Attempting to confuse Dall-E by adjusting the prompt for AI-generated art a few words at a time
I started out with a prompt I saw a friend use that was already pretty lengthy (but sensical), and started changing and adding to until until Dall-E couldn’t manage to produce something that encompassed the whole prompt.
But, spoiler alert, at the very end I rephrase the final result prompt one more time, which results in some art that kinda makes sense!
1. “10,000 hot air balloons, sunrise, desert, cinematic, hyper detailed 8k”
This was the original prompt I saw a friend use. the resulting image is pretty and makes sense.
2. “10,000 hot air balloons, sunrise, eating dessert, cinematic, hyper detailed 8k”
The balloons started to look misshapen, possible a little bit like cupcakes.
3. “10,000 hot air balloons, sunrise, eating dessert cake, cinematic, hyper detailed 8k”
After adding the word “cake” I’m really enjoying the things on the ground that look like a cross between deflated balloons and cakes.
4. “10,000 hot air balloons, sunrise, eating dessert cake, candyland board game, cinematic, hyper detailed 8k”
Adding the mention of candyland, we seem to have lost the “hyper detailed 8k” aspect of the art, although the words are still in the prompt.
5. “10,000 hot air balloons, sunrise, eating dessert cake, candyland board game, Mariokart, cinematic, hyper detailed 8k, award-winning”
Looking like a Candyland themed Mariokart level now.
6. Alternate output from the same prompt (“10,000 hot air balloons, sunrise, eating dessert cake, candyland board game, Mariokart, cinematic, hyper detailed 8k, award-winning”)
I like this style - would definitely enjoy playing a video game like this.
7. Things get confusing by introducing puppies: “10,000 hot air balloons, sunrise, eating dessert cake, candyland board game, Mariokart, cinematic, hyper detailed 8k, award-winning, puppies”
This was when Dall-E really stopped being able to incorporate most of the elements of the prompt in one picture.
8. Alternate output from the same prompt (“10,000 hot air balloons, sunrise, eating dessert cake, candyland board game, Mariokart, cinematic, hyper detailed 8k, award-winning, puppies”)
This version from the same prompt really focused in on the puppy aspect. Not sure where it got the calendar idea from. And it seems to be ignoring all of the stylistic instructions.
9. But order returns by putting the elements into a sentence "10,000 puppies riding in hot air balloons at sunrise while eating dessert cake, over a Candyland board game Mariokart track; cinematic, hyper detailed 8k, award-winning"
It's not exactly right, but definitely a lot better. The final images in this list are all alternative outputs form this prompt, which I view as Dall-E's crowning achievement in this exercise.
10. Alternate output - weird uncanny valley puppies version (""10,000 puppies riding in hot air balloons at sunrise while eating dessert cake, over a Candyland board game Mariokart track; cinematic, hyper detailed 8k, award-winning")
11. Alternate output - army of marshmallow peeps version (""10,000 puppies riding in hot air balloons at sunrise while eating dessert cake, over a Candyland board game Mariokart track; cinematic, hyper detailed 8k, award-winning")
12. My conclusions: Dall-E can handle more elements when they're woven into a story, not just listed out.
Pretty similar to human brains, I would say! If anyone wants to try out a similar exercise with a different starting prompt, tag me in the results! (You may have to sign up for the wait list to get access to Dall-E though: https://openai.com/blog/dall-e/)
No comments.