zlacker

[parent] [thread] 27 comments
1. xg15+(OP)[view] [source] 2022-12-15 12:39:56
If thouse mouse images are generated, that implies that Disney content is already part of the training data and models.

So in effect, they are pitting Disney's understanding of copyright (maximally strict) against that of the AI companies (maximally loose).

Even if it's technically the responsibility of the user not to publish generated images that contain copyrighted content, I can't imagine that Disney is very happy with a situation where everyone can download Stable Diffusion and generate their own arbitrary artwork of Disney characters in a few minutes.

So that strategy might actually work. I wish them good luck and will restock my popcorn reserves just in case :)

The problem I see though is that both sides are billion dollar companies - and there is probably a lot of interest in AI tech within Disney themselves. So it might just as well happen that both sides find some kind of agreement that's beneficial for both of them and leaves the artists holding the bag.

replies(4): >>astran+h1 >>taeric+G2 >>wnkrsh+b5 >>jamesh+lj
2. astran+h1[view] [source] 2022-12-15 12:46:18
>>xg15+(OP)
> If thouse mouse images are generated, that implies that Disney content is already part of the training data and models.

It doesn't mean that. You could "find" Mickey in the latent space of any model using textual inversion and an hour of GPU time. He's just a few shapes.

(Main example: the most popular artist StableDiffusion 1 users like to imitate is not in the StableDiffusion training images. His name just happens to work in prompts by coincidence.)

replies(2): >>Taywee+Z5 >>mcv+wk
3. taeric+G2[view] [source] 2022-12-15 12:53:07
>>xg15+(OP)
This is a bit silly, though? Search Google images for Mickey Mouse, is the results page a possible liability for Google? Why not?

Go to a baker and commission a Mickey Mouse cake. Is that a violation if the bakery didn't advertise it? (To note, a bakery can't advertise it due to trademark, not copyright. Right?)

For that matter, any privately commissioned art? Is that really what artists want to lock away?

replies(4): >>logifa+04 >>crote+Z4 >>wongar+76 >>sigmoi+R6
◧◩
4. logifa+04[view] [source] [discussion] 2022-12-15 13:00:02
>>taeric+G2
> Search Google images for Mickey Mouse, is the results page a possible liability for Google?

In 2018[0], didn't Getty force Google to change how Google Images presented results, following a lawsuit in 2016[1]?

[0] https://arstechnica.com/gadgets/2018/02/internet-rages-after... [1] https://arstechnica.com/tech-policy/2016/04/google-eu-antitr...

◧◩
5. crote+Z4[view] [source] [discussion] 2022-12-15 13:06:25
>>taeric+G2
> Is the results page a possible liability for Google?

Absolutely. Google previously had a direct link to the full-size image, but it has removed this due to potential legal issues. See [0].

> Is that a violation if the bakery didn't advertise it?

According to Disney, it is. See [1].

> Any privately commissioned art?

Not any art, no. Only that which uses IP/material they do not have a license to.

[0]: https://www.ghacks.net/2018/02/12/say-goodbye-to-the-view-im...

[1]: https://en.wikipedia.org/wiki/Cake_copyright#Copyright_of_ar...

replies(1): >>taeric+sF
6. wnkrsh+b5[view] [source] 2022-12-15 13:07:10
>>xg15+(OP)
You can search the LAION5B CLIP-space and you find a lot of mickey in it, lots of fan art between photos of actual merch. If you search with a high aesthetic score, you'll find lots of actual Disney illustrations etc. in the neighbourhood. [0]

[0] https://rom1504.github.io/clip-retrieval/

replies(1): >>xg15+ib
◧◩
7. Taywee+Z5[view] [source] [discussion] 2022-12-15 13:11:29
>>astran+h1
If you can find a copyrighted work in that model that wasn't put there with permission, then why would that model and its output not violate the copyright?
replies(2): >>astran+d6 >>mcv+bm
◧◩
8. wongar+76[view] [source] [discussion] 2022-12-15 13:11:59
>>taeric+G2
The right to citation is already part of the 1886 Berne Convention, a precedent that enables services like Google images.

The matters of the baker and the privately comissioned art are more complicated. The artist and baker hold copyrigh for their creation, but their products are also derived from copyrighted work, so Disney also has rights here [1]. This is just usually not enforced by copyright holders because who in their right mind would punish free marketing.

1: https://en.wikipedia.org/wiki/Derivative_work

◧◩◪
9. astran+d6[view] [source] [discussion] 2022-12-15 13:12:30
>>Taywee+Z5
https://en.wikipedia.org/wiki/The_Library_of_Babel

A latent space that contains every image contains every copyrighted image. But the concept of sRGB is not copyrighted by Disney just yet.

replies(1): >>Taywee+C8
◧◩
10. sigmoi+R6[view] [source] [discussion] 2022-12-15 13:15:32
>>taeric+G2
>is the results page a possible liability for Google?

That's actually a tricky question and lengthy court battles were held over this in both the US and Europe. In the end, all courts decided that the image result page is questionable when it comes to copyright, but generally covered by fair use. The question is how far fair use goes when people are using the data in derivative work. Google specifically added licensing info about images to further cover their back, but this whole fair use stuff gets really murky when you have automatic scrapers using google images to train AIs who in turn create art for sale eventually. There's a lot of actors in that process that profit indirectly from the provided images. This will probably once again fall back to the courts sooner or later.

replies(1): >>red_tr+Od
◧◩◪◨
11. Taywee+C8[view] [source] [discussion] 2022-12-15 13:23:50
>>astran+d6
Sure, but this isn't philosophy. An AI model that contains every image is a copyright derivative of all those images and so is the output generated from it. It's not an abstract concept or a human brain. It's a pile of real binary data generated from real input.
replies(1): >>astran+kb
◧◩
12. xg15+ib[view] [source] [discussion] 2022-12-15 13:38:04
>>wnkrsh+b5
Yes, and probably the copyrighted art of lots of other artists as well. That's the entire point.
◧◩◪◨⬒
13. astran+kb[view] [source] [discussion] 2022-12-15 13:38:11
>>Taywee+C8
StableDiffusion is 4GB which is approximately two bytes per training image. That's not very derivative, it's actual generalization.

"Mickey" does work as a prompt, but if they took that word out of the text encoder he'd still be there in the latent space, and it's not hard to find a way to construct him out of a few circles and a pair of red shorts.

◧◩◪
14. red_tr+Od[view] [source] [discussion] 2022-12-15 13:52:16
>>sigmoi+R6
Europe has no concept of Fair Use. How did the courts argue there?
replies(2): >>FinnKu+lg >>sigmoi+Fn
◧◩◪◨
15. FinnKu+lg[view] [source] [discussion] 2022-12-15 14:03:51
>>red_tr+Od
Not a lawyer, but from how I understand it the German courts argued that if you don't use any technology to prevent web crawlers from accessing the pictures on your website you need to accept that they are used for preview images (what the Google picture search technically is) as this is a usual use case.

-> here is the actual judgement though: https://juris.bundesgerichtshof.de/cgi-bin/rechtsprechung/do...

16. jamesh+lj[view] [source] 2022-12-15 14:16:29
>>xg15+(OP)
There's nothing wrong with the model knowing what Mickey Mouse looks like.

There are noninfringing usecases for generating images containing Mickey Mouse - not least, Disney themselves produce thousands of images containing the mouse's likeness every year; but also parody usecases exist.

But even if you are just using SD to generate images, if we want to make sure to avoid treading on Disney's toes, the AI would need to know what Mickey Mouse looks like in order to avoid infringing trademark, too. You can feed it negative weights already if you want to get 'cartoon mouse' but not have it look like Mickey.

The AI draws what you tell it to draw. You get to choose whether or not to publish the result (the AI doesn't automatically share its results with the world). You have the ultimate liability and credit for any images so produced.

replies(1): >>xg15+2m
◧◩
17. mcv+wk[view] [source] [discussion] 2022-12-15 14:22:23
>>astran+h1
How do you get that coincidence? To be able to accurately respond to the cue of an artist's name, it has to know the artist, doesn't it?

In any case, in the example images here, the AI clearly knew who Mickey is and used that to generate Mickey Mouse images. Mickey has got to be in the training data.

replies(1): >>esrauc+971
◧◩
18. xg15+2m[view] [source] [discussion] 2022-12-15 14:28:26
>>jamesh+lj
Not a lawyer (and certainly no disney lawyer), but my understanding was that copyright is specifically concerned with how an image is created, less so that it is created. Which is why you can copyright certain recordings that only consist of silence. It just prevents you from using this record to base your own record of silence on, it doesn't generally block you from recording silence.

In the same way, making the model deliberately unable to generate Micky Mouse images would be much more far-reaching than just removing Micky imagery from the trainset.

replies(1): >>jamesh+Xn
◧◩◪
19. mcv+bm[view] [source] [discussion] 2022-12-15 14:29:03
>>Taywee+Z5
The idea behind that is probably that any artist learns from seeing other artists' copyrighted art, even if they're not allowed to reproduce it. This is easily seen from the fact that art goes through fashions; artists copy styles and ideas from each other and expand on that.

Of course that probably means that those copyrighted images exist in some encoded form in the data or neural network of the AI, and also in our brain. Is that legal? With humans it's unavoidable, but that doesn't have to mean that it's also legal for AI. But even if those copyrighted images exist in some form in our brains, we know not to reproduce them and pass them off as original. The AI does that. Maybe it needs a feedback mechanism to ensure its generated images don't look too much like copyrighted images from its data set. Maybe art-AI necessarily also has to become a bit of a legal-AI.

◧◩◪◨
20. sigmoi+Fn[view] [source] [discussion] 2022-12-15 14:34:55
>>red_tr+Od
Fair use is just a limitation of copyright in case of public interest. Europe has very similar exclusions, even though they are spelled out more concretely. But they don't make this particular issue any less opaque.
◧◩◪
21. jamesh+Xn[view] [source] [discussion] 2022-12-15 14:36:01
>>xg15+2m
Most Mickey Mouse image usage problems will be trademark infringement not copyright.

Copyright infringement does generally require you to have been aware of the work you were copying. So for sure there's an issue with using AI to generate art where you could use the tool to generate you an image, which you think looks original, because you are unaware of a similar original work, so you could not be guilty of copyright infringement - but if the AI model was trained on a dataset that includes an original copyrighted work that is similar, obviously it seems like someone has infringed something there.

But that's not what we're talking about in the case of mickey mouse imagery, is it? You're not asking for images of 'utterly original uncopyrighted untrademarked cartoon mouse with big ears' and then unknowingly publishing a mouse picture that the evil AI copied from Disney without your knowledge.

replies(1): >>xg15+Mt
◧◩◪◨
22. xg15+Mt[view] [source] [discussion] 2022-12-15 14:56:10
>>jamesh+Xn
> But that's not what we're talking about in the case of mickey mouse imagery, is it? You're not asking for images of 'utterly original uncopyrighted untrademarked cartoon mouse with big ears' and then unknowingly publishing a mouse picture that the evil AI copied from Disney without your knowledge.

I think this is exactly the problem that many artists have with imagine generators. Yes, we could all easily identify if a generated artwork contained popular Disney characters - but that's because it's Disney, owners of some of the most well-known IP in the world. The same isn't true for small artists: There is a real risk that a model reproduces parts of a lesser known copyrighted work and the user doesn't realise it.

I think this is what artists are protesting: Their works have been used as training data and will now be parts of countless generated images, all with no permission and no compensation.

replies(2): >>jamesh+p21 >>TchoBe+1C1
◧◩◪
23. taeric+sF[view] [source] [discussion] 2022-12-15 15:38:04
>>crote+Z4
I started to go down the rabbit hole of commissioned fan art. To say that that is a quagmire is an understatement. :(
replies(1): >>Macha+9C1
◧◩◪◨⬒
24. jamesh+p21[view] [source] [discussion] 2022-12-15 17:12:57
>>xg15+Mt
Right.

So Disney don’t need to worry about AI art tools - so ‘attacking’ them with such tools does nothing.

◧◩◪
25. esrauc+971[view] [source] [discussion] 2022-12-15 17:34:07
>>mcv+wk
For other artist cases the corpus can include many images that includes a description with phrases like "inspired by Banksy". Then the model can learn to generate images in the style of Banksy without having any copyrighted images by Banksy in the training set.

The Mickey Mouse case though is obviously bs, the training data definitely does just have tons of infringing examples of Mickey Mouse, it didn't somehow reinvent the exact image of him from first principles.

◧◩◪◨⬒
26. TchoBe+1C1[view] [source] [discussion] 2022-12-15 19:57:36
>>xg15+Mt
This already happens all the time in the current status quo with no need for AI.
◧◩◪◨
27. Macha+9C1[view] [source] [discussion] 2022-12-15 19:58:17
>>taeric+sF
I mean, isn't most of that "It's trademark infringement, but it is both financially tedious and a PR disaster to go after any but the most prominent cases"

Which is why e.g. Bethesda is not going to slap you for your Mr House or Pip-Boy fanart, but will slap the projects that recreate Fallout 3 in engine X.

replies(1): >>spooki+QU2
◧◩◪◨⬒
28. spooki+QU2[view] [source] [discussion] 2022-12-16 03:45:19
>>Macha+9C1
The tables turn when it's not just some fans doing it, which takes time and effort. AI generated images can be pumped out by the thousands, and big companies are behind these services. See the problem?
[go to top]