That's not what I meant, guess I should have given an example. So if we take this picture
I had to put in the description that there is a staircase or it would morph to something unknown. Also I had to describe those trees as cypresses trees - otherwise they would become dead trees or shadows. This is the type of the description I need: what and where.
This is aboslutely useless and makes it worse:
" the air thick with the scent of oil and smoke "
"The entrance to the Serpentine Quarter" - AI doesnt know what is "Serpentine Quarter"
" The image also captures various smaller details " - this exact text makes it worse.
It should look something like this: An old worn building on the left with a big neon sign saying "Virgin's Hip". Two men in a futuristic armor are standing at the entrance. They have glowing red eyes and hold an energy spear with the glowing tip.
And so on.
Anyway I've still generated some images with the current prompt:
Some details were lost bc of bad prompt + the whole quality is not great bc a second (and maybe even 3d) pass needed. Basically the whole idea is turn the original image into something that AI made and then improve on it.
Or... We can generate from 0:
Results are not great bc the prompt is filled with the details that AI doesn't understand - this is why it can't follow the description properly. The longer the description - the less weight each detail has. Add something that AI can't understand how to draw and u get the results above.
Even so... IMO all of them are better than the original.