There's some fun bouba/kiki stuff in these diffusion models. "Doddaparaavalambi" is a nonsense word in Kannada -- valid construction but not a name or a word in use. Base SD1.5 treats it very differently than say DALL-E 

"A photo of doddaparaavalambi", base SD1.5 https://i.imgur.com/viu35qa.jpg

"A photo of doddaparaavalambi", base SD1.5 https://i.imgur.com/M1KT5Ym.jpg

"A photo of doddaparaavalambi", DALL-E https://i.imgur.com/gXKK1ly.jpg

Both models are producing outputs that are plausibly images from the subcontinent. They look like people and places I've known or been. But the words shouldn't exist in the training set

DALL-E seems to be focusing on "dodda" ("big"), producing mountains. What if we swap that one piece out? The opposite of "dodda" is "chikka" (small)

"A photo of chikkaparaavalambi", DALL-E https://i.imgur.com/QSxlMgM.jpg