Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[frap]

DeviantArt, forgetting who their entire userbase is, has announced their own AI art tool DreamUp.


They automatically opted in everything on the website, and the users had to manually opt-out each art submission, one-by-one.


So they added a new way that's less time intensive, still to opt-out rather than opt-in.


And started a livestream to discuss DreamUp.


But limited who can participate in chat.


It's going well.

 

[The Line]

DeviantArt, forgetting who their entire userbase is, has announced their own AI art tool DreamUp.
View attachment 3849075

They automatically opted in everything on the website, and the users had to manually opt-out each art submission, one-by-one.
View attachment 3849081

So they added a new way that's less time intensive, still to opt-out rather than opt-in.
View attachment 3849057

And started a livestream for discuss DreamUp.
View attachment 3849063

But limited who can participate in chat.
View attachment 3849069

It's going well.
View attachment 3849078

After enormous backlash, multiple members of the deviantart team held a twitter space with RJ Palmer. RJ is a professional artist who has worked on projects like detective pikachu. He is most well known for his tweets about art-related things, large social media following, and vocally dislikes AI image generators.


(Link/Archive)



The twitter space was deleted shortly after for unknown reasons but is archived below and was basically a 1 man Q&A.

audio archive

update:
Inclusion in AI art is now opt-in for all deviantart uploads

(Link/Archive)

 

[AmpleApricots]

Probably not that major. You don't scrape indiscriminately and there's a pre-classification step too (methods for which will also continue to improve) so if they suck they're basically in the same metaphorical bucket as whatever low quality amateur art makes it in there that nobody wants to gen anyway, and if they infiltrate sources you aren't expecting to contain non-human art then they're probably close to being indistinguishable anyway which is part of that self-selection you mention, yeah. In terms of feedback effect, I'd guess the bulk of art is from generic artists nobody could name who are imitating one another anyway, so it's probably either neutral or possibly even a positive effect insofar as it drifts away from that (especially given the benefit of users selecting "creative" flukes that stand out).

Validation schemes are a whole separate thing, there's a bunch of them and they do stuff like try to index human aesthetic preference, and those'll keep developing in their own direction too. Ultimately a self-training process (like how chess bots are trained by playing against themselves) might even take over entirely? People can only make so much art just like there can only be so many real chess matches in the book; imagine simulating a thousand years of speculative artistic development according to some art↔preference model in a day.

The dataset it was trained on actually isn't *that* well tagged and pretty much the definition of "random pictures off the internet", it certainly is diverse though. LAION-5B, by some german uni iirc. They didn't use the entire thing, just a filtered subset, but still. Stable Diffusion is notable mostly on being so incredibly efficient, not being necessarily the best possible. There'd be quite some room for improvement if you'd throw more computing power at it. That's why I'd just stop trusting pictures on the internet at this point.

The problem I could see with a self-training process is that you eventually end up with "optimized" results (= same-y art) this doesn't have to be necessarily bad though. Different models would be like different artists, having a specific style. It seems to be something people would want to begin with, as popularity of stuff like hypernetworks and dreambooth showcase. People don't want random art, the ideal situation would be "predictable" art. (dreambooth btw. is really fascinating)

vocally dislikes AI image generators

I think they all do. It's funny to see the different reactions to AI generated creativity. Writers barely reacted to text generation, the few that did basically said "cool". Artists have been vocally shitting themselves since day one. I'm sure eventually there will be legal regulation but seeing how the pictures actually get created and how such models get trained, if it's treated honestly there's a fat chance for artists being "legally" able to opt-out of having their art crawled for such purposes, and if you'd give them the capability legally, it'd also open a huge can of worms for other content.

I'm glad SD is out there and free. If it was in the hands of some organization it'd just end up getting censored and gatekeeped. We've seen the behavior of OpenAI and how these tech companies basically cannot be trusted.

 

[Involuntary Celebrity]

The dataset it was trained on actually isn't *that* well tagged and pretty much the definition of "random pictures off the internet", it certainly is diverse though. LAION-5B, by some german uni iirc. They didn't use the entire thing, just a filtered subset, but still. Stable Diffusion is notable mostly on being so incredibly efficient, not being necessarily the best possible. There'd be quite some room for improvement if you'd throw more computing power at it. That's why I'd just stop trusting pictures on the internet at this point.

Yeah I've skimmed the paper, but I figured the context (and so my answer) was more talking in the sense of stuff like NAI which continues training on stuff like the danbooru corpus. That's a lot more focused and makes more sense to be worried about re: downstream artistic pollution. Stuff like LAION is a standard for stuff like S-D trying to beat competitive benchmarks but (though that part of it is still relevant to my point about classification/validation improving in the future if someone is worried about some sort of line between pre-AI and post-AI pollution because the heuristic value of that is hypothetically different if you are) beyond a functional threshold real users mainly care about waifu aestheticism.
Funny right after I posted that I saw this: https://openreview.net/pdf?id=lRUCfzs5Hzg

 

[Markass the Worst]

DeviantArt, forgetting who their entire userbase is, has announced their own AI art tool DreamUp.
View attachment 3849075

They automatically opted in everything on the website, and the users had to manually opt-out each art submission, one-by-one.
View attachment 3849081

So they added a new way that's less time intensive, still to opt-out rather than opt-in.
View attachment 3849057

And started a livestream to discuss DreamUp.
View attachment 3849063

But limited who can participate in chat.
View attachment 3849069

It's going well.
View attachment 3849078

Isn't DeviantART the same site that alerted people if their art was being sold as NFT's? What person thought this wouldn't piss everybody the fuck off?

 

[Stephanie Bustcakes]

After enormous backlash, multiple members of the deviantart team held a twitter space with RJ Palmer. RJ is a professional artist who has worked on projects like detective pikachu. He is most well known for his tweets about art-related things, large social media following, and vocally dislikes AI image generators.

View attachment 3849297
(Link/Archive)

View attachment 3849351View attachment 3849363

The twitter space was deleted shortly after for unknown reasons but is archived below and was basically a 1 man Q&A.

audio archive

update:
Inclusion in AI art is now opt-in for all deviantart uploads
View attachment 3849435View attachment 3849438
(Link/Archive)

Changing it to opt-in won't solve all the problems. There are people that upload other people's art without their permission. (There is quite a bit of stolen artwork on DeviantArt.) If they opt-in then that art will be used in their database without the original artist's permission.

 

[frap]

Isn't DeviantART the same site that alerted people if their art was being sold as NFT's? What person thought this wouldn't piss everybody the fuck off?

Yeah artists were praising them a few months back and now they're deleting their accounts.

I'm thirty minutes into the twitter space audio archive and the DA people are mostly just speaking over RJ, negating the entire point of asking him to be there. On two occasions the CEO (who sounds like he's eating) has been frustrated enough to almost outright say that he needs RJ to push this DreamUp thing.

 

[The Line]

I don't know anything about coding or advanced technology, but is there any possibility for a built in anti-bot, anti-whatever the hell the AI machine is, thing for websites to add? I'm imagining something similar to an anti DDoS or bot page. The images are probably already in a giant database, but I'm very surprised I haven't heard anyone talking about the possibility for something like that catered specifically to prevent image generation from grabbing new images from a select site, given the backlash it's received.

I'd say there's potential for big money for whoever makes that and either markets their own site as having a protection or sells it to other image gallery sites. Could be the next "captcha" if it's actually possible.

 

[Sinner's Sandwich]

LMAO Deviantautists crying about AI art while fanarts can be considered stealing too and Deviantart is full of fanart. Artists on Twitter, Tumblr etc. are earning money with creating art of characters that don't belong to them. I don't think they want to open that can of worms.

 

[Kendall Motor Oil]

o prevent image generation from grabbing new images from a select site, given the backlash it's received.

Not possible as you need your images accessible from a web browser. You may be able to detect if an artist's art was used in the model but I don't think there is a legal precedent and everyone steals someone else's styles.

 

[gildersleeve]

What I find interesting is despite the backlash. There isn't anything preventing an organization or individual from scraping these sites for images. And then sorting and tagging them for training. The only reason this stopped was because it was that users platform. I'm sure models have already sorted through art via DA, newgrounds and the like.
Actually now that I think about it. I wonder what hellscape would be created if you trained a model on livejournal or tumblr. Or even specific fanfiction communities. Hell at that point the guy still writing that SSB:B fanfic could train a model on just his single work alone and have it forever spit out new chapters.

 

[SomeBirds]

Is anyone else training their own AI?

Here's some of the stuff I've generated that's unrelated to what we've been training for just 'cause I liked these. The one with the lady might have been impacted a bit by the training. I was running a prompt test after a training session and got it. The bird was just for funsies.

 

[supremeautismo]

LMAO Deviantautists crying about AI art while fanarts can be considered stealing too and Deviantart is full of fanart. Artists on Twitter, Tumblr etc. are earning money with creating art of characters that don't belong to them. I don't think they want to open that can of worms.

It is pretty funny that Deviantart thought this was a brilliant idea while the entire community of spergs has freaked out about tracing since the start of the website.

NOT MY CHARACTERS!! OC DONUT STEAL!!

 

[Troonos]

DeviantArt, forgetting who their entire userbase is, has announced their own AI art tool DreamUp.
View attachment 3849075

They automatically opted in everything on the website, and the users had to manually opt-out each art submission, one-by-one.
View attachment 3849081

So they added a new way that's less time intensive, still to opt-out rather than opt-in.
View attachment 3849057

And started a livestream to discuss DreamUp.
View attachment 3849063

But limited who can participate in chat.
View attachment 3849069

It's going well.
View attachment 3849078

What's the problem? This is fully transformative to satisfy Fair Use. It's creating entirely new content. Artists cope, seethe, dilate.

 

[Involuntary Celebrity]

It's just dumb; they should have known their users better.
I mean they're sitting on a goldmine of data so absolutely exploit it but they should have thought about how people would react to the roll-out. Like giving people an opt-out button guarantees they're going to freak out about having to opt-out, but an alternative since IIRC it's completely within their terms of use to do this (not that I've looked at it in 20 years) was to just use it all and tell everybody to get fucked.
Less discussion, less engagement (especially versus compelling people to fume for hours while they angrily opt-out one post at a time lol), and people would accept it faster like our bitch consumer conditioning tells us to.

I don't know anything about coding or advanced technology, but is there any possibility for a built in anti-bot, anti-whatever the hell the AI machine is, thing for websites to add? I'm imagining something similar to an anti DDoS or bot page.

Y'know that's actually a more interesting question than it seems. You maybe could conceivably design anti-models to do some imperceptible-to-the-audience image processing that would sabotage use in a dataset. Training that adversary would be on a per-system basis and probably mutually exclusive per-image though, and there'd be ways to sanitise it so it would just become a bit of an arms race, but deterrence is a possibility if you make it known you're deliberately planting poison pills so maybe you wouldn't be someone's first choice to scrape.
I don't think it'd be very feasible or worthwhile though and ironically it'd probably barely work against current-gen systems versus future ones training at full res.

 

[AnOminous]

Like giving people an opt-out button guarantees they're going to freak out about having to opt-out, but an alternative since IIRC it's completely within their terms of use to do this (not that I've looked at it in 20 years) was to just use it all and tell everybody to get fucked.

It's still a douche move that will anger a lot of people.

 

[Solid Hyrax]

I love seeing artists being ass hurt by AI art. It just shows me the real good artists who are embracing the future of technology and the Luddites who have shit art and will never manage to improve more than what AI is capable of.
All the anger is from people who won't be able to monetize their work. What are you drawing for? Money? Get real. It's enough that people had to deal with crazies. Some just want out and get their art in peace.

 

[get_ur_gamon]

I love seeing artists being ass hurt by AI art. It just shows me the real good artists who are embracing the future of technology and the Luddites who have shit art and will never manage to improve more than what AI is capable of.
All the anger is from people who won't be able to monetize their work. What are you drawing for? Money? Get real. It's enough that people had to deal with crazies. Some just want out and get their art in peace.

At least the luddites had respectable jobs, rather than chasing coomer cash.

 

[AnOminous]

What are you drawing for? Money? Get real. It's enough that people had to deal with crazies. Some just want out and get their art in peace.

They're mad this will seriously cut into their coom commissions when any idiot can tailor make their own gross porn.

 

[Eddie Riggs]

Sci-fi, fantasy-ish stuff