Hacker News new | ask | show | jobs
by croes 622 days ago
If that's the case then they are sampling previous artists and need to pay them.

You can't have your cake and eat it.

2 comments

I don't necessarily disagree, but I don't necessarily agree either.

I think the most compelling reference work right now is probably a collage.

> a piece of art made by sticking various different materials such as photographs and pieces of paper or fabric on to a backing.

And the law there is actually pretty nuanced and tangled: A collage, while not a compilation, may be a collective or a derivative work. A collective work has copyright protection, but a derivative work does not.

So basically, given the current environment I think we're honing in on a super clear and very helpful "sometimes"...

It's about who created the work (book, painting, code, etc). If it's a human, then it's copyrightable. If it's an AI, then it's not. That's the stance the copyright office has been taking, going so far as to grant partial copyrights if a human substantially changes portions of an AI generated work.
> If that's the case then they are sampling previous artists and need to pay them.

Adobe's models are 100% trained on works they own or license the copyright to.

Synthetic datasets are being built in Unreal Engine and with automated photo turn tables.

This argument is going to lose merit soon.

> This argument is going to lose merit soon.

When midjourney and all other models built upon copyrighted work they had no license to disappears (and any models which were in turn trained on the output of Midjourney et.al.), perhaps.

> Adobe's models are 100% trained on works they own or license the copyright to.

I have to ask, was the license broadly interpreted to include AI training, or was that part of the original license? Because the former is what Google is doing right now with their video AI and scraping youtube.

For completeness sake, Adobe does not appear to have taken Google's lead in this.