5 ms·

People are way overvaluing this thing right? AI "art", prompt "engineering". All I see is people with too much time throwing words at a parser without understan

by zorked 4y ago

People are way overvaluing this thing right? AI "art", prompt "engineering". All I see is people with too much time throwing words at a parser without understanding what it does until something sticks. And the result is probably a copyright violation of some poor artist's work.

O__________O 4y ago
Little strange to comment that users do not understand a technology, then comment that outputs are copyright violations; which to me says you neither understand copyright, nor how the technology works.
- astorsnk 4y ago
  Legal or illegal, there's a big ethical debate to be had. Swallowing up the work of others and using that to generate derivative/new work that _a lot_ of people (I've seen half a dozen stock photo AI sites this week alone) are then trying to use to replace the original creators doesn't feel right.
  - O__________O 4y ago
    Art history is literally full of transformation of prior works, technology is clearly transformative (if you understand it), copyright law clearly allows for it, and indexing publicly available data is also legal — there’s no reasonable basis for debating the topic being illegal or unethical.
    - tsbinz 4y ago
      And yet, every time GitHub Copilot comes up, people on HN are in arms about their code being used to train a model.
      - O__________O 4y ago
        Already responded to comparison to GitHub CoPilot in this thread here: https://news.ycombinator.com/item?id=33079278 https://news.ycombinator.com/item?id=33079278
    - falcolas 4y ago
      Legal and ethical are two wholly different concepts. There's quite a few things which are legal and not at all ethical. Today in many states it's perfectly legal to shoot a native american if you're in a wagon circle. Is that ethical? Fair use is also a positive defense, which means that only the courts can really say if something is fair use or not.
      - O__________O 4y ago
        First line, from the post you replied to: >> “Art history is literally full of transformation of prior works” There’s no reasonable ethical basis to take issue with. You’re examples and counter points to me add substantive your claims. What am I missing?
        falcolas 4y ago
        This isn't really transformation though. It's compositing pixel-accurate bits and pieces of multiple pieces of art to make a new piece (as proven by stock photo watermarks being replicated in some pieces). That difference is enough to assert that there is a wholly reasonable ethical basis. Especially when they're obviously not just training it on images with relaxed copyrights.
        O__________O 4y ago
        A collage of prior works is protected as a transformative work, regardless of if you’re able to recognize the seams between the pieces. Even use of trademarks is not protected in works of art if there is no clearly defined trademark violation; for example, putting fries in a box that looks identical to McDonald fries box and offering them for sell would not be legally justifiable by just saying it is art. Ethically speaking, for the second time, you ignored that within art world, or real world for that matter, making derivative works is completely ethical if done within any related legal constraints; as such, to me, you’re not making a good faith effort, which is actually unethical if intentionally done, so I will not be replying any further.
      - dotancohen 4y ago
  - simion314 4y ago
    Someone else can skip training with copyrighted images, so this copyright argument would be unusable there. Competent creators will not be replaced, they will use the tools to do their job faster/better. I remember soem tiem ago artists having reservations on "digital art". Someone said it better, a person with a camera is not a photographer. the problems I have with Microsoft Copilot is that the fuckers did not used MS proprietary code train but only trained with other people GPL/MIT code, this was shitty and implies that their implementation is "dangerous" to use with proprietary code but their expensive lawyers can handle a bunch of open source small developers.
    - O__________O 4y ago
      GitHub’s terms of service explicitly allow them to do what they did, if anyone is in violation of the licenses you’re referencing it’s the people that loaded the code into GitHub.
      - tsbinz 4y ago
        That is not how copyright law works. If someone uploads an image as public domain, or code as a BSD license, without having the right to do so, that doesn't give others the right to use those things.
        O__________O 4y ago
        Unless I am missing something, nothing I said prior conflicts with your point, nor shows that GitHub is willfully violating the law. Are you aware of any copyrighted material that was illegally loaded to GitHub that has been proven to be used by CoPilot, it’s use was not transformative — AND - GitHub has not removed it from the model?
        simion314 4y ago
        The code used in GitHub copilot is all copyrighted with GPL, MIT,BSD. Also lots of the code was not uploaded to GitHub by the authors so GitHub shity ToS won't apply so you can't claim that the authors agreed to some vague terms that we can now intepret as allows only GitHub to create copilot. You did 1 MS did not violate copyright since is open source code, andfor soem reason you think that only proprietary code should be removed from copilot 2 even if MS violated the license the guilty person is the dude that uploaded the code to GitHub and click I Agree not he ToS, implying that somehow the GitHub ToS has the ability to change open source license(but t proprietary ones, since you claim MS "removed" such code when it was revealed). As I mentioned above, MS are cowards, iof their thing is real creating new stuff then put their ass on the line and put their own code in too, put all proprietary code in GitHub in copilot too and then have their well paid lawyers defend their tech.
- Bakary 4y ago
  The law hasn't caught up, and might never do. But surely a generator that even sometimes inputs the artists' signature (or at least a bizarro version of it) in the output meets the spirit if not the letter of what copyright law is trying to achieve? At the very least you have artists in a more precarious position than they were before.
  - O__________O 4y ago
    Copyright law allows for transformative use, signature alone in some from would not negate the work being transformative. If it was claimed to be work of art by an artist associated with a signature, that would be fraud, which is completely different. If signature was trademarked, that might be another basis for legal claims, but unrelated to copyright.
    - Bakary 4y ago
      Again, I am aware that legally there's nothing they can do, but the point of copyright law is to facilitate making a living from art (let's ignore the House of Mouse shenanigans for the sake of the argument). A service that is able to pop out a new piece that will be indistinguishable from the artist's style is going to be disastrous for hundreds of thousands of artists around the world. We can't hide, at least not in good faith, behind the concept of 'transformative' art. Besides, there clearly hasn't been enough time for the legal system to catch up so arguments about the letter of the law might change too. If you look at musical cases you'll be surprised to see that the 'transformative' part is up to interpretation. Just ask the Verve...
      - O__________O 4y ago
        Copyright law protects finished works of art. It does not protect things like facts, ideas, procedures, or an artist's style, no matter how distinct. And legal system is up to date in any material matter related to the topic. If is not, you not only need to provide a specific notable claim, but back it up with evidence to support it and state why it’s not addressed by current legal code, case law, etc.
- falcolas 4y ago
  When you remix someone's art, it's fair use, but when you remix someone's song, they get royalties? Seems like perhaps copyright is more complicated than folks defending AI Art believe it is.
  - O__________O 4y ago
    Transformative use of prior works requires no royalties.
    - falcolas 4y ago
      With all due respect, you're wrong. See Tailor Swift's recent, successful, copyright judgement. And frankly, unless you are a judge overseeing a specific copyright infringement case, you have no basis with which to make these assertions.
      - O__________O 4y ago
        As someone who has personally litigated cases, even gotten judges to issues orders by phone, I have every right to present reasonable claims; as do you, unfortunately you’re not doing that. Case you referenced is scheduled for hearing on January 2023 and I don’t see any reason to review it until it’s finalized. If I am missing something, please feel free to explicitly state what and why it is related to current topic.
andrepd 4y ago
You're getting downvoted for a decidedly un-hn opinion, but that is my perception as well. You throw 30 prompts at gpt3 until you get what you know you wanted. Then you post saying "look! gpt3 has domain knowledge". Please...
literalAardvark 4y ago
Not really. The model doesn't store or reproduce anyone's art, maybe except with some intentional img2img prompting. And styles aren't copyrightable. There's some really impressive stuff being created out there, and it's still very early days. People take sets of personal images, train a mini model on that and use that model to prompt Stable Diffusion to create something that it wasn't even trained on. There are interesting short animations, etc. It's a brave new world to explore.
sanitycheck 4y ago
I've been having immense fun with SD in the past few days, but I think you're right. Prompt "engineering" is educated guesses plus experience, a good way to get a good (txt2img) image seems to be to iterate on prompts until something promising appears then leaving it to churn out 50 similar images with varying parameters. Using the "by {artistname}" string in prompts does seem to produce a lot of stuff which could potentially be copyright violation - it'll even output a signature/initials fairly frequently. In the next year or two I'm sure we'll see some a refined version of something like img2img integrated into Photoshop, maybe Adobe will finally justify their "Cloud" approach with a big stack of GPUs in a data center.

by zorked 4y ago

O__________O 4y ago

Little strange to comment that users do not understand a technology, then comment that outputs are copyright violations; which to me says you neither understand copyright, nor how the technology works.

astorsnk 4y ago
Legal or illegal, there's a big ethical debate to be had. Swallowing up the work of others and using that to generate derivative/new work that _a lot_ of people (I've seen half a dozen stock photo AI sites this week alone) are then trying to use to replace the original creators doesn't feel right.
- O__________O 4y ago
  Art history is literally full of transformation of prior works, technology is clearly transformative (if you understand it), copyright law clearly allows for it, and indexing publicly available data is also legal — there’s no reasonable basis for debating the topic being illegal or unethical.
  - tsbinz 4y ago
    And yet, every time GitHub Copilot comes up, people on HN are in arms about their code being used to train a model.
    - O__________O 4y ago
      Already responded to comparison to GitHub CoPilot in this thread here: https://news.ycombinator.com/item?id=33079278 https://news.ycombinator.com/item?id=33079278
  - falcolas 4y ago
    Legal and ethical are two wholly different concepts. There's quite a few things which are legal and not at all ethical. Today in many states it's perfectly legal to shoot a native american if you're in a wagon circle. Is that ethical? Fair use is also a positive defense, which means that only the courts can really say if something is fair use or not.
    - O__________O 4y ago
      First line, from the post you replied to: >> “Art history is literally full of transformation of prior works” There’s no reasonable ethical basis to take issue with. You’re examples and counter points to me add substantive your claims. What am I missing?
      - falcolas 4y ago
        This isn't really transformation though. It's compositing pixel-accurate bits and pieces of multiple pieces of art to make a new piece (as proven by stock photo watermarks being replicated in some pieces). That difference is enough to assert that there is a wholly reasonable ethical basis. Especially when they're obviously not just training it on images with relaxed copyrights.
        O__________O 4y ago
        A collage of prior works is protected as a transformative work, regardless of if you’re able to recognize the seams between the pieces. Even use of trademarks is not protected in works of art if there is no clearly defined trademark violation; for example, putting fries in a box that looks identical to McDonald fries box and offering them for sell would not be legally justifiable by just saying it is art. Ethically speaking, for the second time, you ignored that within art world, or real world for that matter, making derivative works is completely ethical if done within any related legal constraints; as such, to me, you’re not making a good faith effort, which is actually unethical if intentionally done, so I will not be replying any further.
    - dotancohen 4y ago
- simion314 4y ago
  Someone else can skip training with copyrighted images, so this copyright argument would be unusable there. Competent creators will not be replaced, they will use the tools to do their job faster/better. I remember soem tiem ago artists having reservations on "digital art". Someone said it better, a person with a camera is not a photographer. the problems I have with Microsoft Copilot is that the fuckers did not used MS proprietary code train but only trained with other people GPL/MIT code, this was shitty and implies that their implementation is "dangerous" to use with proprietary code but their expensive lawyers can handle a bunch of open source small developers.
  - O__________O 4y ago
    GitHub’s terms of service explicitly allow them to do what they did, if anyone is in violation of the licenses you’re referencing it’s the people that loaded the code into GitHub.
    - tsbinz 4y ago
      That is not how copyright law works. If someone uploads an image as public domain, or code as a BSD license, without having the right to do so, that doesn't give others the right to use those things.
      - O__________O 4y ago
        Unless I am missing something, nothing I said prior conflicts with your point, nor shows that GitHub is willfully violating the law. Are you aware of any copyrighted material that was illegally loaded to GitHub that has been proven to be used by CoPilot, it’s use was not transformative — AND - GitHub has not removed it from the model?
        simion314 4y ago
        The code used in GitHub copilot is all copyrighted with GPL, MIT,BSD. Also lots of the code was not uploaded to GitHub by the authors so GitHub shity ToS won't apply so you can't claim that the authors agreed to some vague terms that we can now intepret as allows only GitHub to create copilot. You did 1 MS did not violate copyright since is open source code, andfor soem reason you think that only proprietary code should be removed from copilot 2 even if MS violated the license the guilty person is the dude that uploaded the code to GitHub and click I Agree not he ToS, implying that somehow the GitHub ToS has the ability to change open source license(but t proprietary ones, since you claim MS "removed" such code when it was revealed). As I mentioned above, MS are cowards, iof their thing is real creating new stuff then put their ass on the line and put their own code in too, put all proprietary code in GitHub in copilot too and then have their well paid lawyers defend their tech.
Bakary 4y ago
The law hasn't caught up, and might never do. But surely a generator that even sometimes inputs the artists' signature (or at least a bizarro version of it) in the output meets the spirit if not the letter of what copyright law is trying to achieve? At the very least you have artists in a more precarious position than they were before.
- O__________O 4y ago
  Copyright law allows for transformative use, signature alone in some from would not negate the work being transformative. If it was claimed to be work of art by an artist associated with a signature, that would be fraud, which is completely different. If signature was trademarked, that might be another basis for legal claims, but unrelated to copyright.
  - Bakary 4y ago
    Again, I am aware that legally there's nothing they can do, but the point of copyright law is to facilitate making a living from art (let's ignore the House of Mouse shenanigans for the sake of the argument). A service that is able to pop out a new piece that will be indistinguishable from the artist's style is going to be disastrous for hundreds of thousands of artists around the world. We can't hide, at least not in good faith, behind the concept of 'transformative' art. Besides, there clearly hasn't been enough time for the legal system to catch up so arguments about the letter of the law might change too. If you look at musical cases you'll be surprised to see that the 'transformative' part is up to interpretation. Just ask the Verve...
    - O__________O 4y ago
      Copyright law protects finished works of art. It does not protect things like facts, ideas, procedures, or an artist's style, no matter how distinct. And legal system is up to date in any material matter related to the topic. If is not, you not only need to provide a specific notable claim, but back it up with evidence to support it and state why it’s not addressed by current legal code, case law, etc.
falcolas 4y ago
When you remix someone's art, it's fair use, but when you remix someone's song, they get royalties? Seems like perhaps copyright is more complicated than folks defending AI Art believe it is.
- O__________O 4y ago
  Transformative use of prior works requires no royalties.
  - falcolas 4y ago
    With all due respect, you're wrong. See Tailor Swift's recent, successful, copyright judgement. And frankly, unless you are a judge overseeing a specific copyright infringement case, you have no basis with which to make these assertions.
    - O__________O 4y ago
      As someone who has personally litigated cases, even gotten judges to issues orders by phone, I have every right to present reasonable claims; as do you, unfortunately you’re not doing that. Case you referenced is scheduled for hearing on January 2023 and I don’t see any reason to review it until it’s finalized. If I am missing something, please feel free to explicitly state what and why it is related to current topic.

andrepd 4y ago

You're getting downvoted for a decidedly un-hn opinion, but that is my perception as well. You throw 30 prompts at gpt3 until you get what you know you wanted. Then you post saying "look! gpt3 has domain knowledge". Please...

literalAardvark 4y ago

Not really. The model doesn't store or reproduce anyone's art, maybe except with some intentional img2img prompting. And styles aren't copyrightable. There's some really impressive stuff being created out there, and it's still very early days. People take sets of personal images, train a mini model on that and use that model to prompt Stable Diffusion to create something that it wasn't even trained on. There are interesting short animations, etc. It's a brave new world to explore.

sanitycheck 4y ago

I've been having immense fun with SD in the past few days, but I think you're right. Prompt "engineering" is educated guesses plus experience, a good way to get a good (txt2img) image seems to be to iterate on prompts until something promising appears then leaving it to churn out 50 similar images with varying parameters. Using the "by {artistname}" string in prompts does seem to produce a lot of stuff which could potentially be copyright violation - it'll even output a signature/initials fairly frequently. In the next year or two I'm sure we'll see some a refined version of something like img2img integrated into Photoshop, maybe Adobe will finally justify their "Cloud" approach with a big stack of GPUs in a data center.