Continuing to play with generating images from text, I made this album of pictures on Flickr using prompts describing Dinosaurs doing touristy things around London.
I'm interested in what possible biases one might infer about the training sets used with these models. What pictures of dinosaurs and scenes in London have they seen? How else can you test for bias (ie an unconsciously unrepresentative set of images).
But also how would you know that no bias is present, that it's a "good" representation (whatever that means) What is a generic dinosaur, if there is any? What about an image tells you that it's Hyde Park rather than any of the other Royal Parks? Are these pictures any more or less valuable than giving a prompt to a five year old child and asking them to draw it (or to a 55 year old adult for that matter?)
