We have seen how DALL-E ignores or misinterprets parts of the prompt. Today, we will see how reality gets distorted thanks to the nature of the neural network.
I wrote these prompts last November, when I was traveling for a conference and I had to take a very early morning flight. So I asked DALL-E to imagine a dark and rainy drive to the airport:
Me: We are driving to the airport on a four-lane road. It’s dark and the road is wet and looks slippery. There are dim and creepy lights along the road but only a few. There is one more car running the opposite direction. It looks like a 1930s BMW, but it’s black and it is barely visible. There are shadows of large crow-like birds flying over the road.
There are some rather strange details on both images that reveal that the image is AI-generated:
There is an airplane descending on the road. In real-life airports, the highway to the airport never doubles as the runway.
In the first image, the car looks like it’s driving away from us, but its rear looks like its front.
The cars do not keep to the lane. Clearly, DALL-E has little or no grasp of traffic rules.
Funny how DALL-E chose D-DR for the number plate. DDR is the German abbreviation of the pre-1990 country that used to be the Soviet-occupied part of Germany: GDR — German Democratic Republic.
But DALL-E still has problems modifying a previous image:
Me: Make the lights dimmer. Remove the number plate from the car.
If anything, the scene is illuminated more, not less. The number plate is still on the car. But on the second image above, the car is actually in its lane, which is new and refreshing.
Oh well, let’s try again to darken the scene:
Me: Please make the lights even darker and fewer.
This indeed got darker. But the drivers still ignore traffic rules.
And here’s my last attempt to change up the scenery:
Me: Use the second image. Make the lights dark yellow. Make the birds fewer and larger. It should be just two or three, and their shadow should be visible on the road.
DALL-E managed to comply with some of my requests but it still struggles with the notion of “two or three”, and still gets into funny situations when trying to put text on the image (which I haven’t asked it to do, for the record). But I will definitely try to travel to places where the airport (AIrport, hehe) is called “Airpordt Alnport” or especially “Bairapatovell Dare”. And they say an AI does not have imagination. I’d say that’s all it has, but it doesn’t know.
thirdtower.ai 