Commons:Deletion requests/Algorithmically generated AI artwork in specific styles by User:Benlisquare

From Wikimedia Commons, the free media repository
Jump to navigation Jump to search
This deletion discussion is now closed. Please do not make any edits to this archive. You can read the deletion policy or ask a question at the Village pump. If the circumstances surrounding this file have changed in a notable manner, you may re-nominate this file or ask for it to be undeleted.

Algorithmically generated AI artwork in specific styles by User:Benlisquare

[edit]

According to the creator's provided info on each, these images were generated through an AI using a prompt specifically asking to emulate the styles of specific artists (Hakurei Reimu, art style of Sophie Anderson, and greg rutkowski, and albert lynch. illustration of attractive girl, cleavage, wide hips, thick thighs, full body shot, dynamic posing, soft skin, beautiful, detailed, emphasis mine). It is not yet legally established that doing so respects the artists' copyright, and there is no way to verify that all of these hundred images are sufficiently distinct from existing works to receive independent copyright protection. Further, it presents an ethical nightmare of the wholesale copying of specific artists' styles, a practice which Commons should not endorse. Edit: The first and last artists are long dead. But the second is very much alive, and has protested the use of his name in AI-generated works: [1].

--Ovinus (talk) 02:01, 8 October 2022 (UTC)[reply]

  • Keep: An artist cannot copyright a brushstroke style, and there has been no legal precedent where mere motifs can be protected under copyright. Furthermore, using the example provided in the deletion nomination, a hybrid of Sophie Anderson's brushstroke style, Albert Lynch's brushstroke style, and Greg Rutkowski's brushstroke style would be an amalgamation of the three, which ultimately would resemble nothing of any of the three original styles. If nebulous visual ideas such as motifs can be copyrighted, then no artist on the planet would be able to create art any more, since anything that is made today would inevitably slightly resemble something else made 20 years ago, 40 years ago, or 60 years ago, by any one of thousands of artists. Works can be copyrighted: if they were alive today, the author of the Mona Lisa can copyright the Mona Lisa, but they cannot copyright the appearance of brushstroke patterns that resemble what is seen on the Mona Lisa. --benlisquareTalk•Contribs 02:40, 8 October 2022 (UTC)[reply]
  • Delete: This is outside of the scope of Wikimedia Commons as this user-generated 90+ image collection of "Prompt: busty young girl ..." images is "Self-created artwork without obvious educational use." These images are also outside the scope of Wikimedia Commons because they were posted here for self-promotional purposes, created and uploaded here so that the editor who created these 90+ images could spam their user-generated "busty young girl" images into Wikipedia. That this is also a likely copyright violation is clearly explained by Ovinus above with their citation to MIT Technology Review. As the MIT Technology Review article says, "Some artists may have been harmed in the process" of creating images such as this. We should not be hosting potentially harmful images. Elspea756 (talk) 16:16, 9 October 2022 (UTC)[reply]
  • The above allegations are not based in Commons policy; Commons is full of objectionable smut, whether or not the image contains attractive young women is unrelated to Commons scope. The images were earlier removed from an English Wikipedia article due to the concerns that they might be a copyvio, which is exactly what this deletion request aims to discuss and find a resolution to. If it is determined that the image is not a copyvio, it would still fall under Commons scope as it's clearly educational; from day one I've included a very detailed and thorough file description which instructs the reader on how to make an exact 1:1 replica of the image using open-source AI software, for anyone who is interested in the topic of AI machine learning; this is prime material for future expansion on Wikiversity or Wikibooks is it not? I'd also like to add that this user has been wikihounding me for days now. I don't understand what the obsession might be, but I'm getting rather bored of it. "Self-promotional purposes" - what am I self-promoting? Have I left a Patreon link or a DeviantArt link somewhere? I don't see anything of the sort on my userpage, or on the file description. This allegation is as nonsensical as it is dishonest. For the record, I am not an artist, and my dayjob is a software developer; I make content like this to teach other people how to use AI to create good looking images, it's as simple as that. --benlisquareTalk•Contribs 17:50, 9 October 2022 (UTC)[reply]

Detailed analysis of image content, and rationale why it is not a derivative work of artists named within AI generation prompts

[edit]

As outlined in the file description, each row corresponds to a particular text prompt that I fed into the AI image generation software known as Stable Diffusion; this model then interprets the text that I have input, and generates an image based on this interpretation. The text prompts I have laid out in the file description are there for the purpose of transparency of how I ultimately generated the image, in addition to giving readers a chance to learn through the instructions provided.

Let's go through each row, one by one. Consider the following example:

The prompt for this row is art style of artgerm and greg rutkowski, specifically focusing on the namedropped artists. If the AI was perfect (and I'll explain why it's not later below), then in theory this would mean that every image would be a rough amalgamation that's somewhere between Artgerm's artworks and Greg Rutkowski's works. Greg Rutkowski is a modern digital artist from Poland, and he showcases his art throughout his social media, so let's take a look at some of his works:

As you can clearly see, none of Greg Rutkowski's actual works resemble any of the AI generated images from the first row, including the example posted above. The brushstrokes are much more intricate and thicker, the lighting is not as flat, and his human facial anatomy is much more realistic compared to the animesque apperance within the AI generated image - just look at how massive the eyes are in comparison.

Now let's look at Artgerm, also known as Stanley Lau.

Looking at all three, but Supergirl in particular, his facial style is very much stylised in the "Western/American comic front cover" aesthetic, most particularly in terms of things like eyelashes and noses. None of these motifs are present in any of the AI generated images from row 1. His light shading in Supergirl is very dissimilar, while one might say it's slightly similar in the Widowmaker example, however it's preposterous to state that mere lighting can be copyrighted, since identical lighting techniques will have considerable overlap among many hundreds of different artists, no single artist "invented" this kind of lighting.

Onto the second row, consider the following example:

This prompt is art style of makoto shinkai and akihiko yoshida and hidari and wlop. Makoto Shinkai is a famous film director and animator, so let's take a look at his most well known works:

Sorry, but I'm just not seeing the resemblance. The chin in the AI-generated row is way too pointy, and the eyes resemble more of Youyume-kun's (where you see the eye sclera to the left and right of the iris) compared to Makoto Shinkai's (where you see the eye sclera all around the iris). Not to mention, Makoto Shinkai's animation style is much more reliant on bright rays of light hitting the subject. Then let's look at Akihiko Yoshida:

Sorry, not even in the same zipcode, especially the eyes and the technique used for skin colouring. Next, onto Hidari (illustrator):

Completely different linework, and the colouring almost resembles watercolour (even though it's digitally drawn). Finally, WLOP (also known as Wang Ling):

Very photorealistic in terms of human anatomy, while also going for the rough, thick brushstroke aesthetic. Essentially the antithesis of the animesque aesthetic from this entire row.

Onwards, we have the third row for art style of Michael Garmash:

Michael Garmash draws anatomically correct humans in a highly stylised, thick brushstroke aesthetic that feels very textured, if the painting was on a canvas in front of you, you'd be able to feel every single bump of the paint:

The AI generated row has neither the thick textured brushwork of Garmash's paintings, and while most of the body is anatomically correct, the giant eyes definitely ar not.

Next, we have Charlie Bowater and Lilia Alvarado and Sophie Gengembre Anderson and Franz Xaver Winterhalter, by Konstantin Razumov, by Jessica Rossier, by Albert Lynch:

Oh boy, a lot to unpack here. If I knew I'd be standing before a tribunal, I'd have used a simpler prompt, but I digress. Charlie Bowater draws anatomically correct humans, however the "pencilwork" (or at least the digital equivalent of it) in his shading is very visible and intentional, he makes no attempt to blur or hide them away and sticks to his style. Lilia Alvarado does action pose shots of human subjects, and her photography heavily features colour motifs. You cannot copyright the idea of her use of colour, you can only copyright her works. The works of Sophie Gengembre Anderson (1823 - 1903) are oil paintings that have much more texture and less use of overwhelming lighting compared to the AI images from this row, but otherwise I can see a slight similarity in face anatomy. Franz Xaver Winterhalter (1805 - 1873) paints realistic looking people, however, again, there is much less use of strongly contrasting lighting. Konstantin Razumov draws photorealistic women, but again, heavily emphasises his brushwork and does not blend them away. Jessica Rossier is a landscape artist and does not draw human portraits. Albert Lynch (1860 - 1950) paints anatomically correct humans, but also is more modest with his use of lighting just like the other artists mentioned earlier.

Onto row five, featuring art style of Jordan Grimmer, Charlie Bowater and Artgerm as the prompt:

We've already covered Artgerm and Charlie Bowater earlier (and the same points apply here as well). Jordan Grimmer is a landscape artist, and when he does draw humans, they're done in a heavily stylised, almost comic-like manner, a style which is completely absent from this row.

Row six has the prompt art style of ROSSDRAWS, very detailed deep eyes by ilya kuvshinov:

Rossdraws, also known as Ross Tran, draws excessively sexy women with thick lips and well-defined noses:

Honestly, I can see the resemblance from a face anatomy perspective. However at the same time, the execution does have its dissimilarities as well - Ross makes excellent use of lighting and shadows, while row six is considerably more lifeless by comparison; the example I've posted above is the best example of lively use of lighting, and even then it's still not akin to how Ross executes it, and the other images from this row are worse than this one. Ilya Kuvshinov's common motif across all their works involves round eyes:

This row similarly has round eyes. However, eyes alone aren't enough to warrant any claims of copyright violation, you can't copyright ideas like this, only works.

For this next one, we have little devilish face, yandere, insane, crazy, large devious smile, game cg japanese anime Jock Sturges Kyoto Animation Alexandre Cabanel Granblue Fantasy light novel pixiv; this time I opted to see what would happen if I mashed together nonsensical word salad, and the AI seemed to handle it pretty well:

All of these keywords I pulled out of a random prompt generator, without even looking into what some of these are. Jock Sturges is known for taking nude photographs of adolescents and then showcasing them at art galleries across the United States. I have no intention of looking up his work, but I assume none of these people look like smug anime girls. Alexandre Cabanel (1823 - 1889) paints anatomically realistic humans:

None of them which look like a smug anime girl.

Onwards:

This prompt is art style of Sophie Anderson, and greg rutkowski, and albert lynch. All three have already been covered above, and even in this example I don't see the resemblance.

This prompt is art style of Konstantin Razumov, and Jessica Rossier, and Albert Lynch. All three have already been covered above, and the closest resemblance I can see is Albert Lynch's use of posing, however Lynch's works significantly lack colour, whereas this AI row is utterly saturated with bright reds and shiny yellows.

Another animeqsue row, this prompt is hyper realistic anime painting (((sophie anderson))) (atelier meruru josei isekai by Krenz cushart by Kyoto Animation). For context, parentheses are used as emphasis markers in text prompts fed into Stable Diffusion; the more parentheses, the more you want that particular keyword to affect the final output. We've covered Sophie Anderson before (and her work is not animeqsue); Krenz Cushart draws art like this:

Very human-looking, very well defined anatomy, skillful use of lighting, and uses the aesthetic of a handbrushed painting (despite being digital art) with blended shades. Doesn't particularly match this row, especially in terms of anatomy and face shape. Kyoto Animation does works like this:

None of which match the style of eyes; the closest I can think of would be the chin shape (slightly similar to Hyouka), but you can't copyright a chin, especially since this style of chin is common throughout the entire anime industry (see this example), not just anime by KyoAni.

Next, we have art style of wlop and michael garmash:

Both artists covered before. This row lacks the texture of Michael Garmash's works, and it lacks the fine light detail of WLOP's works.

This one is art style of greg rutkowski and alphonse mucha:

We've covered before how Greg Rutkowski draws thick brushstrokes (completely absent from this row); Alphonse Mucha (1860 - 1939) draws stylised portraits such as this cigarette poster, and paints scenes such as this battle scene. This row's art features a more realistic looking aesthetic, and does not involve stylistic motifs like with the cigarette poster; the battle scene also lacks the kind of background detail seen in this row.

Final one:

This prompt is (((by Donato Giancola))), (Sophie Anderson), (Albert Lynch). The faces are quite similar to Sophie Anderson's paintings, and the posing resembles that of Albert Lynch. Donato Giancola is a science fiction and fantasy artist who specialises in imaginative settings:

I'm not seeing the resemblance, regardless whether it's painting technique, visual themes, or use of motifs. As with a lot of other artists mentioned earlier, Donato Giancola's brushwork has a lot of proper texture to it; the AI's generated art is blended and almost lifeless.

Now that we've looked at each AI-generated row, along with the artists mentioned within the generation prompt, let's wrap up the findings. Do you see any definite resemblance to the works that these artists have put out, or are we simply drawing quick conclusions because of the text prompts being fed into the AI? There is enough random noise in the AI image generation that the prompts will never perfectly line up with the output image. Where does this noise come from? The commas. The punctuation. The capitalisation. The spacing. The order in which words are arranged. Emphasis markers. Filler words. The output image literally changes, even if you use the same keywords, if commas are replaced with full stops. All these things add additional noise, and as we've seen above, this noise is enough to ensure that the output does not have a 1:1 match with all keywords within the prompt. Not to mention, even if there was no noise, it's impossible to have the AI model perfectly trained, there will be errors in the image-keyword pairs fed, and there will be skewed weights that arbitrarily favour some visuals over others.

To address Ovinus' line "there is no way to verify that all of these hundred images are sufficiently distinct from existing works" - We should not be concerned with "what-if" considerations where an AI generated piece of art might be a copyvio. If any piece of AI art is indeed a copyvio, then prove it; if the evidence is brought forward, then I will happily concede. The onus is on the accuser to show the original piece of art that is being derived from and prove that the uploaded file does not have a valid free licence, rather than the "what-if" that the image might coincidentally resemble an existing artwork too closely and that we just "haven't found it yet". If we hinged on what-ifs like this as a valid reason for deletion, then this would significantly harm the prospects of future free content.

As for this MIT Technology Review article, it is expected for human artists to feel threatened by new technology and voice their concerns; this is a very normal and understandable reaction, and the invention of the photography camera didn't do family portrait artists any favours either. However, the cogs of technology must keep on turning, and free content respositories such as Wikimedia Commons must reflect our everchanging world, regardless if the world changes for the better or for the worse. Once upon a time, computer was the name of a human occupation, someone who turned slide rules and scribbled numbers into notepads to put humankind on the moon; now, with the invention of the personal computer, you'll never see a job listing for one of these positions ever again. Should we feel sad for the human computers who were left out of a job? Or should we accept that the world keeps turning as new inventions are born? I'd even argue that savvy artists would exploit this new AI technology to push their abilities even further, rather than become replaced by it. But I digress - none of this is related procedurally to why these files should remain on Commons, and the same applies to the concerns raised by artists in that article. --benlisquareTalk•Contribs 10:33, 11 October 2022 (UTC)[reply]


  • Found this via ANI at enwp. There are probably two questions: COM:SCOPE and COM:DW. I think the extent to which AI art can be considered derivative is a giant open question. I saw another case where someone was actually seeding DALL-E with a copyrighted image and asking for alternative version of it, as though that doesn't obviously create a derivative. When it's pulling pieces from its vast library of references, however, it's unclear. I just sent an email to Wikilegal to see if it's something they'd weigh in on, since we're going to be asking this question more and more frequently. For now, I'm inclined to say it's PD unless you can point to the specific images it's a derivative of. Scope is harder. These were created for use in a Wikipedia article. Whether or not other editors agree they're useful in that article (I tend to think they're not appropriate for such purposes, but that's a conversation for enwp), it seems hard to argue based on SCOPE. No boldtext !vote from me, but probably keep for now. — Rhododendrites talk |  20:18, 22 December 2022 (UTC)[reply]
  • Delete: on the grounds that these serve no serious practical educational value as detailed in COM:NOTHOST. While Rhododendrites is right to say enWiki had to have a conversation, the conversation established that these images are unsuitable for any educational purposes. Additionally, I would prefer we wait until we know for a fact an image can be uploaded legally before allowing it, rather than being stuck checking after the fact; that's just part of the copyright policy. ~ Pbritti (talk) 03:59, 23 December 2022 (UTC)[reply]
  •  Delete this is an extremely long winded discussion for an unambiguous derivative work of ZUN’s copyrighted character. Dronebogus (talk) 03:35, 24 December 2022 (UTC)[reply]
  •  Weak keep There has been a ridiculous amount of discussion about this user (see their talk pages here and on enwiki and the linked admin noticeboard discussions on both projects); I am sure I have missed something. But from what I can tell, the user’s uploads were deemed unsuitable for the English Wikipedia. Of course, Wikipedia doesn’t appreciate a biased collection of dozens of images being packed into one of their articles. That does not mean that the images have no educational value; the uploader themselves said that they would be suitable for future work on Wikiversity or Wikibooks, and I am inclined to agree. Also, I haven’t seen anything approaching a consensus on the copyright issues.
These files might represent an extreme example of bias, but they are not the only such example. I saw generic female characters from about four other uploaders in the category Stable Diffusion; I saw no generic male characters. Brianjd (talk) 12:55, 24 December 2022 (UTC)[reply]

Deleted: Because of the detailed description these files can have an educational value. Due to the similarity to the original works they must be considered as non free derivative works. COM:PCP applies. --GPSLeo (talk) 12:04, 27 December 2022 (UTC)[reply]

Retrieved from "https://commons.wikimedia.org/w/index.php?title=Commons:Deletion_requests/Algorithmically_generated_AI_artwork_in_specific_styles_by_User:Benlisquare&oldid=731382724"