Why is stability getting all the ire? What about midjourney and dall-e and other systems, or even chatgpt? Does stability just seem easier to fight than the other companies?
Stable diffusion uses an open dataset (latest being LAION-5B I believe) so perhaps it's just the easiest to prove which images were definitely used in the training?
Then once the fight is rolling against one adversary, it seems intuitive to me other groups would pile on to increase odds of success.
So, perhaps it is just a matter of time until all of these major AI working groups are targeted this way, depending on how these cases shake out.
If an AI can’t legally be the author of a patent, can an AI legally be the author of a work - and by extension can it violate copyright?
If a machine can copy images, is the machine, the manufacturer or the user liable?
It seems to me that courts have leaned in favour of the argument that AI is just a tool, and the person using the tool (rather than the manufacturer) is liable.
For what it's worth, the argument made by another notable stable diffusion lawsuit [1] is that diffusion is basically a form of compression no different than an MP3 or a JPEG. They've clearly put a lot of effort into materials to explain how diffusion works, in order to attempt to convince a courtroom of that.
I think they've got a decent shot at it, personally, but who knows. I can see it, but it is hard for me to look at the insane stuff these things generate and try to think "that's just a compression of existing works and not something new."
These will certainly be interesting precedents either way.
It's not just a compression of existing works. But when it can retrieve near-verbatim copies of existing works with a textual description of them? It is a compression of existing works and that's the problem.
> diffusion is basically a form of compression no different than an MP3 or a JPEG
Well that’s going to be contested. AI are similar to compression algorithms, not compressed data.
Compressed data would record the difference between the AI prediction and the data you’re trying to compress. You need both to reproduce the original image.
"Defendant would have you believe that when you pick and choose from a palette of colors and you remix them, you get only a muddled grey. That is not how this technology works."
9 comments
[ 3.2 ms ] story [ 34.0 ms ] threadThen once the fight is rolling against one adversary, it seems intuitive to me other groups would pile on to increase odds of success.
So, perhaps it is just a matter of time until all of these major AI working groups are targeted this way, depending on how these cases shake out.
If a machine can copy images, is the machine, the manufacturer or the user liable?
It seems to me that courts have leaned in favour of the argument that AI is just a tool, and the person using the tool (rather than the manufacturer) is liable.
I think they've got a decent shot at it, personally, but who knows. I can see it, but it is hard for me to look at the insane stuff these things generate and try to think "that's just a compression of existing works and not something new."
These will certainly be interesting precedents either way.
[1]: https://stablediffusionlitigation.com/
Well that’s going to be contested. AI are similar to compression algorithms, not compressed data.
Compressed data would record the difference between the AI prediction and the data you’re trying to compress. You need both to reproduce the original image.
It will happily produce endless pages of the Wab Si Jrbl, filled with letter-like hieroglyphs and pictures of traffic or almost-faces.
Not surprised you can do the same thing with a watermark.
Getty Images is suing the creators of Stable Diffusion - https://news.ycombinator.com/item?id=34411187 - Jan 2023 (83 comments)