I've seen very obvious Sarah Michelle Gellar and Kim Kardashian faces at times. Not sure if it's just chance in those cases, or whether like in the Bollywood scenario it's due to the volume of of celebrity faces distorting the algorithm.
I've been playing with how well it recognises IP-specific terms.
It was able to interpolate from 'blood angels space marine' to get all the correct canonical armour and colour scheme without putting any actual bloody angels in the picture (it didn't nail the details of the heraldry, unsurprisingly!), but it didn't understand 'plaguebearer daemon' even when i added 'of nurgle'.
View attachment 301997
There seems too be a very hard threshhold where it ceases to understand the specialised terms, probably due to a lack of training material, and just makes stuff up or ignores items it doesn't recognise. It can understand 'space wolf space marine', but not 'genestealer' or 'baneblade'. I do like the little pink space marines sitting around having a tea party with Barbie though.
I had the same problem trying to get it to generate an image of a Zulu warrior with one of those distinctive pointed-oval cowhide shields. It has a very specific idea of what shape a shield is, and doesn't understand any of the specific terms - isihlangu, ihawu, ingubha, ikhaka or ikhawu (according to google)
View attachment 302079
View attachment 302080
View attachment 302070
View attachment 302076
It understood 'jabba', but not 'hutt', although it did put the stormtrooper in front of a very jabbas-palace building in the first case.