Do you see what I see?
Hummingbird defeats AI
So I took this picture of a cloud. To me, it looks like a hummingbird.
Don’t see it? Here’s a line drawing. (I left the watermark - it was not worth paying to remove it.)
Of course, my cloudy friend is missing wings. I’d like him to look more like this:
Unfortunately, I am neither artist nor graphic designer. So I thought I would use this as an opportunity to try AI. The results were… poor. Arguably, that’s my fault for not knowing how to prompt properly. But, ultimately, my experiments demonstrate very limited image understanding by the models.
Let’s explore.
___________
This is what I got from imgtoimg.ai when I prompted with:
make the cloud look like a hummingbird
Some nice hummingbirds, but both images:
While mottled white, are anything but cloudlike,
Have lost the artistic feel of the bird stretching to the right that’s in the original, and
Have made the bird fatter/rounder.
It’s not what I want. I made a few more attempts, but all of them failed on those points.
So, on the ChatGPT!
This time, I tried to compose bottom up. That is, I asked it for a picture of a hummingbird as a starting point, and eventually got this far:
That’s pretty impressive for bottom up, but it’s remarkably like the top-down one from imgtoimg. Which is to say, it’s still not right. You can read the whole session here.
Maybe Sora can help. Here’s a bottom up…
I give up.
Conclusion
I am very impressed with how well the AI got to images. I am disappointed, albeit not surprised, that the images were not what I wanted, lacking both an artisitic eye and the dynamism of a hummingbird in flight. Good enough for mock-ups, but I expect skilled graphic artists to be employed for a while.