Donmai

Should we allow AI art?

Posted under General

This topic has been locked.

Hello, I would like to preface that I am the author of the Waifu Diffusion models which were trained on images and tags downloaded through this site to produce images from tags. I stand on the side of banning AI generated images on Danbooru. My reasoning for this is that with how image generative models are trained, an image model trained on Danbooru will only be a distillation of Danbooru. Therefore uploading AI generated images onto Danbooru wouldn't be contributing anything of substance even if the content is high quality since the images are just distillations of what's already on the site.

Jigsy said:

I did notice that StableDiffusion is classified as an artist, not as something else/the tool. Should this really be the case, though?

It's like classifying photoshop as the artist, in my opinion.

I think it might be better to classify Stable Diffusion as a medium instead, although that would open up a new can of worms to find out if writing a prompt for an AI image generator makes you an artist or not.

Jigsy said:

I did notice that StableDiffusion is classified as an artist, not as something else/the tool. Should this really be the case, though?

It's like classifying photoshop as the artist, in my opinion.

It's hard to make accurate analogies for AI art because it's new. The best one I can come up with is:
If someone were to upload an image created with Mematic, do we tag the guy who generated it?

I don't think the person running the generation should be tagged as the artist. Perhaps ai-generated should have the same function as artist request.

The person who generated the image should be tagged as the artist. Otherwise you lose all information about who made the image. At the very least, it allows you to find other ai-generated posts by the same person.

This would be like making koikatsu (medium) an artist tag and nuking the actual creator of each post. Or looking at a photograph and saying the creator is the camera, not the photographer.

evazion said:

The person who generated the image should be tagged as the artist. Otherwise you lose all information about who made the image. At the very least, it allows you to find other ai-generated posts by the same person.

This would be like making koikatsu (medium) an artist tag and nuking the actual creator of each post. Or looking at a photograph and saying the creator is the camera, not the photographer.

In that case should stable diffusion etc be meta (medium) tags instead?

Veradux said:

It's hard to make accurate analogies for AI art because it's new. The best one I can come up with is:
If someone were to upload an image created with Mematic, do we tag the guy who generated it?

The best analogy I can think of are NFTs, the typical Bored Ape and other derivatives are Ai generated. NFTs didn't explode in danbooru because they are low quality repetitive art and rapidly became an unpopular subject to the public opinion.
Posts featuring NFT art would problably have no artist tag.

Veradux said:

I don't think the person running the generation should be tagged as the artist. Perhaps ai-generated should have the same function as artist request.

Definitely agree they should not be tagged as the artist, because it might led to users believing these arts were drawn by them. It would feel like, tagging a commissioner as the artist in a commissioned art post, instead of the actual artist themselves.

mongirlfan said:

Definitely agree they should not be tagged as the artist, because it might led to users believing these arts were drawn by them. It would feel like, tagging a commissioner as the artist in a commissioned art post, instead of the actual artist themselves.

Indeed. Using evazion's photograph example, the photographer still has to do the work of choosing the angle, lens, zoom etc., setting the scene if required, dictating the pose of any models and so on. Even with a very descriptive AI prompt detailing all of these things and more, the user running the generation is just telling the AI how to do the work.

As for the main question this topic raises, my opinion is ultimately unchanged from when I first responded in the other thread, so I'll just copy my post from there.

I strongly dislike the idea of allowing 100% AI-generated art on the site, even if it's good enough. The entire decade and a half of Danbooru's existence until now has been dedicated to cataloguing years of manmade anime history; allowing AI art (even if it passes quality standards) spits in the face of this progress. It also opens two metaphorical Pandora's Boxes - uploaders may adopt a similar mentality to self-uploaders in this respect, being less able to judge the quality of AI art because they "made" it. Also, when someone creates a publicly available AI capable of making anime art that often passes our quality standards, users will be free to generate and upload it en mass. I can easily see this overtaking human-created art if it becomes prevalent enough, with more and more uploaders jumping on the bandwagon of infinite, instantly available uploads.

I have no issues with ai-assisted art - it's being used by humans as a tool, similar to 3D posing software or even a wooden posing doll. I also have no ideological issues with purely AI-generated art. However, it doesn't belong on Danbooru. If people want a curated gallery of AI art, then they should consider making an AIbooru.

I don't think purely ai-generated art should be on danbooru. Even if it's high quality, there has to be a point where we can decide the scope of the site. There is a lot of art that I like, but uploading something by Ilya Repin would be seen as off-topic even if its good. I want ai-generated art to be included in the same tier as stuff like headshops and nude filters and upscales.

Ai-assisted is a bit of a different beast, but I would want it in borderline. Something like these, where the AI did most of the "heavy lifting", should be at least scrutinized and discouraged.
https://twitter.com/Amou_Fushi2/status/1571781664438915073
https://seiga.nicovideo.jp/seiga/im11039642

Tbh I don't think AI generated art is a problem. If the problem appears, that the main page is flooded with AI art, we have to consider measurements, so that we users can hide it. Not like the blacklist does, more like a hard-hiding tool.
Right now discussing an "problem" that is 92 images big is IMO kinda weird on a website with over 5 million images.

Nacha said:
...

I disagree with this notion. It's good to discuss something before that something actually becomes an issue. Otherwise you end up in a situation where people are complaining about something that could have not been an issue at all. It's also clear that ai-generated art has been accelerating in it's popularity and upload frequency with roughly one third of the tag being from the last week or so.

I'll throw my 2 cents in and say on one hand I feel like art is art. However on the other there's the part of me that dislikes the idea of this place hosting human done art beside art done by an AI with some human prompting regardless of how high quality it may be. I like the idea of banning completely AI generated art and allowing AI-assisted art because there is still the significant human element in it. Yea I realize that some of this AI generated art is way better than most of the art that gets uploaded here or even on the Hot page, but so what? We don't allow every single type of art here.

Nacha said:

Tbh I don't think AI generated art is a problem. If the problem appears, that the main page is flooded with AI art, we have to consider measurements, so that we users can hide it. Not like the blacklist does, more like a hard-hiding tool.
Right now discussing an "problem" that is 92 images big is IMO kinda weird on a website with over 5 million images.

92 images but it is only a small number of people so far, however chances are as it gets both easier to use (devs have mentioned all sorts of models and such for it, including realistic and anime style) it's use will probably grow exponentially so it is worth considering our options now since there could well be a lot of debate.

I'm somewhat mixed myself, right now the program does seem to require some skill but if better templates became a standard feature it would likely become ridiculously easy for even a first time user to create quite good art. Then again if some people's fears are right, art could very well die and this is the only art so banning it might be pointless if 90% of people just stop

I'll make my statements here too since I'm also an active and practicing artist, and the topic is indeed of importance to those like me.

I'm not saying artworks done via AI programs and software are inherently abhorrent, especially since we artists also use existing material whether online or analog like someone had already said in this discussion, but there has to be a strict and well-defined limit to where and when we can consider AI creations here. And now is the best time to flesh out our plans regarding how we treat artwork done via AI programs since it will only be more problematic in the future (3-5 years, give or take).

Like many, we also see AI-generated art as a potential challenge to our trade, but in my opinion it is of concern if the scope of AI generation involves the "total shortcut" approach of someone typing a prompt and then letting the program "take inspiration" from existing works already online, especially if it's from image boards like how Waifu Diffusion operates. JP artists were already up in arms regarding a different AI art generator, Mimic, and the mere thought of having a program lift from works to "create new material" already led to at least one of them asking for a hard ban/takedown of their works from Danbooru because of such fears.

We can argue about whether or not we'll exclude all manner of AI art on the image board but also consider the fact that many of us wouldn't want to spend too much time on detailed background or composition and would definitely need a shortcut (case in point: post #5713941) so long as the artist had most of the input process. We can treat AI at this moment the same way how the Japanese treated firearms in the Sengoku Period: a new piece of tech that would cause an uproar, but eventually come to terms with it as long as the users (artists) know the line between positive assisting and plain laziness.

ArcieA said:

We can argue about whether or not we'll exclude all manner of AI art on the image board but also consider the fact that many of us wouldn't want to spend too much time on detailed background or composition and would definitely need a shortcut (case in point: post #5713941) so long as the artist had most of the input process. We can treat AI at this moment the same way how the Japanese treated firearms in the Sengoku Period: a new piece of tech that would cause an uproar, but eventually come to terms with it as long as the users (artists) know the line between positive assisting and plain laziness.

Yeah, in general I don't have problem with AI-assisted.
But if AI art gets to a point where making a beautiful artificial art is easy as pressing a button, it will certainly become a huge problem.

He's talking about people on Pixiv using AI to generate photorealistic CP (or people pretending it's AI when it's just CP with filters on). We've not had a problem with content like that so far, but we'll deal with it if it ever shows up here.

On one hand I think post #5713542 is good enough that, if you didn't look at the AI-generated tag, most people wouldn't think twice about allowing it; while it is un-anatomical around the armpit, there's far worse and more prominent bad anatomy out there. And I'd have reservations about allowing or banning art based on "meta" concerns outside the merits of images themselves.
But at the same time, it's a real threat because of how fast it is. Not the easiness, or amount of skill or effort, but the time. On an ordinary consumer level GPU, Stable Diffusion can crank out half a dozen images every minute or two. It would be incredibly easy to end up with the site getting flooded, and even saying "hold it to high standards" could still, potentially, be unworkable due to volume.
I'm really not sure. It has potential to make great stuff (I'm quite pleased with some landscapes I've generated, such as old ruins "drawn by Caspar David Friedrich", though they wouldn't really be the right fit for Danbooru even if they were human-painted) and banning them all regardless of quality doesn't sit entirely right with me. But it may be necessary. I guess my preference would be to allow it with strict standards for now, waiting to see how common it actually gets, and either continuing that way or banning depending on what happens.

iori98 said:
I guarantee you that the image used another artist as a base to create that work.

Almost certainly. In my playing around with SD, getting appealing faces requires pointing the network to an artist it knows, such as Bouguereau or Artgerm.
But real artists can do worse. Current AI works by pattern recognition, but it still combines those patterns in original ways, and nobody has actually drawn post #5713542 regardless of which artists were referenced in the prompt. There are human-made images on Danbooru that are outright traced from other artists (post #2644021 and others by that artist, for example). I don't think the derivative part is a problem.

Updated

I will give credit about flow, that was one concern that actually didn't occur to me at the time. I'm unfamiliar with how robust the Dan servers are, so I can't really hold a strong opinion in either direction, but if they start uploading them by the thousands I can easily see that being a major problem.

The artist sampling issue is pretty much irrelevant though. Real artists steal so often that more than a few can make a *living* just off of other peoples style. Dan is quality police, not style police.

RingyThingy said:

I will give credit about flow, that was one concern that actually didn't occur to me at the time. I'm unfamiliar with how robust the Dan servers are, so I can't really hold a strong opinion in either direction, but if they start uploading them by the thousands I can easily see that being a major problem.

I wasn't concerned about the technical side when it comes to volume. The problem I was thinking about was for the human side. When I said "flooded" what I meant was the possibility that you could end up with half the images on any given page being AI-generated.

1 2 3 4 5 6 7