I took this photo in an upscale interior design shop in Brussels (I promise I didn’t buy anything there, I was just intrigued):
Then I asked ChatGPT to describe it:
“The image shows a small, metallic sculpture of a penguin sitting on a dark wooden surface. The sculpture has a smooth finish with subtle color variations, possibly due to patination. Its form is stylized and abstract, with a rounded body, a pointed beak, and no distinct facial features. Beside the sculpture is a cut-out metal piece that resembles a crown or a stylized flower, also with a patinated finish. In the background, there is a part of a digital screen and a rectangular box made of light brown material. A barcode tag is visible on the surface, suggesting the item might be in a retail or exhibition setting.”
I then asked DALL-E to produce a photorealistic image using the very same description. I got this:
I repeated the exact same prompt four months later:
It’s close enough but far from the exact same image, showing (again) that you simply can’t use generative AI to make a repeatable experiment.
Back to four months ago: I gave in to the usual temptation to start changing things in the image and gradually introduce the absurd. First, let’s try something simple:
Me: Can you change the lighting to light blue?
This was reasonably well done, but we are reminded that there is no changing just one thing about an image.
Let’s then edge away from the laws of physics (although magnets can do this):
Me: Make the penguin hover in the air, above the tabletop.
And now I managed to gave the AI two instructions that were completely ignored:
Me: Turn the penguin upside down. Make the beak emit light.
Also, you can’t make DALL-E listen even if you implore it.
Me: Please do turn the penguin upside down.
I give up. Let’s try something simpler.
Me: Add rain to the image. And me again: Make the rain heavier. Add mist to the bottom.
What follows can also be done by introducing an alternating magnetic field, only DALL-E ignored my instruction again.
Me: Make the penguin look like it is spinning.
But if you tell it to do something completely different, it will catch on to the previous prompt:
Me: Make the scene darker.
The following will make DALL-E forget some of what was done in the conversation:
Me: Please make the tabletop completely flat and shiny so the bird is reflected in the surface.
OK, I think it’s time to finish — but before we do, let’s change the scene up a bit, shall we?
Me: Please replace the bird with a large metal raven.
I don’t know how much the new bird is raven-ish but it sure looks more ravenous than ravishing. And that concludes our story for tonight.
thirdtower.ai 