>>36492871>the fact that it can indeed understand "prose" prompts like "many thick cables and conduit in the background" or "thin antlers with leaves on them",I think this might be just a novelAI thing? I know it does some sort of language processing on prompts because it accepts far longer prompts than the stable diffusion webUI does. I tried throwing in the full set of tags you used on the first marine there and (aside from getting back something that looks nothing like yours), I also get this:
Warning: too many input tokens; some (110) have been truncated:
twintails , red hair , yellow eyes , red eyes , red hair ribbons , roboco - san cables , plug , office chair , medium breasts , mechanical arms mechanical legs , impossible leotard , lanyard , id card , cyberpunk , skirt suit , many thick gauge cables behind chair looking at viewer , android , cyborg , humanoid robot , kusanagi motoko , cyberpunk , dolby vision , make my eyeballs cum , red eyes , yellow eyes , heterochromia , gloves , mechanical hands , latex bodysuit , navel , midriff
which is more or less the entire prompt, so no surprise the result looks nothing like marine.
picrel is
>(((hololive, houshou marine))), houshou marine outside on a starry night looking up at the sky so I guess even stable diffusion has some language processing but I'm not sure if there's actually any understanding going on here or if it's just ctrl-f on danbooru tags. If I pick someone weaker of a tag than marine there's about a 50% change that the image is just a starry sky with no characters anywhere in it
