Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Open-sourcing AudioCraft: Generative AI for audio

ai.meta.com

Post reply on HN
Print view View original

101–110 of 335 posts

Re: Open-sourcing AudioCraft: Generative AI for audio

#101

Post by emporas »

gilmore606 wrote: post #72
Pannoniae wrote: post #69

"generating new music in the style of existing music" will probably be a huge field soon. I can't wait for it to happen, it's a low-cost way of producing even more music to listen to.

> I can't wait for it to happen, it's a low-cost way of producing even more music to listen to. I can't really understand this. I'm a DJ and a huge music nerd, and I spend a lot of time every week discovering new music from the past 100 years and all over the world, and I'm constantly struck by _how much of it there is_. I've spent weeks just digging through psych-funk records from West Africa from the 1970s. How can…

There is a lot of human music for sure, great music from all eras, but just the other day i generated a song which was pure crystal harp. So, how many songs of crystal harp are out there? 1000 all n all? 10.000 maybe? Now i can generate one thousand crystal harp songs per day.

Re: Open-sourcing AudioCraft: Generative AI for audio

#102

Post by bottlepalm »

Anyone feel like with the flood of AI generated content there's a risk of the past being 'erased'. Like in 10 years we won't be able to tell if any information from the past is real or fake - sounds, pictures, videos, etc.. Like we need to start cryptographically signing all content now if there's any hope of being able to verify it as 'real' 10 years from now.

Re: Open-sourcing AudioCraft: Generative AI for audio

#103

Post by seydor »

bottlepalm wrote: post #102

Anyone feel like with the flood of AI generated content there's a risk of the past being 'erased'. Like in 10 years we won't be able to tell if any information from the past is real or fake - sounds, pictures, videos, etc.. Like we need to start cryptographically signing all content now if there's any hope of being able to verify it as 'real' 10 years from now.

The past ended in 2022

Re: Open-sourcing AudioCraft: Generative AI for audio

#105

Post by kmeisthax »

PaulDavisThe1st wrote: post #79
kmeisthax wrote: post #74

Earlier quoted context omitted.

Legal acquisition does not matter for AI training. If training is fair use then you can train on pirated material (e.g. OpenAI GPT). If it's not fair use then buying the material does not matter, you have to negotiate a specific license for AI training for each work in the training set, which is impractical at the scales most AI companies want to work.

This seems to distort the issue a little bit. If you purchase the music, you have a (sometimes explicit, sometimes implicit) license to do certain things with the music, entirely independent of any concept of "fair use". The question is not "is training part of fair use?" but "is training part of, implicitly or explicitly, the rights I already have after purchase?" Given that "training" can be done by simple playing…

In the US, exceptions to copyright come across in two distinct bundles: first sale and fair use. They exist specifically because of the intersection between copyright law and two other principles of the US constitution:

- First sale: The Takings Clause prohibits government theft of private property without compensation. Because copyright owners are using a government-granted monopoly to enforce their rights, we have to bound those rights to avoid copyright owners being able to just come and take copies of books or music you've lawfully purchased.

- Fair use: The 1st Amendment prohibits government prohibitions on free speech. Because copyright owners are using a government-granted monopoly to enforce their rights, we have to bound those rights to avoid copyright owners being able to censor you.

If you hinge your argument on "I bought a copy", you're making a first sale argument.

Notably, first sale is limited to acts that do not create copies. This limit was established by the ReDigi case[0]. Copyright doesn't care about the total number of copies in circulation, it cares about the right to create more. So an AI training defense based on first sale grounds would fail because training unequivocally creates copies.

Fair use, on the contrary, does not care if you bought a copy of a work legally. It only cares about balancing your right to speech against the owners' right to a monopoly over theirs. And it has so far been far more resistant to creative industry attempts to limit exceptions to copyright - to the point where I would argue that "fair use" is an effective shorthand for any exception to copyright, including ones in countries that have no fair use doctrine and do not respect judicial precedent.

The courts won't care how the training comes about, just if the act of training an AI alone[1] would compete with licensing the images used in the training set data.

[0] https://en.wikipedia.org/wiki/Capitol_Records,_LLC_v._ReDigi....

[1] Notably, this is separate from the act of using the AI to generate new artistic works, which may be infringing

Re: Open-sourcing AudioCraft: Generative AI for audio

#106

Post by stale2002 »

kmeisthax wrote: post #81
eropple wrote: post #34

Earlier quoted context omitted.

> MusicGen's weights are licensed CC-BY-NC, which is effectively a nonlicense as there is no noncommercial use you could make of an art generator How do you figure? Have you never just...made stuff to make stuff?

In copyright law the use of the work itself is considered a commercial benefit, so "noncommercial use" is an oxymoron. Consider these situations: - If I use AudioCraft to post freely-downloadable tracks on my SoundCloud, I still get the benefit of having a large audio catalog in my name, even if I'm not selling the individual tracks. I could later compose tracks on my own and ride off the exposure I got from posting…

This is a weird comment.

Do you think that non commercial use simply doesn't exist or something?

Because non commercial use isn't some crazy concept. It is a well established one, that doesnt disclude literally everything.

Also, you are ignoring the idea that Facebook will almost certainly not sue anyone for using this for any reason, except possibly Google or Apple.

So if you aren't literally one of those companies you could probably just use it anyway, ignore the license completely, and have zero risk of being sued.

Re: Open-sourcing AudioCraft: Generative AI for audio

#107

Post by jeffwass »

bottlepalm wrote: post #102

Anyone feel like with the flood of AI generated content there's a risk of the past being 'erased'. Like in 10 years we won't be able to tell if any information from the past is real or fake - sounds, pictures, videos, etc.. Like we need to start cryptographically signing all content now if there's any hope of being able to verify it as 'real' 10 years from now.

I’ve been wondering about this and real video evidence (eg dashcam or cctv) being refuted in court for inability to show it’s not deepfaked.

Re: Open-sourcing AudioCraft: Generative AI for audio

#108

Post by shon »

seydor wrote: post #103
bottlepalm wrote: post #102

Anyone feel like with the flood of AI generated content there's a risk of the past being 'erased'. Like in 10 years we won't be able to tell if any information from the past is real or fake - sounds, pictures, videos, etc.. Like we need to start cryptographically signing all content now if there's any hope of being able to verify it as 'real' 10 years from now.

The past ended in 2022

This ^^

Re: Open-sourcing AudioCraft: Generative AI for audio

#109

Post by dragonwriter »

archontes wrote: post #85
dragonwriter wrote: post #38

The license of the model weights is CC-BY-NC, which is not an open source license. The code is MIT, though.

It's unlikely that model weights can be copyrighted, as they're the result of an automatic process.

> It’s unlikely that model weights can be copyrighted, as they’re the result of an automatic process.

If they can’t for that reason alone, then the model is a mechanical copy of the training set, which may be subject to a (compilation) copyright, and a mechanical copy of a copyright-protected work is still subject to the copyright of the thing of which it is a copy.

OTOH, the choices made beyond the training set and algorithm in any particular training may be sufficient creative input to make it a distinct work with its own copyright, or there may be some other basis for them not being copyright protected. But the mechanical process one alone just moves the point of copyright on the outcome, it doesn’t eliminate it.

Re: Open-sourcing AudioCraft: Generative AI for audio

#110

Post by jsheard »

PaulDavisThe1st wrote: post #25
LewisVerstappen wrote: post #17

Earlier quoted context omitted.

The Record labels are far , far more litigious than the art community.

They can't litigate a person doing this at home, and never redistributing. I suppose they might try, anyway.

Is training a "pirate model" something you'd reasonably be able to do at home though, given the compute requirements? The analogous "image generation at home" is only possible due to a for-profit entity with significant resources choosing to (a) play fast-and-loose with the provenance of their training set and (b) giving away the resulting model for free, if the open source community had to train their models from scratch then as best as I can tell they would still be stuck in the dark ages generating vague goopy abominations.
Post reply on HN