Shponglefan wrote on 2024-10-17, 15:04:
vvbee wrote on 2024-10-17, 03:34:
Many human artists fall under that criticism, could even use it to argue that AI is a real artist.
I disagree strongly with the latter simply because of the difference in how AI generates images versus how a real artist does it. In the latter case, artists typically use underlying structure to build up an image in layers. In contrast, a typical AI is just copying the "surface" of an image based on an LLM (which is really just fancy probability modelling).
I'd agree with that. an artist (like most people) have a model of the world. we know that placing a cup on a table means it will rest there and may cast a shadow in relation to the source of light.
A generator 'knows' nothing, it simply 'looks' at what has gone before and uses that with some algorithmic weighting of color/pixel choices, mixing of sources and some randomness to produce something according to various inputs
that's what makes it amazing to me, that you can say "a person on a beach" and from that it can synthesise a superficially reasonable image showing a distinct person on what looks like a beach, for instance, without the generator needing to understand (in any human sense) the concepts of 'person', 'beach' or any intuitive knowledge of up/down, gravity or light at all.
the more often a word is 'image associated' the better the representation; very famous and much pictured objects like certain buildings, people and so on are well represented. apparently 1990's computers weren't photographed enough though, as the range depicted seems more limited and 'wrong' somehow.
while there will undoubtedly be improvements i dont think they will happen in a linear fashion, i.e. the improvements in the last 3 years won't be equalled in the next 3, from now its just refinements and more and more data