OpenAI’s latest image AI system, DALL-E 2, is picking up in popularity. It allows users to create realistic images from natural language prompts. Alan Resnick decided to use it but with a unique idea:
Every single image in this video was created by the AI called DALL·E 2 from OpenAI from just one initial text prompt “A bad photo”.
The first photo of the blurry hand is the “bad photo” it came up with.
From there all I did was request variations, and choose the variant I liked the most, then ask more more variants, repeat. I never gave it any more text prompts. I would subtly try to “guide” it toward aesthetics I like. I also tried to pick variants that I thought would change things.
I’m really shocked by the vast range of styles and aesthetics it’s capable of. Especially without any explicit direction. Some of the images are ugly but I love many of these, I wish I had made them.
The results are remarkable and almost evolutionary in nature. It starts with a hand and varies between the real and the hyperreal, meandering through surreal and abstract realms with each iteration. Image AI makes me physically uncomfortable (if there’s a legit phobia for these sorts of images, I have a mild-to-severe case of that). I wince a little watching through but was fascinated overall. I still think this technology is harmful but in this instance, it was interesting to see.