Regarding AI-generated Imagery

• The “I can’t believe it’s not” tags get used also when the style resemblance isn’t intentional. And, consider the reverse, where someone meant to emulate a style but didn’t do so good a job of it - would style tags be used based on intention or on result, and if the latter, how would the judgement be defined? (The “I can’t believe it’s not” tags feel a lot more lighthearted to me and don’t seem troublesome to leave unformalized.)
• What you get from AI image generators may or may not have any resemblance to speak of to the artists used in a prompt:
  • The influence on the image might be as little as “the named artist tends to work in grayscale, so the outputs will be grayscale”, for example. Or, “the named artist tends to make paintings with fine detail and shading, so the outputs will be detailed and shaded, not flat-colored”. It definitely wouldn’t be appropriate to attribute style to every name in a prompt.
  • And, depending on the generator used, it might end up emulating an artist that isn’t named at all, if that artist is heavily associated with something else - like if you ask for a cubist picture, you might end up with Pablo Picasso’s style.

• If there are many images that imitate specific styles, such information should be easy for viewers to navigate.
• It should be easy for artists to show off all images based on their style so they can showcase their creative impact. This is only possible if it’s easy for artists to find derived work based on their style.
• I don’t believe viewers are substantially different from people generating images with AI. If people want to generate images of a particular style using AI, then it makes sense that viewers will want to navigate images by style. This is only possible if image styles are tagged appropriately.

• The point is to make the chain of creative contribution more clear. I believe this is a much more important point to address than the classification of styles, which is what style similarity tags would accomplish.
• Whether two styles are actually similar is a judgement call, and that involves a very fuzzy line. Tags should be reasonably black-and-white.

… Nobody is even currently considering the training set authors as potential copyright owners.

• If the resulting image looks very similar to a copyrighted work, then it’s probably considered derivative regardless of the AI that was used to generate it. (Though AIs usually try to avoid this.)
• If an AI is used to run an artist out of business by training off of their works, it is likely the AI’s creation violated the artist’s copyrights.
• The paper’s authors were unsure whether copyright might apply if an AI copied “an authors’ copyrighted expressions (Sobel, 2017), learning from their creative and expressive choices (Bonadio and McDonagh, 2020).” That is, in many ways, the big question when it comes to these works: if a machine emulates a specific artist’s style, especially if the AI is framed as a tool trained off of those works, are they derivative?

• Images generated at random (with no human input for the specific generated images) have no copyright, since no human was really involved in the creative process.
• Images where a substantial amount of creativity comes from the individual (i.e. the tool is like a really fancy Photoshop brush), will likely grant copyright to the individual.
• If a bunch of images are generated, then hand-filtered by a human, that might be enough “creative input” to grant a copyright.

@Cocaine
visually, conceptually, and technologically interesting creations that humans have decided to share with one another