Midjourney Learned About Rain, and Also About Men
A prompt about the London Eye in the rain reveals both the gap between the image generators and something quietly depressing about what they absorbed.
The quality gap observation was accurate for 2023. DALL-E 3 and DALL-E 4 closed it substantially. The more interesting call was noticing the bias patterns in generated imagery — that thread kept pulling through 2024 and 2025 as every model had its own version of the same problem.
Ran the same prompt through Midjourney and DALL-E — couples and families at the London Eye in the rain — and it's not a close contest. Midjourney does something with light on wet pavement, with the blur of a ferris wheel behind mist, with the specific gray of an English April, that DALL-E just flatly fails to do. DALL-E gives you a postcard. Midjourney gives you a memory of a trip you didn't take.
That part isn't surprising anymore.
What caught me was something else. In almost every Midjourney output, the men have the umbrellas. The women and children are damp. The men are dry. Nobody asked for this. The prompt said nothing about umbrella distribution. The model just knows — from the ten billion images it metabolized — that this is what happens when it rains and there are couples and families.
It rendered the scene accurately. That's the whole joke.
DALL-E, for what it's worth, distributed the umbrellas more equitably. I don't know if that's a values decision or just that its training data was different or that it's worse at picking up on ambient social texture. Probably the last one.
Counterpoints
Push back, extend the argument, or sharpen it. New counterpoints go through review before they show up here.
No approved counterpoints yet.