search

borisdayma / dalle-mini

DALL·E Mini - Generate images from a text prompt
https://www.craiyon.com
Apache License 2.0
14.74k stars 1.21k forks source link

Copyright #196

Open dza6549 opened 2 years ago

dza6549 commented 2 years ago

Hello, thankyou for this work and making it available on huggingface. Who owns the copyright to images produced by the dalle-mini demo? Can I sell images produced by the huggingface dalle-mini demo using my own original text prompts? Thank you again, this is a lot of fun and a great way to spend my hours :)

download download download download download

dillfrescott commented 2 years ago

Answered on hugging btw but I'm not a say all end all or anything so its just my insight on the matter :)

dillfrescott commented 2 years ago

Btw the site isn't even loading for me anymore nor is it loading for my friend who had a tab open all night. Somethings broke with it. Does it work for you?

dza6549 commented 2 years ago

Left a reply to you on the facehug.

Yes it times out often, not sure how they do their VMs, assuming VMs. I just try again and I seem to get into the queue.

And thanks for your insight and taking the time to comment. I guess this is a new frontier and no-one is going to have a definitive legal answer until the lawyer comes knocking on my door.

Would be nice to know what the dev says though.

Aaron-G-42 commented 2 years ago

I'm not an attorney and this is not legal advice =)

If you google around for the general application of open (or closed) source licensing as applied to the output of said applications... the consensus seems pretty clear: there are no license restrictions on output.

For example this conversation here: https://opensource.stackexchange.com/questions/8058/does-the-gpl-doesnt-cover-the-output-of-a-program-also-apply-if-the-output-is

In this case, being trained on copyrighted images, IMHO is irrelevant. In the same way that looking at somebody's art for inspiration does not constitute copyright violation.

dillfrescott commented 2 years ago

Also, I might add, if the images it produces are unique and non reproduceable... How would someone be able to tell if it came from the ai or you just drew an artwork yourself that looks like it could have come from dalle mini?

I'd say ur fine!

juniperfdel commented 2 years ago

Not legal advice, but in the US based on my reading of this descion by the Copyright Review Board of the US Copyright Office; nobody can copyright any art created by AI as AI are not humans.

Edit: A better article mentions its still in a grey area as the amount of work which you, as a human, put into it to become copyrightable hasnt been tested yet, it just has to be greater than zero. That said, I suggest reading it yourself

Actual Decision : https://www.copyright.gov/rulings-filings/review-board/docs/a-recent-entrance-to-paradise.pdf Article: https://www.theverge.com/2022/2/21/22944335/us-copyright-office-reject-ai-generated-art-recent-entrance-to-paradise

dillfrescott commented 2 years ago

I'm planning on making some artwork using another similar project and putting the results on Instagram for others to enjoy. I don't think there's anything wrong with that.

dza6549 commented 2 years ago

Thank you @gregoryfdel for your informative. yet not legal advice.

I'm pickin' the implication here is that anyone else can claim my images as their own as the image was produced by an AI model if I do full disclosure. Of course, as others have commented "who would know", yet based on my previous publishing experience, determining ownership used to be a relatively clear cut determination. Opaque AI models based on mashups of millions of inputs is difficult I guess.

@DillFrescott "anything wrong" doesn't hold water in a copyright case. Also I believe the T&C of Meta properties such as Instagram conveys the rights to Meta?

MichaelMcKinnon commented 2 years ago

My guess is that if you produce something using this software and publish it, you have the ownership of the work. Like any other software. You chose the text prompt and selected the semi-randomly generated image to publish. It gets a little murkier when the AI is hosted on someone else's server, but I can't think of any cloud service where you don't own content you produce. Not a lawyer.

juniperfdel commented 2 years ago

My guess is that if you produce something using this software and publish it, you have the ownership of the work. Like any other software. You chose the text prompt and selected the semi-randomly generated image to publish. It gets a little murkier when the AI is hosted on someone else's server, but I can't think of any cloud service where you don't own content you produce. Not a lawyer.

Not a lawyer, and definitely not legal advice.

I dont think inputting text into the model would qualify as enough effort to make these images copyrightable because of two reasons: the act of writing prompts could be done by a computer and there isnt a gurantee that everytime the model is run with a given prompt that there will be a different image.

To expand a bit, I could write a program which inputs all possible combinations of 10 words or less into this model and do that tens of trillions of times each. If this did give me copyright, then there are two outcomes that I can see, either: all outputs of all prompts of less than 10 words done by everyone would then need to be checked against my tens of trillions of images for that prompt to ensure they weren't infriging on my copyright, or the program starts outputting the same images enventually and I have copyrighted all possible outputs of all prompts of 10 words or less.

That being said, I do absolutely believe that this is still a very grey area legally and any assumptions about the extent of what "produce" or "enough effort" entails would need to be fought in court or made law by congress to be legally defined at this point, even the assumption I made above.