks as real as a real ass painting. And since I asked for polaroids of it,

it comes with a white border. Which makes erasing the watermark very easy.

Let me show you what it gave me.

Kept the watermarks on, but as you could imagine, it would not be hard to remove them.

Maybe on images without borders, but even that's possible with some phone apps now.

I also asked it for alcoholic horses in dark abysses. It gave me this.

yeah. I could submit these to the art portal and pretend I made them. Nobody would know.

Well, they would now that I posted this. But had I not, I'd get away with it if I tried.

Photographs of human faces of course always kinda go wrong.

Also, don't ask it to write any words. It can't spell.

[Cerbskies]

At 8/25/22 04:04 PM, Deity-Donkus wrote:

At 8/25/22 03:38 PM, Cerbskies wrote: AI will weed out the weak

Whos Al?

Our overlords

[NitroOrient]

At 8/25/22 08:44 PM, Cerbskies wrote:

At 8/25/22 04:04 PM, Deity-Donkus wrote:

At 8/25/22 03:38 PM, Cerbskies wrote: AI will weed out the weak

Whos Al?

Our overlords

No They are the creators of the whole damn universe, it's just that we didn't know that... until now.

[Cerbskies]

At 8/25/22 08:59 PM, Alvesteffer19 wrote:

At 8/25/22 08:44 PM, Cerbskies wrote:

At 8/25/22 04:04 PM, Deity-Donkus wrote:

At 8/25/22 03:38 PM, Cerbskies wrote: AI will weed out the weak

Whos Al?

Our overlords

No They are the creators of the whole damn universe, it's just that we didn't know that... until now.

We could just be rampant code, a virus, plaguing a computer. What we percieve as time is miliseconds in the universe AI.

[Nabella]

Wow, what a coincidence, I was debating about making a thread on this exact topic for the past week and a half.

Has anyone here seen some of the images generated by Stable Diffusion, the free open-source model anyone can download and modify? With enough patience and a beefy enough computer, it seems like you can create some very impressive results.

The following images were plucked while I was lurking on /sdg/ on /g/ and it's associate threads on /vg/, and /h/.

Prompt that was attached to this specific lady above me.

00002-1371391821_Female_Pope,_Big_Tits,_Big_Boobs,_Oil_on_Canvas,_Highly_Detailed,_Intricate,_award_winning,_magic_the_gathering,_d&d,_top_rated, copy

This was just the cream of the crop, tho, there are definitely handfuls of images with weird (but admittedly entertaining) fuck-ups.

I'm very surprised it took this long for the rest of the internet to become aware of this tool, but that's probably because actually understanding how to install and use this, even with the retard guides provided, is difficult to understand if you're a tech dumbass like me.

[Nabella]

At 8/25/22 03:38 PM, Cerbskies wrote: AI will weed out the weak

Guess that's me. Damn.

[S3C]

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

It already exists. I don't really have much concern. It does cheapen the artist's job a bit...but whether they work hard on pen to tablet or use a complex algorithm, the end product is what matters. Good art is good art. Also AI generated and human art will always have their own distinct domains.

The ability to make lifelike AI photos has existed for some time too. @asiahlucytalitha pfp is AI generated

[Gimmick]

It already has to an extent. I believe a new field of research is training neural networks to detect fake images and doctored videos.

Perhaps in the future there'll be services that then detect whether a given image is fake or not.

[Nabella]

At 8/26/22 01:55 AM, Gimmick wrote: It already has to an extent. I believe a new field of research is training neural networks to detect fake images and doctored videos.

Perhaps in the future there'll be services that then detect whether a given image is fake or not.

Bro, I'm super anxious about the future implications this kind of technology has on an internet where it's already difficult for people to discern what's actually true or not. What's stopping some bad actors from picking this up, feeding it a curated image gallery of their opponents, and churning out propaganda to flood websites with at an inhuman rate?

Hold me.

[Dolorious]

At 8/26/22 01:09 AM, S3C wrote:

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

It already exists. I don't really have much concern. It does cheapen the artist's job a bit...but whether they work hard on pen to tablet or use a complex algorithm, the end product is what matters. Good art is good art. Also AI generated and human art will always have their own distinct domains.

The ability to make lifelike AI photos has existed for some time too. @asiahlucytalitha pfp is AI generated

Well at least it won't be able to animate that shit.

(inb4 it can...)

[SakuraStarNG]

At least I generated this image:

...which will be used in my upcoming Dave OS!

Expect the OS to be ready on either September or October this year!

[Gimmick]

At 8/26/22 02:27 AM, Nabella wrote: What's stopping some bad actors from picking this up, feeding it a curated image gallery of their opponents, and churning out propaganda to flood websites with at an inhuman rate?

Nothing. Given that GPT-3 is so convincing at appearing real --as long as you don't prod too deeply-- it's probably likely that it IS being used to generate propaganda.

On the plus side, maybe if it becomes commonplace enough that we recognize that it's being used for propaganda, then we can finally return to the old days of the internet where you should never trust anything that you read without verifying it yourself independently.

[S3C]

At 8/26/22 03:03 AM, Dolorious wrote:

At 8/26/22 01:09 AM, S3C wrote:

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

It already exists. I don't really have much concern. It does cheapen the artist's job a bit...but whether they work hard on pen to tablet or use a complex algorithm, the end product is what matters. Good art is good art. Also AI generated and human art will always have their own distinct domains.

The ability to make lifelike AI photos has existed for some time too. @asiahlucytalitha pfp is AI generated

Well at least it won't be able to animate that shit.
(inb4 it can...)

it can

it just takes magnitudes more computation power and input data

there's some deep learning algorithms for face swaps in video and it can work frighteningly well.

but like I said, it takes a lot of input data to train the model. and typically it's non-trivial and takes a certain degree of skill.

[misterbright]

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

it already is

[TurkeyOnAStick]

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

There’s a few comments on this pic that Midjourney may have been used. IMO I think something’s amiss as I find the hair and the hands a bit odd.

[Positron832]

At 8/27/22 01:55 AM, TurkeyOnAStick wrote:

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

https://www.newgrounds.com/art/view/kel-chan/retro-futurism-neotokyo-style
There’s a few comments on this pic that Midjourney may have been used. IMO I think something’s amiss as I find the hair and the hands a bit odd.

The author replied to one of the reviews saying it's 3d and photo manipulated. Who knows tho :/

[Skoops]

As an artist by trade, and kind of an enthusiast for AI generation, I think the panic is overblown. Yes, it makes very pretty images if you stay with it and coax it into making something coherent over dozens if not hundreds of human-curated refinements. Yes, sometimes it's freaky how well it "understands" your query every once in a while. I think people look at the best of the best of the best (to the human eye) of what you can eventually get with AI generation and assume that the AI even knows how it got there, or knows what you like about it. Since a lot of people are just looking at the results and aren't making them themselves, they don't see the hundreds of thousands of results that are absolute junk, the minuscule amount that come back even workable, and how even when you get something you like, because it doesn't know why you like it, refining it never gets easier. You are always at square one, rolling dice, and the more I use it, the less afraid of it I am.

That's all before talking about how this wild west phase is going to be short lived. As soon as AI starts approaching genuine broad appeal for use in an art pipeline, there's going to be a reckoning regarding how these AIs are trained. AI can't generate images mimicking art without being trained on a vast library of preexisting art to chop and screw. Most quality art and named styles are under copyright, creative commons or something like it. Right now, it's pretty clear that a lot of these generators, especially the free and public ones, are just crawling Google Images with your query terms, slapping things together with enough smudging and obfuscation that it can't be called plagiarism, then giving the rights to that image either to the AI company or the query-maker. That's not going to sit well with most copyright holders. Nintendo's going to see people using AI to get images of mpreg Mario and they're gonna say A: your for-profit AI service has to be stealing copyrighted material for it to know what Mario is and what he looks like, and B: that image of pregnant Mario belongs to Nintendo. Cease and desist.

The closer an AI gets to knowing what copyrighted art looks like, and the better it gets at replicating it, the more concerned that copyright holder is gonna be. It might eventually be decided that AIs can legally be trained however the engineers want, and everything it pumps out is legal and fair use, there's still going to be a long process where those kinds of legal disputes gets hashed out. We're a ways off.

But like, tricking people on the internet? You can do that now. I dunno how much that'll be good for, or for how long.

[JizzyJazz]

At 8/27/22 01:55 AM, TurkeyOnAStick wrote:

At 8/24/22 12:42 PM, JizzyJazz wrote: How much longer do we have left until the AI becomes so good that you could generate an image and upload it to the art portal without anyone batting an eye? I have a feeling that we're slowly heading in that direction with all these generators becoming so accurate.
You probably couldn't do it today, as you can still kinda tell what's AI generated and what isn't, but what's going to happen if we give the tech another 5 years to develop?

https://www.newgrounds.com/art/view/kel-chan/retro-futurism-neotokyo-style
There’s a few comments on this pic that Midjourney may have been used. IMO I think something’s amiss as I find the hair and the hands a bit odd.

i looked at some more of the pics and i think they might be right. the eyes look extremely weird, a telltale sign of midjourney or DALL E 2. this guy didnt draw that stuff at all

[Dolorious]

At 8/27/22 03:52 AM, Skoops wrote: As an artist by trade, and kind of an enthusiast for AI generation, I think the panic is overblown. Yes, it makes very pretty images if you stay with it and coax it into making something coherent over dozens if not hundreds of human-curated refinements. Yes, sometimes it's freaky how well it "understands" your query every once in a while. I think people look at the best of the best of the best (to the human eye) of what you can eventually get with AI generation and assume that the AI even knows how it got there, or knows what you like about it. Since a lot of people are just looking at the results and aren't making them themselves, they don't see the hundreds of thousands of results that are absolute junk, the minuscule amount that come back even workable, and how even when you get something you like, because it doesn't know why you like it, refining it never gets easier. You are always at square one, rolling dice, and the more I use it, the less afraid of it I am.

That's all before talking about how this wild west phase is going to be short lived. As soon as AI starts approaching genuine broad appeal for use in an art pipeline, there's going to be a reckoning regarding how these AIs are trained. AI can't generate images mimicking art without being trained on a vast library of preexisting art to chop and screw. Most quality art and named styles are under copyright, creative commons or something like it. Right now, it's pretty clear that a lot of these generators, especially the free and public ones, are just crawling Google Images with your query terms, slapping things together with enough smudging and obfuscation that it can't be called plagiarism, then giving the rights to that image either to the AI company or the query-maker. That's not going to sit well with most copyright holders. Nintendo's going to see people using AI to get images of mpreg Mario and they're gonna say A: your for-profit AI service has to be stealing copyrighted material for it to know what Mario is and what he looks like, and B: that image of pregnant Mario belongs to Nintendo. Cease and desist.
The closer an AI gets to knowing what copyrighted art looks like, and the better it gets at replicating it, the more concerned that copyright holder is gonna be. It might eventually be decided that AIs can legally be trained however the engineers want, and everything it pumps out is legal and fair use, there's still going to be a long process where those kinds of legal disputes gets hashed out. We're a ways off.

But like, tricking people on the internet? You can do that now. I dunno how much that'll be good for, or for how long.

The thing can't make obscure art of oc's or anything that is brand new as you said.

Like if you're making a new character or the character you made is not well known.

That AI is going to give completely different results.

[Skoops]

At 8/27/22 04:20 AM, Dolorious wrote:

At 8/27/22 03:52 AM, Skoops wrote: As an artist by trade, and kind of an enthusiast for AI generation, I think the panic is overblown. Yes, it makes very pretty images if you stay with it and coax it into making something coherent over dozens if not hundreds of human-curated refinements. Yes, sometimes it's freaky how well it "understands" your query every once in a while. I think people look at the best of the best of the best (to the human eye) of what you can eventually get with AI generation and assume that the AI even knows how it got there, or knows what you like about it. Since a lot of people are just looking at the results and aren't making them themselves, they don't see the hundreds of thousands of results that are absolute junk, the minuscule amount that come back even workable, and how even when you get something you like, because it doesn't know why you like it, refining it never gets easier. You are always at square one, rolling dice, and the more I use it, the less afraid of it I am.

That's all before talking about how this wild west phase is going to be short lived. As soon as AI starts approaching genuine broad appeal for use in an art pipeline, there's going to be a reckoning regarding how these AIs are trained. AI can't generate images mimicking art without being trained on a vast library of preexisting art to chop and screw. Most quality art and named styles are under copyright, creative commons or something like it. Right now, it's pretty clear that a lot of these generators, especially the free and public ones, are just crawling Google Images with your query terms, slapping things together with enough smudging and obfuscation that it can't be called plagiarism, then giving the rights to that image either to the AI company or the query-maker. That's not going to sit well with most copyright holders. Nintendo's going to see people using AI to get images of mpreg Mario and they're gonna say A: your for-profit AI service has to be stealing copyrighted material for it to know what Mario is and what he looks like, and B: that image of pregnant Mario belongs to Nintendo. Cease and desist.
The closer an AI gets to knowing what copyrighted art looks like, and the better it gets at replicating it, the more concerned that copyright holder is gonna be. It might eventually be decided that AIs can legally be trained however the engineers want, and everything it pumps out is legal and fair use, there's still going to be a long process where those kinds of legal disputes gets hashed out. We're a ways off.

But like, tricking people on the internet? You can do that now. I dunno how much that'll be good for, or for how long.

The thing can't make obscure art of oc's or anything that is brand new as you said.
Like if you're making a new character or the character you made is not well known.
That AI is going to give completely different results.

True, and this presents another challenge to anyone trying to pull a long-term grift or any studio trying to develop consistent visuals. Try using different AIs to generate stylized designs; even with the same exact prompt entered dozens of times, rarely do you find two images that look like they could have been made by the same artist.

Is that caused by deliberate self sabotage to hide what preexisting images they're using? Is it because they only have so many pictures of so many things drawn in so many ways, and they just grab things with matching tags at random until you like one of them? Maybe both? Doesn't really matter. Either way, it means they might be able to fool you once, but it gets exponentially harder to fool you again every subsequent time.

[Dolorious]

At 8/27/22 05:47 AM, Skoops wrote:

At 8/27/22 04:20 AM, Dolorious wrote:

At 8/27/22 03:52 AM, Skoops wrote: As an artist by trade, and kind of an enthusiast for AI generation, I think the panic is overblown. Yes, it makes very pretty images if you stay with it and coax it into making something coherent over dozens if not hundreds of human-curated refinements. Yes, sometimes it's freaky how well it "understands" your query every once in a while. I think people look at the best of the best of the best (to the human eye) of what you can eventually get with AI generation and assume that the AI even knows how it got there, or knows what you like about it. Since a lot of people are just looking at the results and aren't making them themselves, they don't see the hundreds of thousands of results that are absolute junk, the minuscule amount that come back even workable, and how even when you get something you like, because it doesn't know why you like it, refining it never gets easier. You are always at square one, rolling dice, and the more I use it, the less afraid of it I am.

That's all before talking about how this wild west phase is going to be short lived. As soon as AI starts approaching genuine broad appeal for use in an art pipeline, there's going to be a reckoning regarding how these AIs are trained. AI can't generate images mimicking art without being trained on a vast library of preexisting art to chop and screw. Most quality art and named styles are under copyright, creative commons or something like it. Right now, it's pretty clear that a lot of these generators, especially the free and public ones, are just crawling Google Images with your query terms, slapping things together with enough smudging and obfuscation that it can't be called plagiarism, then giving the rights to that image either to the AI company or the query-maker. That's not going to sit well with most copyright holders. Nintendo's going to see people using AI to get images of mpreg Mario and they're gonna say A: your for-profit AI service has to be stealing copyrighted material for it to know what Mario is and what he looks like, and B: that image of pregnant Mario belongs to Nintendo. Cease and desist.
The closer an AI gets to knowing what copyrighted art looks like, and the better it gets at replicating it, the more concerned that copyright holder is gonna be. It might eventually be decided that AIs can legally be trained however the engineers want, and everything it pumps out is legal and fair use, there's still going to be a long process where those kinds of legal disputes gets hashed out. We're a ways off.

But like, tricking people on the internet? You can do that now. I dunno how much that'll be good for, or for how long.

The thing can't make obscure art of oc's or anything that is brand new as you said.
Like if you're making a new character or the character you made is not well known.
That AI is going to give completely different results.

True, and this presents another challenge to anyone trying to pull a long-term grift or any studio trying to develop consistent visuals. Try using different AIs to generate stylized designs; even with the same exact prompt entered dozens of times, rarely do you find two images that look like they could have been made by the same artist.

Is that caused by deliberate self sabotage to hide what preexisting images they're using? Is it because they only have so many pictures of so many things drawn in so many ways, and they just grab things with matching tags at random until you like one of them? Maybe both? Doesn't really matter. Either way, it means they might be able to fool you once, but it gets exponentially harder to fool you again every subsequent time.

I think it might be the ladder. The only way it can achieve what you described is for the AI to become sentient. Which then yeah, they'll be able to pull those grifts easy.

[DamnedByFate]

At 8/27/22 05:54 AM, Dolorious wrote: I think it might be the ladder.

Bloody sepps.

You mean latter. Ladder is the thing you use to climb up to your beloved's window.

[olskoo]

At 8/27/22 06:29 AM, DamnedByFate wrote:

At 8/27/22 05:54 AM, Dolorious wrote: I think it might be the ladder.

Bloody sepps.

You mean latter. Ladder is the thing you use to climb up to your beloved's window.

nobody remembers what sepps means