It's basically like AI Minecraft. You know, that project where someone trained a stable diffusion model to predict the next video frame based on the previous frame and the player inputs. Then people started speedrunning it, like by looking down into a lava pool and back up, now they're in the Nether (underground lava dimension). That's how I think dreams are, except you don't even get a full video frame, just one object of focus.
For me the real surprise is the fidelity of the visual I _do_ get. Sounds, sight, and touch feel just like real life (in lucid dreams especially now that I think of it). If the quality of our waking visualization is really meant to be the same as it is in our dreams, for me it is nowhere close.