Does anyone have a good technical-context breakdown for how Midjourney does what it does?

Not my focus, but would love to read through something that connects "This is how it breaks down input/example images, identifies relevant high-level themes and characteristics, then rebuilds those into novel output images."

Specifically around image generation, although I assume there's some text metadata in the background.

F.ex. "What makes bomb shelter images look like bomb shelters?" is a question I could answer, but apparently one Midjourney et al. can do a pretty good job answering too.