LemonDrop 
C++ Crazed
[@Cosmicscribbles](/images/3024337#comment_10371898)
That is not how AI works. Learn how AI works first and maybe you'll have a better appreciation for what it is doing. In short though, AI learns to generalize various constructs of an image on every level, from the overall composition (e.g. what arrangement of different high level concepts will create a "pony") down to the finest detail (stuff like texture, shading, lines, curves, etc). It is quite literally learning in a way very similar to how humans conceptualize things which is one of the one of the goals of modern machine learning in a way as things that see the world in a fuzzy way similar to how we do are quite useful for automating various workflows (e.g. an AI that can recognize images can take the place of a human content moderator or something like that). AI obviously is different from us and especially for art it generates art in a much different way (though do note stroke-based AI art is a thing, it's just a really stupid way for a computer to go about making art when brushes and strokes like that are more a necessity due to limited human/computer interaction technology like drawing tablets vs hooking your brain up to a computer directly), but it is not just taking images and pasting them together, it is learning the fundamental aspects of what make up images and art and that is provable in its generalization ability. You can say its inspiration is a tad lacking however since once learning these concepts you are in a way just interpolating around in a huge high-dimensional space of all the things that make up "art", but it still is able to be creative despite that as much of that space is unexplored, and humans do the same thing a lot of the time as well to discover new combinations of existing ideas that work well (it's hard to say how much human inspiration truly comes from some magical complex brain BS vs how much is us just mimicking things in the environment and other content we've seen).
As for signatures that's really just a limitation of the model. It is finite sized and can only encode so much information in it, and a full understanding of language and writing and even a concept of an "identity" are beyond its capabilities (when its primarily focused on artwork not writing), so you'll either get just jumbled attempts at things that look like a mishmash of signature styles (since there's virtually no stylistic correlation between most dataset labels and the style or content of a signature) or reproduction of a specific artist's signature (when asking for something in a specific artist's style since the signature if drawn in the same way on each image will be highly correlated with that label and its near-identical reproduction usually will train the AI to focus on memorizing it as it doesn't really understand its purpose and thinks its just a critical feature of the art).
Tldr weird signature stuff an undesirable artifact of the training/generation process but also a minor one not worth a ton of time fixing when there's bigger things to be working on (like actually developing the model's artistic skills). You can expect future models though to remove signatures from their datasets to prevent the AI from thinking they have any importance in art (this is already done for watermarks for similar reasons).
That is not how AI works. Learn how AI works first and maybe you'll have a better appreciation for what it is doing. In short though, AI learns to generalize various constructs of an image on every level, from the overall composition (e.g. what arrangement of different high level concepts will create a "pony") down to the finest detail (stuff like texture, shading, lines, curves, etc). It is quite literally learning in a way very similar to how humans conceptualize things which is one of the one of the goals of modern machine learning in a way as things that see the world in a fuzzy way similar to how we do are quite useful for automating various workflows (e.g. an AI that can recognize images can take the place of a human content moderator or something like that). AI obviously is different from us and especially for art it generates art in a much different way (though do note stroke-based AI art is a thing, it's just a really stupid way for a computer to go about making art when brushes and strokes like that are more a necessity due to limited human/computer interaction technology like drawing tablets vs hooking your brain up to a computer directly), but it is not just taking images and pasting them together, it is learning the fundamental aspects of what make up images and art and that is provable in its generalization ability. You can say its inspiration is a tad lacking however since once learning these concepts you are in a way just interpolating around in a huge high-dimensional space of all the things that make up "art", but it still is able to be creative despite that as much of that space is unexplored, and humans do the same thing a lot of the time as well to discover new combinations of existing ideas that work well (it's hard to say how much human inspiration truly comes from some magical complex brain BS vs how much is us just mimicking things in the environment and other content we've seen).
As for signatures that's really just a limitation of the model. It is finite sized and can only encode so much information in it, and a full understanding of language and writing and even a concept of an "identity" are beyond its capabilities (when its primarily focused on artwork not writing), so you'll either get just jumbled attempts at things that look like a mishmash of signature styles (since there's virtually no stylistic correlation between most dataset labels and the style or content of a signature) or reproduction of a specific artist's signature (when asking for something in a specific artist's style since the signature if drawn in the same way on each image will be highly correlated with that label and its near-identical reproduction usually will train the AI to focus on memorizing it as it doesn't really understand its purpose and thinks its just a critical feature of the art).
Tldr weird signature stuff an undesirable artifact of the training/generation process but also a minor one not worth a ton of time fixing when there's bigger things to be working on (like actually developing the model's artistic skills). You can expect future models though to remove signatures from their datasets to prevent the AI from thinking they have any importance in art (this is already done for watermarks for similar reasons).