VR-Man said:
It's not AI. Don't worry.
Thanks. I just have terrible eye for these sorts of things so I wanted to be careful.
Posted under General
This has AI vibes to it. Looking through the artist's twitter, it looks like they mess around with AI generation while also doing their own sketches.
All posts by user #1143214.
There's currently a debate in the comments of post #7205502 on whether the AI-generated or AI-assisted tag should be used for the posts.
post #7209087 and post #7209075: although the artist may not have tagged it as such on Pixiv, they have a very extensive set of AI-generated images. I uploaded those two without looking more closely at the artists' other works.
Hey guys, how are you doing?
Blank_User said:
All posts by user #1143214.
There's currently a debate in the comments of post #7205502 on whether the AI-generated or AI-assisted tag should be used for the posts.
Apparently, the uploader is also the responsible behind the images, because in the comments debate, she shares the link to a Twitter account. And the source of the images directs to a Pixiv profile. Both of these profiles have the same name, and same profile picture. Now, coming back to the comments, it seems like she made a tweet confessing the use of AI-assistance to generate her images, but it's now conveniently deleted.
Using Wayback Machine was useless, so I used Sowte, a Twitter Web Viewer, and I found some evidence. If you take the time to click and see the profile in Sowte, it contains the same profile picture and the same alias from the original Twitter account I shared before. Judge by yourselves.
EvergreenNights said:
post #7209087 and post #7209075: although the artist may not have tagged it as such on Pixiv, they have a very extensive set of AI-generated images. I uploaded those two without looking more closely at the artists' other works.
Thank you for being honest. Yes, the images are definitely AI-generated, because of the time-frame in between. The artist shared both images in their Instagram, and as you can see, both images have only hours of difference.
nonamethanks said:
Those posts are obviously 3D, not ai-assisted. In post #6308255 or post #7127709 you can even still see the 3D background, for example.
So should the AI-assisted tags be removed from their art? It seems there’s no basis for it other than their art style looking AI-ish.
This artist's posts all seem to be AI-generated. All 512x768, lots of weird image artifacts, and their pixiv has some pretty egregious examples. They aren't tagged as AI-generated on pixiv but I think it's very obvious.
rockfish said:
This artist's posts all seem to be AI-generated. All 512x768, lots of weird image artifacts, and their pixiv has some pretty egregious examples. They aren't tagged as AI-generated on pixiv but I think it's very obvious.
Those definitely look AI, lots of artifacts too.
rockfish said:
This artist's posts all seem to be AI-generated. All 512x768, lots of weird image artifacts, and their pixiv has some pretty egregious examples. They aren't tagged as AI-generated on pixiv but I think it's very obvious.
They say they use IbisPaint, which is a drawing app that offers various AI tools along with traditional ones in the drawing canvas. You can see an example of those features here.
It's been explained in the comments of this ai-generated tagged post that it's considered "assisted" actually, but how much retouching has to be done for it to be considered ai-assisted instead of generated? Has enough retouching been done on this one for me to swap the tags? People can't seem to decide which it is.
Updated
Maiden_in_Orange said:
It's been explained in the comments of this ai-generated tagged post that it's considered "assisted" actually, but how much retouching has to be done for it to be considered ai-assisted instead of generated? Has enough retouching been done on this one for me to swap the tags? People can't seem to decide which it is.
In the end it's the admins' judgement to make, but personally I do not agree with letting images off with an assisted tag when they're still visibly overwhelmingly AI. It's all machine jank and no human input (that we can see) to offset it. Nothing about the hair spread out on the right makes sense, continuity collapses under the hair ornament, and some elements such as the hair outline and back share a single continuous line which is typical of AI but not how a human creative process works.
Diet_Soda said:
In the end it's the admins' judgement to make, but personally I do not agree with letting images off with an assisted tag when they're still visibly overwhelmingly AI. It's all machine jank and no human input (that we can see) to offset it. Nothing about the hair spread out on the right makes sense, continuity collapses under the hair ornament, and some elements such as the hair outline and back share a single continuous line which is typical of AI but not how a human creative process works.
Personally, I wholeheartedly agree on this. Which makes the whole thing more baffling, considering the other "works" by this "artist" were deleted for being AI without any complaint (and even has their artist page described as such). I'll hold on flagging for now, but the urge is very tempting, as I personally do think ai-generated is the more accurate tag than ai-assisted.
This is a good reminder for me that nobody who posts in this thread actually has any idea of what AI looks like.
I asked the artist in question for details, and he was gracious enough to actually send me the original AI-generated image. here it is. Open it side by side with the one we have and go back and forth with the tabs.
Diet_Soda said:
It's all machine jank and no human input (that we can see) to offset it. Nothing about the hair spread out on the right makes sense, continuity collapses under the hair ornament, and some elements such as the hair outline and back share a single continuous line which is typical of AI but not how a human creative process works.
Ironically all of the things you have an issue with were not present in the original generation. The back, the hair ornament, the hair were all completely retouched. Nothing of what you said actually makes any sense once you compare the original and the edited version, you just gaslit yourself into thinking you saw things that were never there. If you compare the original and the edited one side by side you'll realize they're pretty much two different pictures.
Now, of course you could claim that the artist only used img2img or another automatic tool instead of drawing by hand, and it's still a fully ai-generated picture. But again, where's the proof? Which spots can you pinpoint as ai-generated, and which as hand-drawn? And the artist this time was nice enough to reply, but what if they don't? Who's to decide whether a user's vibe is good enough?
This topic has gone on in circles for long enough, and I've seen enough witch hunts that turned out to be straight-up lies, or stuff like dikko being mass tagged with ai-assisted on posts that are obviously just traced 3d, so here's the mod post that we should've made months ago.
ai-generated content was banned to avoid getting flooded by thousands of samey-looking images outputted in an instant with zero effort like what happened with Pixiv or Deviantart. That's it. It's not a moral statement, it's not a blanket ban on anything that has had any degree of machine input, as people in this topic would have us believe. It's a rule to preserve the visual quality of our art gallery and to avoid overloading our already overworked approvers.
ai-generated does not mean "I personally don't like AI so I'm going to pull out of my ass that this picture is 90% AI and flag it". It does not mean "the background or the colors are AI generated so it should be deleted". ai-generated means that the image is fully ai-generated. It means that the image was outputted by a model and immediately uploaded to danbooru without any actual redrawing.
To what degree are ai-assisted images allowed? As with anything involving the mod queue, the answer is: it depends on how good they look.
ai-assisted should be treated like anything under the "Borderline Content" section of help:upload_rules (and that's where I'm going to add it to): stuff that you probably shouldn't dump en masse as it's going to make some people angry, but is not strictly against the rules because at the end of the day if it can pass the queue then it can stay.
To keep matters short:
If it doesn't have immediately obvious artifacts, mangled hands or the usual telltale signs of AI generation, it can go through the queue.
If it fools one or multiple approvers, it can go through the queue.
If it's good enough to pass the queue, it can go through the queue.
If the only fault you can find is that the hair or eye color is wrong, it can go through the queue.
If you can't tell if something is ai-generated vs ai-assisted, it goes through the queue. A good indication is whether you want to post it in this thread. If you're so unsure that you need to ask others for opinion, then it can go through the queue.
If you have to take your microscope out and write a dissertation on why no true homegrown organic artist would ever draw a hand at a 47% degree angle to convince people that something has had machine input at some point in its life, then it can go through the queue.
If an artist has "AI" in bio, that's not enough proof to say something is generated vs assisted. Believe it or not, but most sites and people out there don't make the distinction we do about ai-assisted vs ai-generated. A shitty bing generation and something that was fully hand-traced out of an ai image and an img2img iteration are all under the same tag on sites like Pixiv or Twitter. You need to use your eyes to look at the picture, not make up your own logic about someone else's idea of what a word means.
If people end up actually arguing and disagreeing on whether something is generated vs assisted (ignoring trolls and clueless neophytes), then it can go through the queue and approvers will decide the fate of the picture based on its quality as with any other post.
If it fools a contributor and you think it's not good enough, you can flag it so that it goes through the queue and gets put to the test, just like you would any other post that bypassed the queue and that you think is not up to our quality standards.
If you are an approver and you accidentally upload or approve something that you then end up thinking is AI-generated, then you should flag it (if you think it's poor quality), you shouldn't directly delete it. But the rules above apply to you too: you should flag it because it's poor quality, not because you think it's AI-generated.
If it looks like post #7201892 or post #7194647 or post #7175810 or any of those standard blurry one-of-a-million outputs, or if it has the classic glossy look of mass-generated AI such as what you'd see on AIBooru (stuff like this to be clear), it can be directly deleted. That's what the rule was meant to keep out.
If it looks like post #7215649, post #7215503 or post #7035323 then it's allowed to go through the queue and it should be treated like any other post, approved or left to die based on its visual merits, and approvers should not directly delete these.
If the modqueue decides that a post belongs to the site, then it belongs to the site. This applies to any kind of content, including ai-assisted.
The modqueue is the core of this site, and approvers are supposed to be a distilled version of the general userbase. If enough people think an image belongs to the site, then it belongs to the site. That's how it's always worked, even back in the days when we had things like furry or yaoi or gigantic breasts as explicitly forbidden content.
Flagging is obviously fine, as everyone makes mistakes, but I don't want to see people (especially approvers) fighting over technicalities just because they want a post gone on moral grounds.
If you're an approver and you personally don't like AI, that's ok, you can blacklist ai-generated and ai-assisted to avoid seeing them the queue. We don't force people to look at content they don't like, and we explicitly tell them to only approve posts they personally like or would upload themselves.
But the modqueue is not your playground, and you alone do not get to decide whether something is fully banned. If banned content was determined by majority vote then the first thing to go would not be ai-assisted/ai-generated, it would be futanari and yaoi.
To be clear, this is how the rules on AI were always intended to be interpreted. The problem is that technology advanced fast enough that it came back to bite us in the ass after less than two years, not a decade, and people started bastardizing them to the point where now you actually have to prove that something is allowed, rather than it's banned. I hope you can see that it was not a sustainable situation in the long term.
We've gotten to the point where contributors and approvers alike genuinely cannot tell AI content apart from normal art, and it regularly makes it past the queue and multiple approvers without issue. Every time I look at order:rank there's at least half a dozen untagged obvious AI-assisted stuff that nobody cares to tag, let alone flag.
Even with ai-generated fully banned, 0.3% of all of our uploads this year were ai-generated or ai-assisted: (ai-generated or ai-assisted) age:<1y returns 3665 posts, vs 1155 from two years ago (0.1% of uploads in the same timeframe). It was a 200%+ growth compared to the overall 20% growth of all uploads, ten times the growth in short. And that's of course not accounting for the possibly thousands of untagged ai-assisted art that nobody bothered to tag. What do you think that's going to look like in one or two years? Do you really think it's up to us to stop it?
I posted an example of an app integrating AI coloring above, IbisPaint. It's not an isolated case. Photoshop recently revealed generative fill, and Krita has a plugin for it. This stuff is only going to become more and more prevalent as these tools get wider adoption and new artists replace the old ones.
We are, at the end of the day, a repost site, one that values attribution and quality yes, but whose primary purpose has always been one and only one: to host and categorize high quality art. Expecting us to only host fully human-drawn art is quickly becoming equivalent to expecting us to only host traditional media.
We are, by consequence our very nature, a focused and filtered reflection of the greater art community out there, and any trend, any tool that artists decide to adopt will end up getting reflected on danbooru. We regularly see this with app filters or memes. If something becomes popular outside of danbooru then it ends up flooding us too.
If a large amount of artists decide to start using AI tools to create their art, then sucks to be us but that's going to end up reflected in our gallery too, whether we like it or not.
Updated
nonamethanks said:
Ironically all of the things you have an issue with were not present in the original generation. The back, the hair ornament, the hair were all completely retouched. Nothing of what you said actually makes any sense once you compare the original and the edited version, you just gaslit yourself into thinking you saw things that were never there. If you compare the original and the edited one side by side you'll realize they're pretty much two different pictures.Now, of course you could claim that the artist only used img2img or another automatic tool instead of drawing by hand, and it's still a fully ai-generated picture.
Well I am. I thought we were all already operating under the knowledge of img2img as a standard tool for the more ambitious prompters to polish an image before publishing it, and I fully stand by the details I mentioned not being human.
Again, I'm not going to pointlessly argue against admin policies on the matter, but I feel like the "nitpicking" allegedly taking place ITT is being grossly misrepresented. If, until today's clarification, the minimum threshold for a rule-breaking AI gen was three extra arms anyone can notice at a glance, then there would be no need for this thread to exist in the first place.
Diet_Soda said:
Well I am. I thought we were all already operating under the knowledge of img2img as a standard tool for the more ambitious prompters to polish an image before publishing it, and I fully stand by the details I mentioned not being human.
Again, I'm not going to pointlessly argue against admin policies on the matter, but I feel like the "nitpicking" allegedly taking place ITT is being grossly misrepresented. If, until today's clarification, the minimum threshold for a rule-breaking AI gen was three extra arms anyone can notice at a glance, then there would be no need for this thread to exist in the first place.
The crux of the matter is very simple: Danbooru uses a moderation queue to determine whether posts belong here. But the vast majority of our approvers and contributors cannot tell AI-assisted posts and human-made posts apart reliably anymore. Can you provide an objective criteria that people can follow, that can tell whether a post is fully AI-generated or underwent human retouching?
As shown by my example above and the 70 pages of this topic, it's all guesswork and gut feelings. The artist claims they retouched details with photoshop, so it's their word against yours. Can you see how that is an impossible way to deal with this content for us? That's not a valid way to moderate millions of posts every year.
We're already at the point where any policy we could make against ai-assisted content does not matter anymore. I wasn't kidding when I mentioned order:rank. Open it and look at it right now. Are post #7215297, post #7216043, post #7212769, post #7212171, post #7215750 ai-assisted? If yes, to what degree? If not, why? Can you tell without opening the artist's gallery and checking their past style before 2022?
What percentage of acceptableness would we even come up with? How would we verify it? If we say that an image is allowed if at least 50% of it has been retouched, how do we verify it? If the background is fully AI generated and takes 60% of the picture and the character in front is fully hand-drawn, do we allow it? How do we define retouching? Does it have to be an artist using specific tools like photoshop without generative help? Again, how do we prove that they did? And what do we do now that photoshop, gimp, krita and any other software offers plugins or even integrated way to use generative tools inside the program itself?
What if they generated lineart and hand-colored it? What if they fully traced an ai-generated image? What if they spent a week refining every single detail with img2img? Why then should we allow some shitty touhou chibi sketch that took five seconds for someone to draw and was approved only because it's touhou, and not that?
If an image is obviously AI then that's already poor quality for most people. The constant exposure to the same glossy AI style over the past months has ensured that the default style associated with 2022-era novelAI will forever equal "AI slop" in the eyes of most danbooru users. And that's fine, that's the kind of content we wouldn't want either. But you cannot honestly look at post #7182107 and tell me that it's objectively poor quality or that there's anything wrong with the image that would tell you it's AI at first glance. If you had looked at it in 2020 you wouldn't have batted an eye.
The threshold on whether a post can be submitted to the queue has to be the same that any other post undergoes, not because of bias, but because we have no choice. Look at the kind of stuff that people regularly approve: approver:any ai-generated. Even stuff like post #7064608 or post #7012036, which is as blatant AI as they get nowadays, still manages to catch people unaware. If stuff like this is not spotted by approvers what do you think is going to happen in a few months (not even years)?
Maybe you truly can tell these posts apart. Maybe you are a Stable Diffusion or novelAI developer and know how to spot this stuff by heart. But most of our userbase, and most importantly our approvers and uploaders are not. The modqueue is already a hard filter to most people, we can't have another committee sitting above that deciding what's allowed and what's not based on some inscrutable criteria that only they can figure out and that nobody can verify.
Updated
All of their works look very AI, although I don't know if this post in particular is assisted or not.
Can't agree more with nnt here.
And now we finally have a clear stance on how to deal with ai-assisted. I've been waiting for this for ages. Because either you completely allow it or completely disallow it. You can't make a treshold.
As to why this stuff still gets approved: Accidents or these apiprovers genuinely don't care. In the latter case, it's not valid to allow AI more and more, but instead tell these approvers off (if there is a pattern).
Updated
