Spaces:
Running
DALL•E mini biases (in the non-pejorative sense)
DALL•E mini seems to have some interesting biases picked up from its dataset that pop to the surface in the absence of other input.
For example, a blank or single space tends to produce images of individual women from southeast Asia wearing what look like saris. Other results tend to fall into categories like a flat monochromatic color, or a single color circle on a blank background.
Similar output is produced by a period, colon, semicolon, or a comma. Occasionally a southeast Asian man puts in an appearance.
$ of course produces images of money, but other punctuation has its own twists. An asterisk (ie. *) produces anime-like headshots. An apostrophe (ie. ') produces animals.
And over and over in the results we see circles, circular emblems, circular icons or logo-like constructions (usually meaningless, but occasionally edging toward something recognizable).
Any ideas where these particular associations may be coming from? I imagine there must be a lot of images of people from southeast Asia that are unlabeled in the corpus to associate a blank prompt with those results, for example.
Wishing I could upvote this discussion. Still no answers... ;(