I like taking pictures of reflections. That includes photographing puddles in the street that reflect the city above. There is a photo I know I took, not too far from my office, but I cannot find it anymore.
I also became curious how I can get DALL-E to reproduce these pictures, if I can at all. One day I did this:
Me: Create a photo of a close-up of a puddle that reflects the clear sky with a few clouds.
This image tells a lot about the messy relationship between DALL-E and reality. If the puddle is more than just a single drop of water, it cannot have the curved, lens-like surface we see above. In a way, it’s truly Matrixesque: in a computer system, rules of physics can be bent, some can even be broken.
Let’s move closer to reality, methought, and said this:
Me: Can you create a variant where the puddle is larger and is in a cobblestone street?
DALL-E also doesn’t know the difference between intentional and accidental. It doesn’t know that a puddle isn’t built (see the first pic) but formed by accident (rain, that is)?
On the photos I take of water surfaces, I like moving as close as I can, so that most of the photo is taken up by the water. I also tried this with DALL-E:
Me: Can you create one more variant where we are closer to the puddle?
This didn’t make a whole lot of difference: DALL-E apparently doesn’t know what “closer” means. Also, can you see that the second puddle looks entirely built on purpose?
Finally, we arrive at the most peculiar quirk of the post:
Me: Can you re-create the first image as a pencil drawing?
Why does DALL-E think that a pencil drawing must include the pencil, too?
thirdtower.ai 