Commons:Deletion requests/Files uploaded by 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.

More AI artwork by Benlisquare

[edit]

Same derivative work copyright concerns as for Commons:Deletion requests/Algorithmically generated AI artwork in specific styles by User:Benlisquare. Should also be deleted as of COM:PCP.

GPSLeo (talk) 12:17, 27 December 2022 (UTC)[reply]

@GPSLeo File:Algorithmically-generated art of a French maid teasing the viewer.png:
Prompt: (highly detailed, shiny:1.1), (glossy hair:1.2), beautiful detailed eyes, (highly-detailed face:1.2), purple eyes, (open mouth), tongue out, big eyes, realistic, hair bangs, (full color), fully clothed. Image of a maid with an open mouth.
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, mosaic, ahegao, distorted mouth, multiple girls, (disembodied), (poorly drawn hands), (poorly drawn face), (mutation), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), bar censorship, censorship, elongated face, ((black and white))
I see no references to specific artists or artwork, so what is it supposed to be a derivative work of?
Choosing another random file that I haven’t looked at before (File:Algorithmically-generated photorealistic closeup painting of young woman.png), I find:
Prompt: female face, extremely detailed, intricate, clear_focus, perfect face, very deep eyes, round pupils
Negative prompt: (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, ((((mutated hands and fingers)))))
Again, what is this supposed to be a derivative work of? Brianjd (talk) 12:26, 27 December 2022 (UTC)[reply]
The software cuts the original works into small pieces runs some calculations over them and then generates a new image. The original work is part of the new generated file. Like File:WikimediaMosaicCapture.png consists of many photos these files also consist of the input photos. You can also not make a filme out of a playbook without the permission by the playbook author, even when you do major changes to the plot. GPSLeo (talk) 12:41, 27 December 2022 (UTC)[reply]
That's not how the software works. No part of the works fed into the code are literal part of the output. You can certainly make a generic comedy after studying many comedies. There are cases where images fed in repeatedly can be copied, like Mona Lisa, but in general it's just in the style of many works, and in these cases many artists.--Prosfilaes (talk) 13:21, 27 December 2022 (UTC)[reply]
Pinging @GPSLeo. Brianjd (talk) 13:39, 27 December 2022 (UTC)[reply]
If you look at File:This AI-generated woman does not exist.png you can definitely see parts coming directly from the input photos and are not just inspired. We deleted many much less clear cases under COM:PCP and COM:DM policies. If there is a court decisions (and there definitely will be many cases) that this kind of photos do not violate the rights of the original creators we can undelete those files. But for now we should not host them to not bring third party reusers at risk. GPSLeo (talk) 13:51, 27 December 2022 (UTC)[reply]
@GPSLeo I was not sure about this one:
Prompt: Lovely cousin Matilda, fair daughter of England. In Costco. Eighteen years old, a woman full-grown. My timid darling, my shining angel, O good heavens! Her eyes are like the morning sun. Her strange clothes are so amusing. I would trace the contour of her delicate hair. She is my heart's delight. 2018
Negative prompt: toy, B&W, nudity, (painting), outside, greenscreen, studio, ugly child
But you say you can see parts coming directly from the input photos? Can you cite examples?
Pinging @Prosfilaes. Brianjd (talk) 13:55, 27 December 2022 (UTC)[reply]
Lamps, product packing and shelf labels look to realistic to be totally generated. They are generated but the input was a copyrighted photo which was needed to create this. It is not just a simple information of the original photo it is a huge amount of information taken from the original photo. The amount of information taken from the photo is enough to be a derivative work. GPSLeo (talk) 14:04, 27 December 2022 (UTC)[reply]
@GPSLeo But the point made by Prosfilaes is that these are not copies of photos of lamps, packaging and labels, but rather composites of many different photos of those things.
Suppose a person draws a generic set of product packaging. At some level, that drawing would be a copy of various sets of packaging they had seen. But it would probably not be a copy of any particular packaging. And it would not be copyright infringement, despite most packaging being copyrighted. It’s the same idea with AI. Brianjd (talk) 14:09, 27 December 2022 (UTC)[reply]
This is how we end up with the nondescript and unrecognisable products for sale in the store, and the illegible signage, as noted in the file description. Brianjd (talk) 14:13, 27 December 2022 (UTC)[reply]
Would you say the same thing about a human work? The lamps are blurry bright rectangles in the distance, so I don't see why you would have needed to copy them. The packaging is among most "creative" parts of the picture; it shows the AI is trying to create a picture of packaging, but instead of copying anything, it creates its own image that shows it doesn't know what humans look for in packaging.--Prosfilaes (talk) 15:43, 27 December 2022 (UTC)[reply]
@GPSLeo: (from Special:Diff/720361864 by uploader, who is currently blocked) Comment: A typical Stable Diffusion-based model checkpoint file, for instance Stable Diffusion v1.5, is typically around 3.97GB in filesize. The AI model saw 6 billion pictures from its training dataset, and formed a set of algorithms stored in a single file totalling at roughly 4 billion bytes during the AI training process. Please explain to me how it is possible to fit an entire picture within 66% of a single byte of information? For context, the ASCII letter "a" is one byte, and the UTF-8 character "中" is two bytes. AI models contain algorithms that are trained on the training data, they do not store the training data, and you cannot retrieve something to photobash it into a collage if you do not store it. --benlisquareTalk•Contribs 01:51, 28 December 2022 (UTC)[reply]
 Keep ludicrously indiscriminate, you can’t copyright styles, concepts or vague elements. This has been discussed already Dronebogus (talk) 22:41, 27 December 2022 (UTC)[reply]
I’d like to add that prompts mean almost nothing and can only “steer” the AI in a general direction; I might upload some examples of how I used copyrighted media -based keywords that resulted in an output nothing like the thing in question. Dronebogus (talk) 09:25, 28 December 2022 (UTC)[reply]
I also was hesitant to delete the Sakuya and Marisa art because, unlike Reimu’s extremely distinctive design, it’s difficult to copyright things like “blonde witch girl” or “silver-haired french maid”. But YMMV. Dronebogus (talk) 09:41, 28 December 2022 (UTC)[reply]
  •  Keep per Dronebogus. You can't claim massive gross copyvios if you can't even tell us what copyright is being infringed. Especially not when the uploader has just been indef blocked to stop them discussing these images, and admins are using fake SPEEDY reasons and premature DR closures to hide as many of these as fast as possible. Andy Dingley (talk) 23:18, 27 December 2022 (UTC)[reply]
    I agree that admins are clearly abusing privileges here, but the uploader was behaving extremely poorly. Dronebogus (talk) 09:27, 28 December 2022 (UTC)[reply]
  • So what's to be done with the uploader? An indef block for un-Islamic behaviour? Banned from Commons for their behaviour on projects that aren't Commons? All their uploads summarily deleted because "we don't like them" ? Images deleted under CSD F10 as "they're not a contributor to Wikimedia projects", despite 50k edits?
"We're angry at this editor, lets block them, then delete all their content while they can't say anything to stop us" is not any sort of deletion policy we should be following. Andy Dingley (talk) 10:10, 28 December 2022 (UTC)[reply]
Only the first part (plus adding porn to a discussion to make a WP:POINT, which is blatantly disruptive), with no prejudice to an unblock. The rest was gravedancing and admin overreach pure and simple. Dronebogus (talk) 10:25, 28 December 2022 (UTC)[reply]
  • Weak delete. Following the Commons:Project scope/Precautionary principle of Commons #4, just because we can't tell whose copyright has been violated is not a sufficient argument. For example, without knowing the basis of File:This AI-generated woman does not exist.png we don't know if someone just superimposed an image of a woman onto a non-free image someone has from inside a Costco. I presume the woman is completely generated but we don't know if the computer literally took a separate image of a the body, or was this an image of a women inside a Costco and generated just the face. File:Algorithmically-generated AI artwork of Kirisame Marisa.png is specifically an AI-generated artwork based on copyrighted works. Since we don't know what the computer used as a basis for making these images, this isn't very far from "I found it on the internet, trust me, it's okay to be here." -- Ricky81682 (talk) 00:26, 28 December 2022 (UTC)[reply]
    There is precedent that we accept a human-drawn portrait that is derived from so many copyrighted images that no individual link to a particular image can be identified; see Commons:Village pump/Copyright/Archive/2021/07#Illustrating known people? for examples. I think of AI-generated artwork, such as File:This AI-generated woman does not exist.png, as the same. You should not think of generative AI models as cutting and pasting portions of unidentified copyrighted images, but rather constructing the new image from scratch having learned from its training set, much as a human artist studies the work of other artists to improve their skill. With human-drawn art, if the creator asserts that it does not use any copyrightable aspects of existing non-free works, then the burden of proof is on the person requesting deletion to identify a specific work or works that they believe it to be a close copy of; I don't see any reason we should treat AI differently. However, I agree that File:Algorithmically-generated AI artwork of Kirisame Marisa.png should be deleted per COM:FAN since a specific character is named. -- King of ♥ ♦ ♣ ♠ 01:01, 28 December 2022 (UTC)[reply]
    "Because a specific character is named" is a bizarre rationale. If the uploader had not put the name in the title, would you support keeping it? The next uploader will simply have an AI draw copyrighted pictures all day without telling us the source material. Why not feed the software just one image and have it "generate" a new one off that? Ricky81682 (talk) 02:59, 28 December 2022 (UTC)[reply]
    The problem isn't the file name, it's the prompt passed to Stable Diffusion that included "Kirisame Marisa". This isn't a new problem, this is applying the same standards as in 2015 for generic art vs. fan work of copyrighted figures. SnowFire (talk) 05:46, 28 December 2022 (UTC)[reply]
    +1 Exactly per SnowFire, as long as the prompt is not highly suggestive of a specific copyrighted subject then we should have no issues. It's like the difference between asking a human artist to draw Kirisame Marisa vs. asking them to draw an anime girl. -- King of ♥ ♦ ♣ ♠ 06:17, 28 December 2022 (UTC)[reply]
    Such an extreme interpretation of COM:PCP would mean that Commons would not allow any user-created photography either, given that there's always the chance that the uploader could be lying about the origin of the photo, and things such as EXIF aren't a foolproof sign of authenticity given that anyone with 3 minutes of reading time can use readily-available FOSS software to falsify EXIF information. This could even be extended to anything user-generated, from SVG diagrams to Wiktionary pronunciation recordings. Commons operates on the trust and goodwill of its contributors, and yes, from time to time that goodwill will certainly be tested by malicious actors, however this doesn't mean that COM:PCP can be used as a catch-all argument to make blanket deletions; if there genuinely are COM:DW concerns, they can be appropriately handled on a case-by-case basis. --benlisquareTalk•Contribs 23:59, 1 January 2023 (UTC)[reply]
    No, we presume when the uploader says that they created it, it is theirs. The EXIF is a separate argument. If I find the image on another website, that's evidence that it isn't theirs. If it is a screenshot, that's evidence. Evidence is evidence. If the uploader claims they made these pieces, we assume they did. If the uploader says "my friend made this", we say "get permission." If the uploader says "I had a monkey create it for me" we follow the copyright rules and say it's a free image. But if the uploader says "I had a computer create this, no one knows what it used to train on but it should be kept anyways," that isn't much different than "I found it on the internet, it may be copyrighted but I don't know whose copyright so it doesn't matter" to me. Ricky81682 (talk) 00:51, 5 January 2023 (UTC)[reply]
    AI is like the case where the uploader says "my friend made this", except that there is a law stating that any works created by this "friend" are ineligible for copyright. If the artwork were created by an actual human friend, we wouldn't ask them (the friend) to prove that it wasn't closely derived from the works of other artists; we'd just take their (the friend's) word for it unless there is evidence to the contrary. No one knows what previous copyrighted works our human friend has learned from, but we don't care. No reason to treat our AI friend any different from a human friend. -- King of ♥ ♦ ♣ ♠ 01:07, 5 January 2023 (UTC)[reply]
    If you actually believe that, then why not say all these images are public domain and not GFDL? This is the first time you've mentioned it. It would make derivative works someone more straightforward as the human component is required to have any license from then on. I don't think that's been the rulings by court as of yet. -- Ricky81682 (talk) 09:26, 5 January 2023 (UTC)[reply]
    @Benlisquare, King of Hearts, and Ricky81682: That is a good point that I did not notice before. Why not tag these files as {{PD-algorithm}}? Brianjd (talk) 10:02, 5 January 2023 (UTC)[reply]
    It's like why {{Licensed-PD-Art}} exists. Right now, if someone photographs an old 2D painting or prompts an AI image, the official Commons position is that they have no rights, and no license from them is required in order to host the image on Commons. However, there could be jurisdictions where this does not hold, or laws or their interpretation could change in the future, so {{Licensed-PD-Art}} is a fail-safe: maybe some day Bridgeman v. Corel is overturned and we have to delete all our PD art photographed by third parties, but the license will make sure that we can keep Commoner-photographed PD art. Similarly, maybe the legal landscape changes with AI art, so having a license that may or may not be necessary can't hurt but could possibly help. -- King of ♥ ♦ ♣ ♠ 17:54, 5 January 2023 (UTC)[reply]
    @Ricky81682 For example, without knowing the basis of File:This AI-generated woman does not exist.png we don't know if someone just superimposed an image of a woman onto a non-free image someone has from inside a Costco. I presume the woman is completely generated but we don't know if the computer literally took a separate image of a the body, or was this an image of a women inside a Costco and generated just the face. This is just making up accusations of copying – it’s not how AI actually works, as other users here have already explained. Brianjd (talk) 08:21, 5 January 2023 (UTC)[reply]
    I understand that the AI works by being trained on images. We don't know what the "in Costco" portion of the prompt was generated from. It is very specific to me which is why I question how many Costco images the AI is trained on. It wasn't "in a store" but specifically "in Costco." I highly doubt there are thousands of images tagged "Costco" but I don't know if the entire basis was copyrighted images. For example, I can't tell but it looks like the store signs are fictional/machine-generated nonsense rather than a foreign language. It's entirely a black box and a guess that this particular image or that particular image isn't problematic. If the prompt was "a women in the style of Emma Watson" versus "Emma Watson" versus "a women in Disneyland" versus "a women in Disneyland on June 20, 2021," we remain stuck at the black box problem. -- Ricky81682 (talk) 09:26, 5 January 2023 (UTC)[reply]
    @Ricky81682 I question how many Costco images the AI is trained on I don’t know about the AI training data specifically. But to give an idea, I just searched Commons for ‘Costco’ and, just from search result thumbnails, found Costco aisle in N. Plainfield 12.png, Costco aisle in N. Plainfield 1.png, Costco aisle in N. Plainfield 33.png, Costco aisle in N. Plainfield 34.png, Costco aisle in N. Plainfield 26.png, Costco aisle in N. Plainfield 8.png, Costco aisle in N. Plainfield 15.png, Inside Costco Perth.jpg, along with dozens of other images showing the interior of Costco stores. I’m not surprised that the AI recognises the term ‘Costco’.
    it looks like the store signs are fictional/machine-generated nonsense rather than a foreign language Yes, this point was raised earlier and supports the claim that this image was generated, not copied.
    But if you still think the AI is copying images, perhaps you could answer Benlisquare’s question above: How does the AI manage to fit an entire image into roughly ⅔ of a byte? Brianjd (talk) 09:59, 5 January 2023 (UTC)[reply]
  • Keep most, delete fan-art ( File:Algorithmically-generated AI artwork of Sakuya Izayoi.png and File:Algorithmically-generated AI artwork of Kirisame Marisa.png), and stick a pin on art that includes living artists in their prompts pending guidance from Wikimedia Foundation and their lawyers. Nominator has failed to elucidate a deletion rationale; re Ricky81682, I would argue that using something as a reference is not the same as violating their copyright. This is perhaps most clear in analogies of human-created images for, say, map data - scanning in a novel map from a copyrighted book isn't allowed, but a human using that copyrighted map as a reference to draw their own map is fine. On the other hand, drawing fan art of Mickey Mouse or whatever isn't fine because the result is still under copyright. As long as basic factual generic prompts are given like "woman" (rather than stuff like "Marisa Kirisame"), then we're closer to the "human redraws a map using copyrighted map as reference" rather than "human draws Mickey Mouse fan art using copyrighted official Disney art as reference." SnowFire (talk) 00:59, 28 December 2022 (UTC)[reply]
    • On the excursus on living artists - I do not think it is particularly viable to enforce such a restriction. The uploader has been helpful in uploading very descriptive prompts used to generate the art, but that won't stop less scrupulous uploaders from "forgetting" to mention any living artists used as prompts, or even claiming AI art as their own work outright - basically we'd need a requirement that only trackable, online art origins that show the prompt could be used, and even then those websites could go down in the future. And that still wouldn't be perfect once we start adding retouches on top of an initial AI-art layer when the retouching could actually be pasting in stuff generated with "forbidden" prompts. However, my suspicion is that if any restriction on AI art is ever made, it will be on using prompts that ask for the style of certain artists directly, or models largely trained on one artist's work, without there being some sort of opt-in / opt-out list. Quite a few of the prompts here reference specific artists, so if the WMF ever does make a judgment on this and also makes it retroactive, then those images may need to be revisited. (Also, insert obligatory disclaimer here that this should not serve as an endorsement of uploader's rather obvious fetishes, but this will be cured best by encouraging more people to upload descriptive AI Art on a wide range of topics, not just prompts of "large breasts" and the like.) SnowFire (talk) 00:59, 28 December 2022 (UTC)[reply]
      "Keep because someone less scrupulous will lie instead"? At that point, you may as well argue to keep everything that is "I found it on the internet" because so many people lie when they claim their uploads are "their own work." Ricky81682 (talk) 03:00, 28 December 2022 (UTC)[reply]
        • I specifically called out those images for suspicion and for potentially revisiting in the future - but given our current (lack of) guidance from the WMF, it would be premature to make up a policy that doesn't exist yet. If you'd like to create such a policy, you are welcome to lobby for it. (I will say that acknowledging the real world is not a flaw, and Commons both now and pre-AI art already hosted images with false copyright claims by uploaders. Discussing the feasibility of detecting such issues is absolutely a relevant thing to do.) SnowFire (talk) 05:43, 28 December 2022 (UTC)[reply]
  • Delete for only aforementioned fan art; the lack of consensus on AI copyright policy means it'll need to be sorted out at a higher level. ~ Pbritti (talk) 07:25, 28 December 2022 (UTC)[reply]
    So basically keep for now. Dronebogus (talk) 09:42, 28 December 2022 (UTC)[reply]
  • Please keep this discussion on-topic! We already have a separate discussion for the uploader’s conduct and the admins’s response to it. Brianjd (talk) 13:08, 28 December 2022 (UTC)[reply]
  • Keep per Dronebogus. This is a trainwreck nomination with far too many images to evaluate individually and the nominator hasn't provided a real argument for why they are bad. JPxG (talk) 20:09, 28 December 2022 (UTC)[reply]
  • Delete Some of it is pretty amazing stuff (such as the blurry step images) but I can't help but think these are all out of SCOPE, I also agree with Ricky's sentiments regarding the protentional copyright issues such as with this image. — Preceding unsigned comment added by Davey2010 (talk • contribs) 2022-12-30T01:57:45 (UTC)
    pretty amazing stuff detailing how to use a cutting edge technology is out of scope? Brianjd (talk) 03:00, 30 December 2022 (UTC)[reply]
    Agree, Davey2010’s vote is nonsensical on top of repeating the same vague copyright paranoia that has dogged these discussions. Dronebogus (talk) 03:33, 30 December 2022 (UTC)[reply]
    I'm sorry what ? ..... Something being "pretty amazing" doesn't automatically mean it's in scope. But hey we'll debunk my out-of-scope and potential copyright concerns as "nonsensical".
    But to answer Brian's comment - My assumption at that time was that none of the files were being used, none of the files would be used and that there were no categories for these images.... (even the most basic of categories here tend to be missing so it's fair to assume ai-related categories didn't exist here. –Davey2010Talk 12:45, 30 December 2022 (UTC)[reply]
    And yes, if art is of “pretty amazing” quality we generally keep it. Dronebogus (talk) 05:10, 31 December 2022 (UTC)[reply]
    No, we don't. Else this place would be full of artists keeping their artwork here because it looks "pretty amazing" (at least to them). Ricky81682 (talk) 00:47, 5 January 2023 (UTC)[reply]
  •  Info I asked for a statement on the page of WMF legal: m:Talk:Wikimedia Foundation Legal department#Statement on copyright of AI generated images needed. --GPSLeo (talk) 18:03, 30 December 2022 (UTC)[reply]

A bit off topic but i find the tone in this deletion request to be disheartening. --Trade (talk) 21:35, 31 December 2022 (UTC)[reply]

That’s partially my fault, and I apologize. Dronebogus (talk) 21:38, 31 December 2022 (UTC)[reply]
And I apologise for my comments in this DR too. I wish you all a Happy and Healthy 2023. –Davey2010Talk 22:12, 31 December 2022 (UTC)[reply]

 Delete all files where the prompt either specifies a copyrighted character or the art style of an artist whose work is still in copyright, but  Keep the rest. Two files where copyrighted characters are involved (Sakuya Izayoi, Kirisame Marisa) have already been deleted by Yann and I agree with that, as I would have copyright concerns in these cases, too, and Commons:Deletion requests/File:Algorithmically-generated art of a French maid teasing the viewer.png has also been deleted in a separate deletion request with reasonable arguments. But for the other files discussed here, the argument for deletion seems to unspecific, too broad. Commons:Deletion requests/Algorithmically generated AI artwork in specific styles by User:Benlisquare was a different case, as these images were based on the styles of specific artists, one of them living (Greg Rutkowski), and also involved a copyrighted character. However, AI-generated files based on generic prompts should be fine as long as it can't be shown that a specific (still copyrighted) source file has been slavishly copied (or important elements from that file have been directly used). Current AI algorithms such es the ones used for generating these images do much more than creating "derivative works", and so I would argue that for each individual case, it would be necessary to show which copyright is supposed to be violated. Gestumblindi (talk) 14:05, 11 March 2023 (UTC)[reply]

As an addition: Generating images in the style of artists which are long out of copyright, such as en:Sophie Gengembre Anderson who died in 1903, Michelangelo, or Van Gogh, should always be fine, no copyright concerns at all. In the previous request, living artist Rutkowski was included in the prompt, that might be an issue. Gestumblindi (talk) 14:50, 11 March 2023 (UTC)[reply]
You can’t copyright a style. Dronebogus (talk) 15:13, 11 March 2023 (UTC)[reply]
That's true, but if you expressly request the style of a still-copyrighted artist, a higher chance that copyrighted elements from their works crop up in the results is there, IMO, and so this might go against COM:PCP. On the other hand, if no particular artist is specified, of course there might still be copyrighted elements in the results, but the likelihood will not be as high as to trigger the precautionary principle. Gestumblindi (talk) 15:17, 11 March 2023 (UTC)[reply]
This is a very weird assertion. What constitutes a “copyrighted element”? Dronebogus (talk) 16:14, 11 March 2023 (UTC)[reply]
Well, for example, if you prompt an AI to generate an "exciting fantasy image in the style of Greg Rutkowski" and the resulting image contains a figure that looks much like the undead sorcerer in this one, among other things... then it might be a derivative work. Gestumblindi (talk) 17:41, 11 March 2023 (UTC)[reply]
You’ll need to link to a specific image that was used by the AI here, then. My original point stands. Dronebogus (talk) 07:15, 12 March 2023 (UTC)[reply]

 Keep Loss, in first-person view.png. This is a famous exploitable comic and within scope. --Trade (talk) 03:35, 19 March 2023 (UTC)[reply]

 Keep Whose copyright is violated? --Synotia (talk) 08:09, 27 March 2023 (UTC)[reply]

Kept: Apart from the three images that go against COM:FAN, there is no specific copyright being violated. —holly {chat} 18:10, 31 May 2023 (UTC)[reply]

Retrieved from "https://commons.wikimedia.org/w/index.php?title=Commons:Deletion_requests/Files_uploaded_by_Benlisquare&oldid=769154925"