Interesting. How do you see the Getty v stability lawsuit going? That looks much worse for stability. Do you think they will just settle and stability will pay them some licensing fee?
back
1 comments
Getty has a much stronger case, given that warped versions of the Getty logo have shown up in a number of SD-generated images, so it's obvious that there was impermissible copying.
I'm not sure Stability will agree to a licensing fee, since part of the rationale for the last version of SD was to remove the infringing images from their training sets going forward.
> warped versions of the Getty logo have shown up in a number of SD-generated images.
If you create art that has a Pepsi logo on a depicted vending machine etc Pepsi has no copyright claim on your art does it? All it shows is the art was made with the knowledge of the logo and the logo was included as an element inside the art.
When logos are shown in a context that may cause confusion (about who made a product etc) there may be trademark infringement but trademark infringement is not being claimed here so why would warped logos matter?
The issue at hand isn't actually directly about how the output images contained the Getty logo; the lawsuit isn't saying "you're showing our logo on your output, which isn't a Getty image, and we take issue with that". It's whether Getty images can be ingested into the training set without consent or compensation to Getty.
The reason the distorted logos matter is because they make it much more difficult to claim that Getty images were not ingested and used for training — if they weren't, then how come the outputs have those logos? And similarly, they make it much more difficult to claim that these source images were only used as "inspiration" for the generative algorithm and thus fall under fair use — if they're only used for "inspiration", how come they generate/copy easily-recognizable parts of the original images (i.e. the logo) as-is?
> they make it much more difficult to claim that Getty images were not ingested and used for training
Was that claim put forth? Why then does making this difficult matter?
> they make it much more difficult to claim that these source images were only used as "inspiration" for the generative algorithm and thus fall under fair use — if they're only used for "inspiration", how come they generate/copy easily-recognizable parts of the original images
If artists created works of art containing warped logos etc as elements in their art would they be infringing copyright because of these warped logos? But if an artists uses a computer to create the same art instead of real paint that becomes infringement? Because copyright depends on the method of production not just the produced result?
So I'm only guessing here but my thought is because a Pepsi logo just indicates that the input image was an ad, a Getty Images logo means that the input images were owned by Getty and likely used without their consent.
We know input images were used without consent. That’s not in question, it just also isn’t all that relevant.
But a watermark showing up indicates clear, undeniable ownership of those images by a massive company that seems to make most of their money selling licensing to their images.
> so it's obvious that there was impermissible copying.
Impermissible is a legal conclusion that is not the least bit obvious.
I've never seen a getty logo in any of my generated images so I suspect that they used img2img or some weird prompt.
In the early stuffs, Stable Diffusion 1 (not XL 1!) and such, if you prompted for stock photo style images, you regularly (~30%) got something resembling the Getty stock photo watermark in the lower right.
Was quite annoying but adding "public domain, Creative Commons" to the prompt usually got rid of it (the model knows that public domain images have no watermark :-). Since SD2.0 I haven't seen this happening at all.
I've generated thousands of images using SD 1.5 based models and I've never got the getty watermark. That makes my think that some dishonest lawyer used img2img.
I can't find a publicly accessible SD1 model, but CompVis LatentDiffusion also uses the LAION dataset.
(1) go to https://huggingface.co/spaces/multimodalart/latentdiffusion
(2) enter "businesspeople arguing in the boardroom" as the prompt
(3) generate a couple images
(4) find things which look a lot lot lot like watermarks of stock photo companies
This is what I get (first try!): https://ibb.co/1620jT8
Haha, that model is ancient. I doubt that anyone is using it now. Can they reproduce that problem in any recent model?
Hmmm. What do you think the outcome will be then?