Topic: How to know if it is ai work?

Posted under Art Talk

Since AI illustrations appear i've been so lost that I can no longer trust the artists I used to follow.

I love sharing my favorite posts on this site, but I spend too much time figuring out whether it's AI or not.
Here are three examples with Lonna:

https://www.furaffinity.net/view/57893556/ (NSFW)
https://www.furaffinity.net/view/57344449/ (SFW)
https://www.furaffinity.net/view/56939082/ (SFW)

I'm sorry if this has already been explained, but can you help me how to know the differences?
I really need advice to understand how to spot ai.

Updated

None of those look like A.I.. You may just be getting amped up over nothing by doomsday-esque ranting.

If something is AI, you'll probably be able to tell just by seeing multiple comments of people saying it's AI. They're everywhere, people have a major hard-on for detecting and condemning AI art.

If an artist you like is actually making art with AI and not themselves, they'll be found out before long, you won't even have to lift a finger. Just read the feedback.

for someone who was regurgitated with so much ai slop, none of these are ai-generated. they don't have common mistakes that contemporary ai generators that creates and struggles with (highly specific generic lighting, trouble with eyes, nipples, teeth, etc). artists' linework looks more organic and crispier than typical ai images i've seen.
they were also uploaded on furaffinity which vehemently against ai images, even if you scroll through their gallery and compare pieces, it doesn't seem to made with ai at all.

votp said:
None of those look like A.I.. You may just be getting amped up over nothing by doomsday-esque ranting.

You don't help me to understand the differences

I'm going to preface this by saying that there is no perfect, concise, and convenient way to detect what is fully AI-generated, what is a human's interpretation of an AI-generated image, and what is fully human art.

The only truly reliable way you're going to develop this kind of skill is if you are continuously exposed to AI slop and inspect it in detail to see what's wrong. That said, there are a bunch of users who don't know their left from their right and will just haphazardly throw AI accusations anywhere with no regard for what their reasoning/accusation implies.

I have seen it suggested that posts are AI because the artist spelled something wrong.
I have seen it suggested that posts are AI because the artist made two left hands.
I have seen it suggested that posts are AI because a commenter just thought the image felt wrong with no further effort to articulate what they think the issue is.

There's this weird all-consuming AI hysteria that makes people want to smear any artist who doesn't fit into their narrow definition of art, and I'm genuinely not a fan of it because weirdos will derail entire comment sections for their bullshit, and then never apologize or admit a mistake when an image's manual origin is actually proven.

If you're really curious to see every possible sign of AI slop, then:

In practice, the following things are often, but not always and not by themselves, indicators of AI slop.

  • Dimensions are weird multiples of 4, such as 576x832.
    • Pixel artists will occasionally work in certain limited resolutions such as 1028x1028, so this isn't foolproof.
  • Single-color background has a lot of random color noise.
    • Scanned traditional artwork will almost always have random noise, and an artist could manually add noise to the background if they wanted to.
  • Inconsistencies in letter shape or letter placement, unreadable signatures.
    • While the vast majority of image editors have tools for programmatically drawing text... it is perfectly normal for a human to make irregular text if they use a pencil or brush.
  • Details bleeding into each other, poorly-formed details.
    • The entire way that generative AI works is that it's trying to make sense of random noise... but also, some artists just aren't any good at drawing things like hands, sometimes they'll get background details wrong, and so on and so forth.

That said, there is no way to perfectly articulate every kind of AI-related tell. You really do just have to develop a skill for it.

One tip I heard is that if it's a PNG that looks like a JPEG. More advanced image generators MIGHT avoid this, but most I've seen do not.

lafcadio said
I have seen it suggested that posts are AI because the artist spelled something wrong.
I have seen it suggested that posts are AI because the artist made two left hands.
I have seen it suggested that posts are AI because a commenter just thought the image felt wrong with no further effort to articulate what they think the issue is.

There's this weird all-consuming AI hysteria that makes people want to smear any artist who doesn't fit into their narrow definition of art, and I'm genuinely not a fan of it because weirdos will derail entire comment sections for their bullshit, and then never apologize or admit a mistake when an image's manual origin is actually proven.

I've seen people pull a "that's AI" accusation simply because they don't like the art. Or even because they don't like the person. (Though TBF those weren't on THIS site...) "AI" is so often used as just a synonym for "crap", which really muddies the waters.

lendrimujina said:
One tip I heard is that if it's a PNG that looks like a JPEG. More advanced image generators MIGHT avoid this, but most I've seen do not.

Still an absolutely terrible heuristic.
delreason:*png*

Some things that I pick up on are a) linework that looks like it was done with a smudge tool- it's goopy and liquidy b) odd 'edges' on 'brushstrokes'- this also happens with jpg artifacts to a degree c) weird stuff- a clitoris growing in a folded knee, a hole in the coloring in the middle of a character, eyelids with pupils on them. In ai-colored sketches sometimes it dumps weird squiggles in random colors over areas it isn't sure what to do with.

snake-girl said:
for someone who was regurgitated with so much ai slop, none of these are ai-generated. they don't have common mistakes that contemporary ai generators that creates and struggles with (highly specific generic lighting, trouble with eyes, nipples, teeth, etc). artists' linework looks more organic and crispier than typical ai images i've seen.
they were also uploaded on furaffinity which vehemently against ai images, even if you scroll through their gallery and compare pieces, it doesn't seem to made with ai at all.

lafcadio said:
I'm going to preface this by saying that there is no perfect, concise, and convenient way to detect what is fully AI-generated, what is a human's interpretation of an AI-generated image, and what is fully human art.

The only truly reliable way you're going to develop this kind of skill is if you are continuously exposed to AI slop and inspect it in detail to see what's wrong. That said, there are a bunch of users who don't know their left from their right and will just haphazardly throw AI accusations anywhere with no regard for what their reasoning/accusation implies.

I have seen it suggested that posts are AI because the artist spelled something wrong.
I have seen it suggested that posts are AI because the artist made two left hands.
I have seen it suggested that posts are AI because a commenter just thought the image felt wrong with no further effort to articulate what they think the issue is.

There's this weird all-consuming AI hysteria that makes people want to smear any artist who doesn't fit into their narrow definition of art, and I'm genuinely not a fan of it because weirdos will derail entire comment sections for their bullshit, and then never apologize or admit a mistake when an image's manual origin is actually proven.

If you're really curious to see every possible sign of AI slop, then:

In practice, the following things are often, but not always and not by themselves, indicators of AI slop.

  • Dimensions are weird multiples of 4, such as 576x832.
    • Pixel artists will occasionally work in certain limited resolutions such as 1028x1028, so this isn't foolproof.
  • Single-color background has a lot of random color noise.
    • Scanned traditional artwork will almost always have random noise, and an artist could manually add noise to the background if they wanted to.
  • Inconsistencies in letter shape or letter placement, unreadable signatures.
    • While the vast majority of image editors have tools for programmatically drawing text... it is perfectly normal for a human to make irregular text if they use a pencil or brush.
  • Details bleeding into each other, poorly-formed details.
    • The entire way that generative AI works is that it's trying to make sense of random noise... but also, some artists just aren't any good at drawing things like hands, sometimes they'll get background details wrong, and so on and so forth.

That said, there is no way to perfectly articulate every kind of AI-related tell. You really do just have to develop a skill for it.

regsmutt said:
Some things that I pick up on are a) linework that looks like it was done with a smudge tool- it's goopy and liquidy b) odd 'edges' on 'brushstrokes'- this also happens with jpg artifacts to a degree c) weird stuff- a clitoris growing in a folded knee, a hole in the coloring in the middle of a character, eyelids with pupils on them. In ai-colored sketches sometimes it dumps weird squiggles in random colors over areas it isn't sure what to do with.

Thank you all for your advices
I admit that I have become a little paranoid.

They looks better than overal real drawing, but looks with same style over and over again. Mostly solo because it's more easy to get not mutated character on solo without sex scene. Also you will mostly never meet group sex scenes on AI generation because it will be too hard.

For example my AI generation of Veldora dragon that looks awesome to me and better than all existed real draws
https://e6ai.net/posts/63612

And real drawing. Looks more simple, not ideal draws in terms of sharp. But have more variety of scenes.
https://e621.net/posts?tags=veldora_tempest

Updated

lendrimujina said:
One tip I heard is that if it's a PNG that looks like a JPEG. More advanced image generators MIGHT avoid this, but most I've seen do not.

Local AI generator like WebUI ReForge by default saves into *.PNG. So it will be without JPG artifacts. It mostly limitations of online web generators. Also you can delete JPG artifacts by using AI upscaler. But again user should use local AI generator on own PC because online web generators is too limited.

So mostly that advice works versus that online web generator.
https://perchance.org/free-nsfw-ai-generator#edit

Updated

yetanotheraiuser said:
They looks better than overal real drawing, but looks with same style over and over again. Mostly solo because it's more easy to get not mutated character on solo without sex scene. Also you will mostly never meet group sex scenes on AI generation because it will be too hard.

For example my AI generation of Veldora dragon that looks awesome to me and better than all existed real draws
https://e6ai.net/posts/63612

And real drawing. Looks more simple, not ideal draws in terms of sharp. But have more variety of scenes.
https://e621.net/posts?tags=veldora_tempest

yetanotheraiuser said:
Local AI generator like WebUI ReForge by default saves into *.PNG. So it will be without JPG artifacts. It mostly limitations of online web generators. Also you can delete JPG artifacts by using AI upscaler. But again user should use local AI generator on own PC because online web generators is too limited.

So mostly that advice works versus that online web generator.
https://perchance.org/free-nsfw-ai-generator#edit

yetanotheraiuser said:
You can also train own feeling of AI detection by looking into rule34.xxx. These is a lot of such generations with various quality.

https://rule34.xxx/index.php?page=post&s=list&tags=ai_generated+anthro+

Also you can look into e621.net version for AI images
https://e6ai.net

By watching some examples you will better feel if it AI generated or not.

Thank you very much, I learned a lot from reading your answers.

I did not know that e6ai existed, I will use them as comparisons.

I have a request, Is it possible to have a Topic like:Find The Artist/Source/Image Request Thread
But for like : Is This ai ? (Or do you think its ai).

chat_sauvage said:
I have a request, Is it possible to have a Topic like:Find The Artist/Source/Image Request Thread
But for like : Is This ai ? (Or do you think its ai).

I think the staff would like to avoid having such a topic here, but you can converse over at the e6ai forums

chat_sauvage said:Thank you very much, I learned a lot from reading your answers.

I'm skilled with AI image generator and I like both arts. Real and AI generated. On AI generated I got some images that will be mostly impossible with real draws like you will found mostly nothing on tag "deep_sea_king_(one-punch_man)", but AI generated will overperform all existed real draws with easy for solo male scene. Also you can look how ridiculous it will look on IRL version since AI can try to generate it on photorealistic view.

Updated

Also I'm thinking about weird censorship deletion from east images using AI. For example I like that sexy dragon, but I don't like pixelated dick. It will be never removed anyway.
https://e621.net/posts/5051660?q=parent%3A5051659

It would be intresting can I post such non censored image here since some AI was used, but only partially to delete weird pixelated censorship.

snpthecat said:
I think the staff would like to avoid having such a topic here, but you can converse over at the e6ai forums

I'll give it a try.

Staff, I'd like to create a topic where people can share their doubts about drawings.
(Ai or not)

Artificial intelligence is advancing faster than you think, and in the long term it will be difficult to tell the differences.

chat_sauvage said:
Staff, I'd like to create a topic where people can share their doubts about drawings. (Ai or not)
(Ai or not)

When I noticed on e621.net (e6AI.net was not existed at this time) the first decent looking AI generated explicit content I was surprised and tried to generate such image myself. Also I have some weird fetishes. I like demons with such view https://e621.net/posts/2172392 but majority likes mostly cute images. So for me it's hard to found similar demonic images, but I can generate some on AI. I uploaded some of my demons, but really I generated much more of it, but never uploaded. For example https://e6ai.net/posts/57541 or https://e6ai.net/posts/51125

chat_sauvage said:
I'll give it a try.

Staff, I'd like to create a topic where people can share their doubts about drawings.
(Ai or not)

Artificial intelligence is advancing faster than you think, and in the long term it will be difficult to tell the differences.

No matter how "good" it gets, we're even better at sussing out all the flaws of AI-generated artwork, from the artifacting to the sourcing. As a concept it just doesn't work in a professional space and we've been diligent about keeping it out of spaces where it isn't welcome.

Artists and anyone who's serious about art will ensure that it won't dominate artistic spaces so I wouldn't worry too much. It's okay to be skeptical but try not to see everything in bad faith.

chat_sauvage said:
Staff, I'd like to create a topic where people can share their doubts about drawings.
(Ai or not)

There is a high chance it is going to spiral into a mess where people are going to throw in already approved posts that they don't like and making unsubstantiated claims that they are AI-generated, or a place for people to make callouts to artists that they think are using AI to make artwork.

It is best to take it to the e621 Discord helpdesk if you have trouble or doubts identifying AI-generated artwork and to clarify with staff there privately about these sort of matters.

chat_sauvage said:
Artificial intelligence is advancing faster than you think, and in the long term it will be difficult to tell the differences.

This goes both ways, Dood!
╹‿╹)~★

Just like there's a market for AI-generated biz, There is a
MASSIVE untapped market for AI-generation detection software.
Some prototypes are in the work as we speak, and the peep's
behind them have their eyes on the prize. The First person who
makes a successful AI detection software will be as big and
probably more profitable as Ad-Block. The winner of this
race would make thousands,if not a few million in publicity
and marketing opportunities, Dood.
T‿T)~★

I wish there was a nicer reason but Hey,
CEO's would sell souls to be in that great
of a position business-wise. You'd have to
mess up PRETTY HARD to 'not' make a
fortune from that, Dood.
=‿=)~★

chat_sauvage said:
Staff, I'd like to create a topic where people can share their doubts about drawings.
(Ai or not)

I would not trust most users to make these kinds of accusations.

notkastar said:
Just like there's a market for AI-generated biz, There is a MASSIVE untapped market for AI-generation detection software. Some prototypes are in the work as we speak, and the peep's behind them have their eyes on the prize. The First person who makes a successful AI detection software will be as big and probably more profitable as Ad-Block. The winner of this race would make thousands,if not a few million in publicity and marketing opportunities.

I also would not trust any vendor purporting to make definitive statements about the origin of any particular image, video, or text. It's better for sales if they return false positives instead of false negatives, and in the case of artists such as Norman Rockwell, their art is more likely to return a positive because they're part of the training data.

Updated

lafcadio said:
I also would not trust any vendor purporting to make definitive statements about the origin of any particular image, video, or text. It's better for sales if they return false positives instead of false negatives, There's an incentive for them to dupe their customers by returning false positives instead of false negatives, and in the case of artists such as Norman Rockwell, their art is more likely to return a positive because they're part of the training data.

That's the main hurdle when it comes to detection and what keeps most peeps from trying.
I'll be the first to tell ya' I don't have an answer for the logistics aside from some
Metadata/Digital Silver Bullet the program could pick up on and other Fairy Tails, Dood.
-‿-)~★

Though I'll also be the first to tell ya, even if 30% of junk slips through and 30%
are false positives, I'd still have the extension to block that remaining 40%. In
the hope that not only will updates come but also cut off as much of that generated
junk as possible. Even if it's only a little, Dood!
T‿T)~★

We found out how to block ads as big of a leap that was from the "Just deal with it" Arc.
We've just hit another obstacle that's even even EVEN more insurmountable.
We're in the next Arc and I know peeps are rearing to go to not only do it,
but to be the first, Dood!
╹‿╹)~★

Updated

qwazzy said:
If something is AI, you'll probably be able to tell just by seeing multiple comments of people saying it's AI. They're everywhere, people have a major hard-on for detecting and condemning AI art.

If an artist you like is actually making art with AI and not themselves, they'll be found out before long, you won't even have to lift a finger. Just read the feedback.

I have, unfortunately, seen a LOT of "OGM that's AI because it's too good!" aimed at photos taken by professional photographers, provably so since they were first published like a decade or more ago before ANY of these image generators were even close to almost becoming a thing.

People have a hard-on for THINKING they're detecting and condemning AI art, just as they have for a multitude of other white-knight-esque behaviors. It doesn't matter if they're right; they want the dopamine from FEELING that they're right.

The third picture I can tell you pretty definitively isn't AI because the artist already has a piece on here from 2020, which is before AI was capable of producing anything that good. (Ironic that their username literally has "aiart" right in it though.) I've taken the liberty of uploading their Loona picture and one other I liked.

Also, I'm not sure you have to worry too much about people using AI to pose as real artists on sites that either disallow it (like FA) or require it to be tagged. In my experience AI "artists" tend to be pretty open about it, either by staying where they're welcome and following the rules or by openly invading AI-hostile spaces to smugly show off how adeptly their expensive machines are able to replicate and surpass the work of us puny humans. Either that or they're trying to scam people out of commission money, which we know is a thing because they tend to get caught when they do that.

If nothing else you can just stick to only uploading stuff from FA and then blame their admins for not doing their job if any of it turns out to be AI later.

Okay, so, on the note of detection programmes, no, just no. You're trying to tell a machine to detect itself which is the entire way in which these things function in the first place.
Assuming they are of the same quality as written detection, you're going to be detecting literally every image as A.I. anyway.

Also taking notes for if I want something to have a deliberate simulated AI look for stylistic reasons; might have a cool effect if I want to draw a canonically-digital creature like MissingNo. or SCP-1471-A, making them look artificial and out-of-place without resorting to typical tropes like inverted color blocks and chromatic aberration.

Another thing about image resolutions: Some art programs (like Krita) have texture templates, some indecisive artists may use them for easy starting canvas sizes, they're all powers of 2. I used to use those but nowadays I just crop afterwards or use paper sizes to start with. Resolution isn't a good sole indicator, also look for the types of mistakes made in conjunction with each other; for example, an artist super skilled in rendering and anatomy probably won't make weirdly splattered fluids on a pussy.

Nobody's suggested just opening the .PNG itself to look for the AI rendering metadata?

In my experience when people are trying to purposely mislead their AI renders as 'art' they're usually too stupid to remove the metadata.

pocket_erector said:
Nobody's suggested just opening the .PNG itself to look for the AI rendering metadata?

In my experience when people are trying to purposely mislead their AI renders as 'art' they're usually too stupid to remove the metadata.

Who tried to fool you that it his real work also will try to manually fix some AI render errors that definetly points into AI mistakes. Any resave on image editor will delete AI prompt metadata.

alphamule

Privileged

yetanotheraiuser said:
Who tried to fool you that it his real work also will try to manually fix some AI render errors that definetly points into AI mistakes. Any resave on image editor will delete AI prompt metadata.

This is an evil feature of MSPaint that caught me once. Accidentally overwrote a PNG that had rendering code in it. Yeah, got too clever and forgot. Oddly, other editors saved it just fine.

donovan_dmc said:
No.

While I certainly understand why this specific proposal would not, in a million years, be workable: Do we currently have a specific avenue to go through for strongly suspected AI gens on the sight? Or at the very least, do we have a hard rule on how moderation/janitorial staff determine what is and isn't AI here? Because I hate to say it, but gens are getting more convincing, and/or the people who make them are getting better at hiding it. So I feel like we may be reaching a point where certain AI gens might only be detectable by looking back through an "artist's" catalogue of work rather than trying to determine on the merit of individual uploads.

None of this is to say that the staff have been doing poorly in handling it thus far. It's just that are definitely some artists who are straddling that line between AI gens and real art that get in, and that line seems quite blurry right now. Basically just looking for a little insight into how that's processed, and how questions about potential missed uploads might be processed.

Updated

felicity_longis said:
Do we currently have a specific avenue to go through for strongly suspected AI gens on the sight? Or at the very least, do we have a hard rule on how moderation/janitorial staff determine what is and isn't AI here?

AI-generated artwork would never be accepted in the first place, so any unapproved uploads are screened through by the janitorial team (and not by regular members).
I would assume that no janitor would approve a post unless they are completely certain it is not AI-generated. Any grey area cases would have gone through multiple janitors.

If it slips through the janitors or is uploaded by someone with unrestricted uploads (both unlikely scenarios but still possible), then the only way it gets checked again is if someone with Priv+ account flags the upload for not meeting the uploading guidelines (a flag option that is otherwise not accessible for regular members).

For regular members who still want to voice their doubt on suspected AI-generated artworks, they can always approach the helpdesk on the e6 Discord server.

thegreatwolfgang said:
unless they are completely certain it is not AI-generated. Any grey area cases would have gone through multiple janitors.

At least I don't know why you want to fool peoples here on non AI site when you can upload your work on separate site E6AI instead. Same site as E621, but for AI images. If your AI generated work on good quality, you will meet zero problems on approval stage. And you don't need to spend time and energy trying to fool peoples over and over again on e621 and can concentrate making good AI images instead.

thegreatwolfgang said:
AI-generated artwork would never be accepted in the first place, so any unapproved uploads are screened through by the janitorial team (and not by regular members).
I would assume that no janitor would approve a post unless they are completely certain it is not AI-generated. Any grey area cases would have gone through multiple janitors.

If it slips through the janitors or is uploaded by someone with unrestricted uploads (both unlikely scenarios but still possible), then the only way it gets checked again is if someone with Priv+ account flags the upload for not meeting the uploading guidelines (a flag option that is otherwise not accessible for regular members).

For regular members who still want to voice their doubt on suspected AI-generated artworks, they can always approach the helpdesk on the e6 Discord server.

Right, I get that the Janitorial staff looks for AI generated content. That's their job. But what's the line we look for here? Because we see plenty of AI gens being deleted. But then we have a lot of posts with the ai_assisted tag that, aside from having the signature of and/or were uploaded by an established artist, have no indication of being anything but AI generated:
post #4893554 post #4964792
So is it just that after a certain point the artist just gets the benefit of the doubt? Are we looking for certain artistic elements being added? What is the line between "assisted" and "generated", and what's being looked for to establish that? Because based on the wiki post, it seems that any AI gen that I take the time bring into photoshop, adjust the saturation by any amount, and slap a signature on would be acceptable. Which definitely doesn't seem to be the case.

yetanotheraiuser said:
At least I don't know why you want to fool peoples here on non AI site when you can upload your work on separate site E6AI instead. Same site as E621, but for AI images. If your AI generated work on good quality, you will meet zero problems on approval stage. And you don't need to spend time and energy trying to fool peoples over and over again on e621 and can concentrate making good AI images instead.

Because a large number of people who generate AI imagery are trying to pass it off as real artwork. Which, for the context of E621, is less a problem of "X individual keep uploading their slop", and more an issue of "Various users keep uploading images from X individual's twitter (or other similar ai-permissive platform) without realizing what it is". Which plays into the whole problem of the vast majority of users here (and people across the internet in general) being broadly unable to differentiate art from ai gens. Let alone being able to do so accurately and consistently. People who upload their own AI gens here tend to either get nuked pretty quickly, are very new to the site and its guidelines, or some combination of both.

You're free to believe that there is a useful and non-exploitative application for AI. But the fact of the matter is that AI is a tool viewed by many (including those who create these products) for making money; the "art" is wholly incidental to that goal. And it's a lot easier to sell fake "art" that looks like it took effort to people who don't know better (or, to an extent, don't care) than it is to sell real AI gens right now.

Updated

felicity_longis said:
Right, I get that the Janitorial staff looks for AI generated content. That's their job. But what's the line we look for here? Because we see plenty of AI gens being deleted. But then we have a lot of posts with the ai_assisted tag that, aside from having the signature of and/or were uploaded by an established artist, have no indication of being anything but AI generated:
post #4893554 post #4964792
So is it just that after a certain point the artist just gets the benefit of the doubt? Are we looking for certain artistic elements being added? What is the line between "assisted" and "generated", and what's being looked for to establish that? Because based on the wiki post, it seems that any AI gen that I take the time bring into photoshop, adjust the saturation by any amount, and slap a signature on would be acceptable. Which definitely doesn't seem to be the case.

Painting over ai junk is allowed. Those examples (and others that I've seen get approved) aren't that. They're likely a real sketch or lineart (I've seen sketches from Twinkle-sez of pieces where the finished product is definitely mostly ai) that has been rendered by ai and then adjusted/edited/partially painted over. It's why they're filled with ai artifacts but are able to avoid certain anatomy errors ai is prone to.

I think that there needs to be some clear and consistent decisions made on 'ai assisted' works that go beyond backgrounds. This particular process is definitely going to get more and more difficult to spot.

felicity_longis said:
Right, I get that the Janitorial staff looks for AI generated content. That's their job. But what's the line we look for here? Because we see plenty of AI gens being deleted. But then we have a lot of posts with the ai_assisted tag that, aside from having the signature of and/or were uploaded by an established artist, have no indication of being anything but AI generated:
post #4893554 post #4964792
So is it just that after a certain point the artist just gets the benefit of the doubt? Are we looking for certain artistic elements being added? What is the line between "assisted" and "generated", and what's being looked for to establish that? Because based on the wiki post, it seems that any AI gen that I take the time bring into photoshop, adjust the saturation by any amount, and slap a signature on would be acceptable. Which definitely doesn't seem to be the case.

You will come to realise that whenever you ask questions such as "where is the line drawn here?" to the mod team, you will never get a clearcut answer.
And that is intentional, since they want to be flexible with the rules while also not wanting to set arbitrary limits to what is and is not allowed.
After all, nobody wants to sit on their computers all day calculating the exact percentage of AI-assistance used in an artwork when they are already volunteering in their spare time going through thousands of unapproved posts daily.

In addition, just to clarify, al-assisted is for posts that are human-made with minimal use of AI.
More specifically, according to the Uploading Guidelines on AI-generated content:

  • No AI generated, or AI assisted artwork. Exceptions are currently for backgrounds (treated like using a photo as a background, quality rules apply); for artwork that references, but does not directly use, AI generated content; for full paintovers; and for audio in WebM posts.

felicity_longis said: stuff

I already have a separate forum thread about this but I think that letting people just freely tag ai_assisted and ai_generated with no further justification is going to be more trouble than it's worth. We already have users spreading their false accusations to Twitter and Discord when janitors like me make judgments they don't like, and any serious accusation should at least try to point us the right way since there are 33 post approvers who're all going to be looking for completely different things.

I'm surprised we haven't been subject to a Twitter thread where some innocent artist goes "wtf e621 tagged my art as ai_generated! Slander!"

felicity_longis said:
post #4893554

Looking into this post and some of the artist's more recent works.

felicity_longis said:
post #4964792

Gonna need something more concrete than a handful of tag edits. What's the angle here? What sets this apart from the artist's 2022 works?

lafcadio said:
Gonna need something more concrete than a handful of tag edits. What's the angle here? What sets this apart from the artist's 2022 works?

2022:
post #3998736

I think the biggest difference is in the lineart and sketch lines. In the 2022 piece there's areas of sketchy hatching- around the armpit, the lower jaw, and on the genitals. These lines are crisp, distinct, and the coloring under them is consistent with the coloring throughout the rest of the image.

In the strip-club piece these areas of hatching (in the armpit, on the breast, around the chin, and where the ankle is near the thigh) have ai artifacts in them. The lines are blurry and warped- the warping is particularly visible on the thigh and near the back. The coloring around these lines is unusual and has an effect somewhere between chromatic aberration and a 'plastic wrap' filter. There's also a part of the tail that is melting into the floor.

lafcadio said:
Gonna need something more concrete than a handful of tag edits. What's the angle here? What sets this apart from the artist's 2022 works?

Well besides all this:

regsmutt said:
2022:
post #3998736

I think the biggest difference is in the lineart and sketch lines. In the 2022 piece there's areas of sketchy hatching- around the armpit, the lower jaw, and on the genitals. These lines are crisp, distinct, and the coloring under them is consistent with the coloring throughout the rest of the image.

In the strip-club piece these areas of hatching (in the armpit, on the breast, around the chin, and where the ankle is near the thigh) have ai artifacts in them. The lines are blurry and warped- the warping is particularly visible on the thigh and near the back. The coloring around these lines is unusual and has an effect somewhere between chromatic aberration and a 'plastic wrap' filter. There's also a part of the tail that is melting into the floor.

You also have clear artifacting on the deer's teeth and eyes, a random dimple past her jaw, color bleeding through lineword in several places when the artist either blends without lines or uses lines as clean contours between coloration...

I really like Twinkle-sez's work, and have for over 15 years now, which makes it really quite jarring to see this. That's not a "oh boo-hoo, why do they use AI now?" statement; it's to make it clear that this is pretty obviously not comparable to their past artwork. Like just as an example, the mouse's glasses: in essentially all previous work, glasses on anthro characters have either had temples which follow beneath the base of the ear and terminate beneath hair or at the sides of the head, or simply lack temples. Now fair enough, a stylistic change like this isn't an issue on its own, but between "The gen didn't know how to render glasses" and "TS just made this total 180 on a consistent stylistic pattern you can trace back here all the way to 2009"...

I get that it's a bunch of little nitpicky things, but it's a bunch of things that cumulatively point to either use of AI with no real sign of significant human input, or that Twinkle-Sez needs an MRI as soon as possible. I'm not a doctor, but I've seen a lot of slop, so I know where I'm putting my money. To be frank, I have to feel that if this image didn't have a known artist's name attached to it, it would've been removed two months ago.

TBH that deer piece is not the best example. Take for example this one:

post #4805358
This looks to me straight up just lightly touched up AI content. The cow's hair is melting into her back, her calves have really weird stuff in the line-art, the shapes of the fallen papers make no sense, the white cat in bg has really messed up feet, the character in foreground has something really weird going on with hands. I am confused about why this is allowed, when it's so clearly primarily AI that has just been edited a bit.

rupikonna said:
TBH that deer piece is not the best example. Take for example this one:

post #4805358
This looks to me straight up just lightly touched up AI content. The cow's hair is melting into her back, her calves have really weird stuff in the line-art, the shapes of the fallen papers make no sense, the white cat in bg has really messed up feet, the character in foreground has something really weird going on with hands. I am confused about why this is allowed, when it's so clearly primarily AI that has just been edited a bit.

It certainly doesn't help that they have a bunch of these, and some are actually passable. I'm definitely not proposing anything like "Anything X artist posts that even smells of AI should be deleted". But clearly, in this case at least, there's a scale here of how much of a fuck the artist seems to have given about either producing or touching up the gen. Like this one for example:
post #4805151
Sure, the background is weird and all that, but there's not a ton going on that isn't like "Eh, maybe they just weren't feeling it that day". Mistakes that, outside of potential AI use, wouldn't make you question whether or not the artist suffered a severe blow to the head in the past few weeks. But then you have the above two pieces which (again), had they not come directly from this artist, wouldn't for a second be questioned as AI knock-offs trained on their work before being promptly deleted.

When it's something that's so obvious like this, I really don't think the mere presence of a name attached to it should count for anything. Do we know this artist can draw? Yes. Do we know they can produce a workflow that could offer a strong basis for an AI gen but still be called (in some sense) their original work? Sure. But without seeing that, can we really just assume that's what's happening. Because at that point, it should be basically open season for AI posts by any artist tag that's been around for more than two or three years. That's hyperbolic, given that I can only point to this and one other artist here who would fit that, but it serves the point: Does having an existing artist tag alone mean we give them the benefit of the doubt on posts that are so blatantly against the uploading guidelines here?

Updated

rupikonna said:
TBH that deer piece is not the best example. Take for example this one:

post #4805358
This looks to me straight up just lightly touched up AI content. The cow's hair is melting into her back, her calves have really weird stuff in the line-art, the shapes of the fallen papers make no sense, the white cat in bg has really messed up feet, the character in foreground has something really weird going on with hands. I am confused about why this is allowed, when it's so clearly primarily AI that has just been edited a bit.

Yeah I didn't want to bring up Skirt-Ear Ponytail-Back but that one is. Pure slop.

rupikonna said:
TBH that deer piece is not the best example. Take for example this one:

post #4805358
This looks to me straight up just lightly touched up AI content. The cow's hair is melting into her back, her calves have really weird stuff in the line-art, the shapes of the fallen papers make no sense, the white cat in bg has really messed up feet, the character in foreground has something really weird going on with hands. I am confused about why this is allowed, when it's so clearly primarily AI that has just been edited a bit.

also the cat in the background is leaning on a chair arm that isn't attached to the chair and appears to not be wearing pants. add to that that, all three of the chairs in the scene are totally different designs, which is very odd.

EDIT:

felicity_longis said:
post #4805151

this one also looks really weird/bad when you take the furniture into account, all of the stools have different numbers of legs, and colors, and sizes, and shapes. I mean, that one, second from the left, has two different numbers of legs depending on how far down the stool you go.

felicity_longis said:
post #4964792

lafcadio said:
Gonna need something more concrete than a handful of tag edits. What's the angle here? What sets this apart from the artist's 2022 works?

the dude has glasses with subcutaneous temples.

Updated

felicity_longis said:
Well as much as I hate to necro this thread, it's probably relevant to note that Twinkle-sez has just come out and admitted to using significant ai assistance on nearly all of their work since early 2023., and are now realizing that they're now quite upset with the results. So...

Well that explains it. And my assessment that of touched up was somewhat correct here. In light of this, it would be good idea to go through the artwork, considering how much AI was used for the art.

rupikonna said:
Well that explains it. And my assessment that of touched up was somewhat correct here. In light of this, it would be good idea to go through the artwork, considering how much AI was used for the art.

Frankly anything post March 2023 should just be deleted.

regsmutt said:
Frankly anything post March 2023 should just be deleted.

Given that they basically open by asking people to delete applicable submissions on FA in exchange for properly finished work, that would probably be the best option.

all handled

dba_afish said:
also the cat in the background is leaning on a chair arm that isn't attached to the chair and appears to not be wearing pants. add to that that, all three of the chairs in the scene are totally different designs, which is very odd.

EDIT:
this one also looks really weird/bad when you take the furniture into account, all of the stools have different numbers of legs, and colors, and sizes, and shapes. I mean, that one, second from the left, has two different numbers of legs depending on how far down the stool you go.

the dude has glasses with subcutaneous temples.

Just pulling this post back up to mention that they admitted to AI coloring using Clip Studio Paint, not AI composition/linework using DALL-E/Stable Diffusion/Bing/etc., so the claim seems to be that they did, in fact, draw these chairs/furniture/glasses.
The posts are deleted either way, but people will need different evidence (such as the same coloring style) if they want to repeat these accusations in the future.

Watsit

Privileged

lafcadio said:
all handled

Just pulling this post back up to mention that they admitted to AI coloring using Clip Studio Paint, not AI composition/linework using DALL-E/Stable Diffusion/Bing/etc., so the claim seems to be that they did, in fact, draw these chairs/furniture/glasses.
The posts are deleted either way, but people will need different evidence (such as the same coloring style) if they want to repeat these accusations in the future.

It was a bit more than just coloring. They admitted to having had used AI coloring since March 2023, but also toyed with AI in general during that time too:

Since it is generally innate for an artist to experiment and seek out new ways, when such tools became widely available, and not as clumsy as before (...) I couldn't resist trying to somehow use it in my workflow.

At first I was amazed at how accurately the machine saw the flaws in my technique and tried to correct them. So I couldn't think of anything better than to keep running almost finished drawings through it for, as it seemed to me then, improved shading and additional detailing. [...]

Well, and so it was, until one day I came across a certain drawing from an earlier period of these experiments, looked at the details more closely without the usual haste, and was horrified by what I had done. The “enhanced details and shading” I was after turned out to be digital noise artifacts crudely mixed with haphazardly cut out fragments of someone else's drawings. [...]

The worst thing is that this approach makes the drawings look so unnatural that it downplays the artist's actual work to such an extent that it becomes almost impossible for the viewer to distinguish how much of it was actually put into the piece (some passerby might look at such a picture and decide that it was not drawn by a person, but by a command sent to the computer program, and even decide that your humble servant is a scam).

Image-to-image AI can and does change various aspects if you're having it add details, creating some of those AI-isms where it's trying to add details. He even admits he wasn't looking very thoroughly at what the AI did to his art, and when he finally did, he was horrified with what he saw. That's how it tends to go with AI image generation; at a first glance, it can look amazing, almost flawless. Keep looking at it, and you'll find subtle errors, keeping looking, and you'll find more until you can't unsee it and wonder how you missed it all. So while he may have initially drawn the chairs, glasses, etc, the AI likely still messed with them to create those effects, rather than being what he drew.

Updated

lafcadio said:
all handled

Just pulling this post back up to mention that they admitted to AI coloring using Clip Studio Paint, not AI composition/linework using DALL-E/Stable Diffusion/Bing/etc., so the claim seems to be that they did, in fact, draw these chairs/furniture/glasses.
The posts are deleted either way, but people will need different evidence (such as the same coloring style) if they want to repeat these accusations in the future.

They actually don't say specifically what they did. Clip Stufio Paint is mentioned as an example of an early AI tool that delivered bad results:

Since it is generally innate for an artist to experiment and seek out new ways, when such tools became widely available, and not as clumsy as before (ordinary users are unlikely to know this, but for example everyone’s favorite Clip Studio Paint had that “colorize all” AI autocoloring tool way long before stuff like Stable Diffusion even emerged, though I doubt anyone actually used it because it had certain problems related to maths and machine learning, and hence was returning very dull colors.

lafcadio said:
all handled

Twinkle essentially points to all their art from March 2023 onward as being tainted by this problem, but a lot of it is still here, including images with the same obvious ai artifacts from whichever processes they applied. Should we be reporting all of these individually, making a report, or...?

felicity_longis said:
Twinkle essentially points to all their art from March 2023 onward as being tainted by this problem, but a lot of it is still here, including images with the same obvious ai artifacts from whichever processes they applied. Should we be reporting all of these individually, making a report, or...?

Examples? The stuff that's still up that was uploaded post March 2023 was made pre-2023, unless something was mislabeled.

felicity_longis said:
Twinkle essentially points to all their art from March 2023 onward as being tainted by this problem, but a lot of it is still here, including images with the same obvious ai artifacts from whichever processes they applied. Should we be reporting all of these individually, making a report, or...?

Posts with no year tagged: twinkle-sez date:>2023-03-01 -200* -201* -202*
Posts originating from 2023/2024: twinkle-sez date:>2023-03-01 ~2023 ~2024
Posts originating from before 2023: twinkle-sez date:>2023-03-01 ~200* ~201* ~2020 ~2021 ~2022

regsmutt said:
Examples? The stuff that's still up that was uploaded post March 2023 was made pre-2023, unless something was mislabeled.

lafcadio said:
Posts with no year tagged: twinkle-sez date:>2023-03-01 -200* -201* -202*
Posts originating from 2023/2024: twinkle-sez date:>2023-03-01 ~2023 ~2024
Posts originating from before 2023: twinkle-sez date:>2023-03-01 ~200* ~201* ~2020 ~2021 ~2022

post #3998736 has a submission date at source of April 17 2023, and post #3996576 was April 16. Fair enough, that's not exactly "a lot", so poor wording on my part.

I've been reviewing some posts which have had flags or ai_assisted or ai_generated tags removed from them. Some have evidence of AI use, some do not. Check out my recent comments and post changes if you want a bit of insight about what I would look for, as someone making AI-assisted pics and having to run damage control for AI quirks.

Notably, I do NOT consider lighting or shading patterns to be reliable. This is what people usually look to, but it leads to many false positives and false negatives. Especially when you get people using AI as pose references, seeing how shadows and highlights would fall at a particular angle. I know a fair number of artists do that who otherwise have zero AI use.

Ultimately, there is no way to tell when a sufficiently skilled artist uses AI. The tech is there already. There are a number of high-profile artists I suspect to be doing paintovers of AI pics because the poses, proportions, etc change subtly at some distinct point post-2022, becoming more in line with poses/things easy to do with AI. But I have no hard proof (except for a few that are super obvious once pointed out) so it's never worth it to falsely accuse. Plus if they use AI at a specific step, ie converting a sketch into lineart and doing nothing more, how are you going to know?

  • 1