Have you seen image blend / target modification results? Midjourney features are just the tip of the iceberg

There is no way Google is training on every photo that it does this to. It is prohibitively expensive and not necessary to get the results you describe. They can just feed in a set of images, with one to be processed and the others as reference, to an already trained model