>>2234512The Stable Diffusion upscaler works differently.
In a typical case you don't want the upscaler making things up or adjusting the image, since you just want a higher-resolution version of what you have. But with an already-AI generated image, it makes no difference if that happens, so it's what the SD upscaler does. It's not "fake" details since it's coming from the same source as the original. The upscaler actually does make a higher-resolution version of the image, with real details generated to fill it in, to the extent that any details are "real" in an AI image.
Of course, it still has all the flaws of AI-generated images in general, and it's also not perfect, but in enough cases to be worth it, an SD-upscale of an SD image is at least as good as if it had been created in that resolution to begin with. And in many cases better, since re-running the image can tidy up flaws with fiddly areas like hands and faces.
>>2235282Assuming the Automatic1111 UI:
Generate in the largest resolution of appropriate ratio you can handle, with the Highres fix on. Then run it through the SD upscaler script (not the extras tab, but the one at the bottom of the img2img tab). It may take multiple runs, with different prompts (since it works in chunks, you might need to do some runs with a prompt that emphasizes the foreground, and others that emphasize the background, etc), and you'll very likely need to stitch multiple versions together. And it'll take experimenting with different samplers, pre-upscalers, etc.
>>2235407Because that takes a lot of VRAM. Meanwhile the SD upscaler runs in more manageable chunks, so larger images only take longer time without consuming more resources.