>>106197595
>>106197595
You can't extract "originals" from weights. They are statistical weights on patterns about edges, textures, colours, composition which is learned/created/trained from real examples, but those real examples are not used after training. You're also confusing previous llms to diffusion models (image generators/art argument. 2 different things. Llms can output verbatims when it is common in training, they predict the next token. Llms are autoregressive transformer models which can in fact memorise pieces of strings verbatim, and when there'sa lot of them, reconstruct them. Stable diffusion based models are diffusion models, they don't do this, images are generated indirectly and not from stored sequences, you can get in some cases get something close to an original which can happen particularly with overfit poorly trained models. The mechanism is still indirect and it's more akin to someone painting from memory vs tracing over something. People do the same thing too, it crops up from time to time where someone has recreated another work uncannily close.
>unlicensed use
No, the model isn't not using anything unlicensed. The training used something unlicensed, 2 different things. When the model learns statistical patterns and relationships in anatomy, the training data may have been unlicensed, but the generic concepts of human body anatomical relationships etc do not belong to them, let alone in a statistical weights.
>Star Trek IP
They won part of the case and the reason was "potentially harmful to Star Trek's market"
That is the real issue with everything here. It doesn't truly matter if it's infringement on anything or not, the ultimate fundamental desire to opposition is about control of the market, it's more of a "don't you dare take market share away from us" type thing. If there's a legal avenue to cripple competition however small, then it's taken
