sillyysunnyy - Tumblr blog

sillyysunnyy · 9 months

Text

Recognizing AI Generated Images, Danmei Edition

Heyo, @unforth here! I run some danmei art blogs (@mdzsartreblogs, @tgcfartreblogs, @svsssartreblogs, @zhenhunartreblogs, @erhaartreblogs, @dmbjartreblogs, @tykartreblogs, and @cnovelartreblogs) which means I see a LOT of danmei art, and I go through the main fandom tags more-or-less every day.

Today, for the first time, I spotted someone posting AI-generated images (I refuse to call them AI "art" - and to be clear, that's correct of me, because at least in the US it literally LEGALLY isn't art) without any label indicating they were AI generated. I am not necessarily against the existence of AI-generated images (though really...considering all the legal issues and the risks of misuse, I'm basically against them); I think they potentially have uses in certain contexts (such as for making references) and I also think that regardless of our opinions, we're stuck with them, but they're also clearly not art and I don't reblog them to the art side blogs.

The images I spotted today had multiple "tells," but they were still accumulating notes, and I thought it might be a good moment to step back and point out some of the more obvious tells because my sense is that a LOT of people are against AI-generated images being treated as art, and that these people wouldn't want to support an AI-generator user who tried to foist off their work as actual artwork, but that people don't actually necessarily know how to IDENTIFY those works and therefore can inadvertently reblog works that they'd never support if they were correctly identified. (Similar to how the person who reposts and says "credit to the artist" is an asshole but they're not the same as someone who reposts without any credit at all and goes out of their way to make it look like they ARE the artist when they're not).

Toward that end, I've downloaded all the images I spotted on this person's account and I'm going to use them to highlight the things that led me to think they were AI art - they posted a total of 5 images to a few major danmei tags the last couple days, and several other images not to specific fandoms (I examined 8 images total). The first couple I was suspicious, but it wasn't til this morning that I spotted one so obvious that it couldn't be anything BUT AI art. I am NOT going to name the person who did this. The purpose of this post is purely educational. I have no interest whatsoever in bullying one rando. Please don't try to identify them; who they are is genuinely irrelevant, what matters is learning how to recognize AI art in general and not spreading it around, just like the goal of education about reposting is to help make sure that people who repost don't get notes on their theft, to help people recognize the signs so that the incentive to be dishonest about this stuff is removed.

But first: Why is treating AI-generated images as art bad?

I'm no expert and this won't be exhaustive, but I do think it's important to first discuss why this matters.

On the surface, it's PERHAPS harmless for someone to post AI-generated images provided that the image is clearly labeled as AI-generated. I say "perhaps" because in the end, as far as I'm aware, there isn't a single AI-generation engine that's built on legally-sourced artwork. Every AI (again, to the best of my knowledge) has been trained using copyrighted images usually without the permission of the artists. Indeed, this is the source of multiple current lawsuits. (and another)

But putting that aside (as if it can be put aside that AI image generators are literally unethically built), it's still problematic to support the images being treated as art. Artists spend thousands of hours learning their craft, honing it, sharing their creations, building their audiences. This is what they sell when they offer commissions, prints, etc. This can never be replicated by a computer, and to treat an AI-generated image as in any way equivalent is honestly rude, inappropriate, disgusting imo. This isn't "harmless"; supporting AI image creation engines is damaging to real people and their actual livelihoods. Like, the images might be beautiful, but they're not art. I'm honestly dreading someone managing to convince fandom that their AI-generated works are actual art, and then cashing in on commissions, prints, etc., because people can't be fussed to learn the difference. We really can't let this happen, guys. Fanartists are one of the most vibrant, important, prominent groups in all our fandoms, and we have to support them and do our part to protect them.

As if those two points aren't enough, there's already growing evidence that AI-generated works are being used to further propagandists. There are false images circulating of violence at protests, deep-fakes of various kinds that are helping the worst elements of society to push their horrid agendas. As long as that's a facet of AI-generated works, they'll always be dangerous.

I could go on, but really this isn't the main point of my post and I don't want to get bogged down. Other people have said more eloquently than I why AI-generated images are bad. Read those. (I tried to find a good one to link but sadly failed; if anyone knows a good post, feel free to send it and I'll add the link to the post).

Basically: I think a legally trained AI-image generator that had built-in clear watermarks could be a fun toy for people who want reference images or just to play with making pseudo-art. But...that's not what we have, and what we do have is built on theft and supports dystopia so, uh. Yeah fuck AI-generated images.

How to recognize AI-Generated Images Made in an Eastern Danmei Art Style

NOTE: I LEARNED ALL THE BASIC ON SPOTTING AI-GENERATED IMAGES FROM THIS POST. I'll own I still kinda had the wool over my eyes until I read that post - I knew AI stuff was out there but I hadn't really looked closely enough to have my eyes open for specific signs. Reading that entire post taught me a lot, and what I learned is the foundation of this post.

This post shouldn't be treated as a universal guide. I'm specifically looking at the tells on the kind of art that people in danmei fandoms often see coming from Weibo and other Chinese, Japanese, and Korean platforms, works made by real artists. For example, the work of Foxking (狐狸大王a), kokirapsd, and Changyang (who is an official artist for MDZS, TGCF, and other danmei works). This work shares a smooth use of color, an aim toward a certain flavor of realism, an ethereal quality to the lighting, and many other features. (Disclaimer: I am not an artist. Putting things in arty terms is really not my forte. Sorry.)

So, that's what these AI-generated images are emulating. And on the surface, they look good! Like...

...that's uncontestably a pretty picture (the white box is covering the "artist's" watermark.) And on a glance, it doesn't necessarily scream "AI generated"! But the devil is in the details, and the details are what this post is about. And that picture? Is definitely AI generated.

This post is based on 8 works I grabbed from a single person's account, all posted as their own work and watermarked as such. Some of the things that are giveaways only really show when looking at multiple pieces. I'm gonna start with those, and then I'll highlight some of the specifics I spotted that caused me to go from "suspicious" to "oh yeah no these are definitely not art."

Sign 1: all the images are the exact same size. I mean, to the pixel: 512 x 682 pixels (or 682 x 512, depending on landscape or portrait orientation). This makes zero sense. Why would an artist trim all their pieces to that size? It's not the ideal Tumblr display size (that's 500 x 750 pixels). If you check any actual artist's page and look at the full-size of several of their images, they'll all be different sizes as they trimmed, refined, and otherwise targeted around their original canvas size to get the results they wanted.

Sign 2: pixelated. At the shrunken size displayed on, say, a mobile Tumblr feed, the image looks fine, but even just opening the full size upload, the whole thing is pixelated. Now, this is probably the least useful sign; a lot of artists reduce the resolution/dpi/etc. on their uploaded works so that people don't steal them. But, taken in conjunction with everything else, it's definitely a sign.

Those are the two most obvious overall things - the things I didn't notice until I looked at all the uploads. The specifics are really what tells, though. Which leads to...

Sign 3: the overall work appears to have a very high degree of polish, as if it were made by an artist who really really knows what they're doing, but on inspection - sometimes even on really, REALLY cursory inspect - the details make zero sense and reflect the kinds of mistakes that a real artist would never make.

So, here's the image that I saw that "gave it away" to me, and caused me to re-examine the images that had first struck me as off but that I hadn't been able to immediately put my finger on the problem. I've circled some of the spots that are flagrant.

Do you see yet? Yes? Awesome, you're getting it. No? Okay, let's go point by point, with close ups.

Sign 4: HANDS. Hands are currently AI's biggest weakness, though they've been getting better quickly and honestly that's terrifying. But whatever AI generated this picture clearly doesn't get hands yet, because that hand is truly an eldritch horror. Look at this thing:

It has two palms. It has seven fingers. It's basically two hands overlaid over each other, except one of those hands only has four fingers and the other has three. Seeing this hand was how I went from "umm...maybe they're fake? Maybe they're not???" to "oh god why is ANYONE reblogging this when it's this obvious?" WATCH THE HANDS. (Go back up to that first one posted and look at the hand, you'll see. Or just look right below at this crop.) Here's some other hands:

Sign 5: Hair and shadows. Once I started inspecting these images, the shadows of the hair on the face was one of the things that was most consistently fucked up across all the uploaded pictures. Take a look:

There's shadows of tendrils on the forehead, but there's no corresponding hair that could possibly have made those shadows. Likewise there's a whole bunch of shadows on the cheeks. Where are those coming from? There's no possible source in the rest of the image. Here's some other hair with unrelated wonky shadows:

Sign 6: Decorative motifs that are really just meaningless squiggles. Like, artists, especially those who make fanart, put actual thought into what the small motifs are on their works. Like, in TGCF, an artist will often use a butterfly motif or a flower petal motif to reflect things about the characters. An AI, though, can only approximate a pattern and it can't imbue those with meanings. So you end up with this:

What is that? It's nothing, that's what. It's a bunch of squiggles. Here's some other meaningless squiggle motifs (and a more zoomed-in version of the one just above):

Sign 7: closely related to meaningless squiggle motifs is motifs that DO look like something, but aren't followed through in any way that makes sense. For example, an outer garment where the motifs on the left and the right shoulder/chest are completely different, or a piece of cloth that's supposed to be all one piece but that that has different patterns on different sections of it. Both of these happen in the example piece, see?

The first images on the top left is the left and right shoulder side by side. The right side has a scalloped edge; the left doesn't. Likewise, in the right top picture, you can see the two under-robe lapels; one has a gold decoration and the other doesn't. And then the third/bottom image shows three sections of the veil. One (on the left) has that kind of blue arcy decoration, which doesn't follow the folds of the cloth very well and looks weird and appears at one point to be OVER the hair instead of behind it. The second, on top of the bottom images, shows a similar motif, except now it's gold, and it looks more like a hair decoration than like part of the veil. The third is also part of the same veil but it has no decorations at all. Nothing about this makes any sense whatsoever. Why would any artist intentionally do it that way? Or, more specifically, why would any artist who has this apparent level of technical skill ever make a mistake like this?

They wouldn't.

Some more nonsensical patterns, bad mirrors, etc. (I often put left/right shoulders side by side so that it'd be clearer, sorry if it's weird):

Sign 8: bizarre architecture, weird furniture, etc. Most of the images I'm examining for this post have only partial backgrounds, so it's hard to really focus on this, but it's something that the post I linked (this one) talks about a lot. So, like, an artist will put actual thought into how their construction works, but an AI won't because an AI can't. There's no background in my main example image, but take a look at this from another of my images:

On a glance it's beautiful. On a few seconds actually staring it's just fucking bizarre. The part of the ceiling on the right appears to be domed maybe? But then there's a hard angle, then another. The windows on the right have lots of panes, but then the one on the middle-left is just a single panel, and the ones on the far left have a complete different pane model. Meanwhile, also on the left side at the middle, there's that dark gray...something...with an arch that mimics the background arches except it goes no where, connects to nothing, and has no apparent relationship to anything else going on architecturally. And, while the ceiling curves, the back wall is straight AND shows more arches in the background even though the ceiling looks to end. And yes, some of this is possible architecture, but taken as a whole, it's just gibberish. Why would anyone who paints THAT WELL paint a building to look like THAT? They wouldn't. It's nonsense. It's the art equivalent of word salad. When we look at a sentence and it's like "dog makes a rhythmical salad to betray on the frame time plot" it almost resembles something that might mean something but it's clearly nonsense. This background is that sentence, as art.

Sign 9: all kinds of little things that make zero sense. In the example image, I circled where a section of the hair goes BELOW the inner robe. That's not impossible but it just makes zero sense. As with many of these, it's the kind of thing that taken alone, I'd probably just think "well, that was A Choice," but combined with all the other weird things it stands out as another sign that something here is really, really off. Here's a collection of similar "wtf?" moments I spotted across the images I looked at (I'm worried I'm gonna hit the Tumblr image cap, hence throwing these all in one, lol.)

You have to remember that an actual artist will do things for a reason. And we, as viewers, are so used to viewing art with that in mind that we often fill in reasons even when there aren't. Like, in the image just about this, I said, "what the heck are these flowers growing on?" And honestly, I COULD come up with explanations. But that doesn't mean it actually makes sense, and there's no REASON for it whatsoever. The theoretical same flowers are, in a different shot, growing unsupported! So...what gives??? The answer is nothing gives. Because these pieces are nothing. The AI has no reason, it's just tossing in random aesthetic pieces together in a mishmash, and the person who generated them is just re-generating and refining until they get something that looks "close enough" to what they wanted. It never was supposed to make sense, so of course it doesn't.

In conclusion...

After years of effort, artists have gotten across to most of fandom that reposts are bad, and helped us learn strategies for helping us recognize reposts, and given us an idea of what to do when we find one.

Fandom is just at the beginning of this process as it applies to AI-generated images. There's a LOT of education that has to be done - about why AI-generated images are bad (the unethical training using copyrighted images without permission is, imo, critical to understanding this), and about how to spot them, and then finally about what to do when you DO find them.

With reposts, we know "tell original artist, DCMA takedowns, etc." That's not the same with these AI-images. There's no original owner. There's no owner at all - in the US, at least, they literally cannot be copyrighted. Which is why I'm not even worrying about "credit" on this post - there's nothing stolen, cause there's nothing made. So what should you do?

Nothing. The answer is, just as the creator has essentially done nothing, you should also do nothing. Don't engage. Don't reblog. Don't commission the creator or buy their art prints. If they do it persistently and it bothers you, block them. If you see one you really like, and decide to reblog it, fine, go for it, but mark it clearly - put in the ACTUAL COMMENTS (not just in the tags!) that it's AI art, and that you thought it was pretty anyway. But honestly, it'd be better to not engage, especially since as this grows it's inevitable that some actual artists are going to start getting accused of posting AI-generated images by over-zealous people. Everyone who gets a shadow wrong isn't posting AI-generated images. A lot of these details are insanely difficult to get correct, and lots of even very skilled, accomplished artists, if you go over their work with a magnifying glass you're going to find at least some of these things, some weirdnesses that make no sense, some shadows that are off, some fingers that are just ugh (really, getting hands wrong is so relatable. hands are the fucking worst). It's not about "this is bad art/not art because the hand is wrong," it's specifically about the ways that it's wrong, the way a computer randomly throws pieces together versus how actual people make actual mistakes. It's all of the little signs taken as a whole to say "no one who could produce a piece that, on the surface, looks this nice, could possibly make THIS MANY small 'mistakes.'"

The absolute best thing you can do if you see AI-generated images being treated as real art is just nothing. Support actual artists you love, and don't spread the fakes.

Thanks for your time, everyone. Good luck avoiding AI-generated pieces in the future, please signal boost this, and feel free to get in touch if you think I can help you with anything related to this.

3K notes · View notes