If you described the original image (or made CLIP do it) and fed the description to txt2img generation, then you'd probably be fine.
If you described the original image (or made CLIP do it) and fed the description to txt2img generation, then you'd probably be fine.
I think being able to selectively optimize for "magic" random seeds in the diffusion algorithm will be kind of critically important here. Different seeds can produce very different images given the same prompt.
What if an optimizer can find a seed+prompt combos that are just as good as cloning as img2img?
If you instead don't want to clone an image, you can just extract the CLIP image embeddings from it and use them to condition a generative model like Dalle 2, Midjourney or Karlo (open source). The CLIP embeddings extract really well the semantic meaning of an image.
Something about going image -> prompt -> image feels like it "subverts" this somehow, even if the prompt is hyper-optimized to recreate the original image.
Obviously, this is just my feel/impression of it, the real test is how a jury feels about it.
The next few years will be really interesting in exposing that this is a really massive gray area.
That's only because you understand the algorithm in the first example.
At a jury trial most, if not all, of the jurors will find both methods equally opaque and, I believe, will treat them as equivalent.
I can start with van Gogh's Starry Night, get the prompt "starry night, van Gogh" and get Starry Night back.
I'm using Starry Night as an example because Stable Diffusion consistently reproduces it, sometimes with the original frame, even with vaguely related prompts.
I'd say the jury will be making the right decision. Especially if the original image was part of the training set.
A tool that specifically advertises itself at getting around copyright is going to have a very hard time in court…
The vectors are much more opaque because there's no straightforward human equivalent.
Describing to a human.
Describing it to stable Diffusion is very different. If you ask for a starry night, you get van Gogh's starry night, sometime with the original frame.
A prompt can easily trigger it to make a derivative work from something in its training set, and it often does.
Any competent expert will tell that to the jury and easily demonstrate it again and again.
Trying to explain how that "isn't really a copy" by explaining AI concepts isn't going to win the day, not when they can SEE the copy.
On the other hand, if I was to take an existing image had and then used that to transform it, it is a derivative work. If this was my image created with a camera rather than mspaint, its still a derivative work but I, as the creator of the original, can create derivative works of my own (I do it all the time).
The idea that I could use someone else's image and run it through img2img to make it "not a copy" misses the point of copyright and the exclusive right of the copyright holder.
https://www.copyright.gov/circs/circ14.pdf
> Only the owner of copyright in a work has the right to prepare, or to authorize someone else to create, an adaptation of that work. The owner of a copyright is generally the author or someone who has obtained the exclusive rights from the author. In any case where a copyrighted work is used without the permission of the copyright owner, copyright protection will not extend to any part of the work in which such material has been used unlawfully. The unauthorized adaptation of a work may constitute copyright infringement.
---
With starting from text, a description of the Mona Lisa is difficult to get generative art to create without significant effort to get the Mona Lisa (without saying "I want that image").
The question is then "is a description of the Mona Lisa a derivative work" and that is likely sufficiently far away from the original to be "no."