Upscaling existing images works much better because as I understand it AI image generation is basically trained by blurring/adding noise to images that are tagged with descriptions and concepts and having the AI try to undo this until it gets close enough to the original image, because bafflingly that makes something that can take a description and concept tags and unblur an image that's pure noise to produce something that could conceivably have those descriptions and tags. When you've got an image that's just kind of noisy or blurry, like one that's been crudely upscaled, it's very good at sharpening it although it still tends to look off.

This just looks like they prompted a generator with "eagle and American flag flying over wheat field at sunset" and grabbed the first output that looked kind of like a flag, which is an impressive level of laziness and grifting from their support staff.