Conversation

#1: "anime girl in a yellow sundress standing on a hill at twilight" #2: "screenshot of an anime girl in a yellow sundress standing on a hill at twilight from sailor moon crystal (2014)"
Image
Image
4
5
40
#3: "screenshot of an anime girl in a yellow sundress looking soulfully at the viewer, standing on a hill at twilight, from the anime sailor moon crystal (2014)"
Image
2
7
#4: "sailor moon in a yellow sundress looking soulfully at the viewer, standing on a hill at twilight, screen capture from the anime sailor moon crystal (2014)"
Image
Image
Image
1
4
"macro 35mm photograph of a woman cosplaying as sailor moon" (mostly sharing because somehow these came out super photorealistic and *also* super zoomed in for some reason? mystified)
Image
4
10
and i'm out of attempts for today. tentatively i have no idea how to get DALL-E 2 to consistently generate good anime-style pictures. based on what some of the bad attempts look like this might be partly an issue of lots of amateur fanart in the training data?
2
7
Replying to
It seems to have trouble keeping track of multiple properties for things, I've seen people get better results by adding each as separate sentences. eg."Sailor moon looking soulfully at the viewer with big anime eyes. Sailor moon in a yellow sundress. Sailor moon standing on..."