Thursday, November 09, 2023

Mary, don't you weep

My first experiment with generative AI that transforms words into images, something called Firefly which comes with Adobe. For any prompt it conjures up four images, which you can then tweak by choosing art or photo, a range of styles and moods, and of course altering the wording of the prompt. As with ChatGPT's glib words, the images can be pretty cheesy, and the four often offer a demo-graphically balanced set of images (as for "tree teachers" above).
But some of the images can be quite lovely (until you fully process that these are fake photo-graphs), like these gene-rated from a few sen-tences about torrey pine trees. But the coolest thing is that you can add to all this an image of your own for "reference" - colors and textures, I gather. Still, there be monsters. As I played around with more and less specific and more and less obscure prompts, I decided to see if it could generate something like an image I'd recently enjoyed at the Met, of Jesus, his mother Mary, her mother Anna, and her mother, too. Though I just put in Jesus, Mary and Anna this was a bridge too far; instead of Anna I kept getting Joseph. For fun, I added a photo I took of a voluptuous tree spirit I took at the "Tree & Serpent" show of early Buddhist art. 
And got these remarkable painted statues of holy families (the Indian "reference" I sent was just carved red sandstone, not painted). But wait. This Mary is a bearded! What's confounding isn't that Firefly got things so wrong but that it's so convincingly wrong.