Working with AI image generators is both addictive and frustrating. You can get pretty pictures, but it takes persistence. It’s like working with an idiot savant, emphasis on “idiot.” If a model can misinterpret your prompt, it will. It generally takes between six and twelve tries to get a satisfactory image, and the final result is always a compromise.
I mostly play at NightCafé. Sometimes I take part in the “challenges,” in which participants submit pictures they “made” and vote on the entrants. Consequently, I need to quickly evaluate a lot of pictures, too many of which look exactly like other pictures. I’ve developed some criteria for scoring:
Catches my attention: +1
Holds my attention: +2
Is a generic headshot: -1
Includes a kitten: -1
Includes a fox: -1
Includes a dragon: -1
Includes fairies: -1
Includes pointy ears: -1
Includes angels: -1
Includes Batman: -1
Includes Wonder Woman: -1
Includes Catwoman: +1
Is a generic landscape: -1
Is a generic cityscape: -1
Is a generic dystopian cityscape: -10
Features a knight in armor standing on a precipice looking out over a vast wilderness under a dramatic sky: -1
Includes islands floating in the sky: -1
Includes women in boob-plate armor: -1
Is imitation Klimt: -1
Is imitation Art Nouveau: -1
Is imitation Thomas Kinkade: -10
Relies on vulgar humor: -10
Is intentionally grotesque: -10
Includes Donald Trump: -10
Has a message: -10
Has a political message: -100
The astute will notice that many of my own pictures would be down-voted by these criteria. Don’t ever accuse me of consistency.
