/ic/ - Artwork/Critique

>>6091406
>Shouldn't it be possible to hard-code some anatomy books into their database that they can reference for correct proportions and correct number of limbs?
No because these AI are using "self learning" algorithms to associate words with pictures. It does not understand concepts like light physics or anatomy outside what is visually observable it cannot comprehend what a human actually is.

It takes the words it was given, then look at a biillion images that come up on google (or whatever databank/search engine it uses) Then compares them together and looks at what colors and values are common on them and their distribution on the area. Sometimes if there are very commonly repeating patterns or something is an exact match it just literally cuts that out (like say an anime eye) and pastes it in, if not then it tries to create an approximation of them. Finally it tries to create a new picture using all those parameters it gathered from the billions of pictures it looked at and tried to photobash them together while simply putting smudged blurs on places it cannot understand or able to find a good enough reference or if there is too hard edge on to be disturbing. Then if there is a feedback on what is good it tries to use those for future patterns.

At its core it really is just an extremely advanced photobashing tool that is capable of creating results just by brute force from the sheer number of material that is available to it and a list of simple parameters that are given to it.