Plain enlargement (bicubic and the like) interpolates existing pixels to fill gaps, so it gets softer and blurrier the more it is stretched. A trained upscaler, learned from pairs of low- and high-resolution images, instead "guesses" plausible new detail — hair strands, fabric weave — that was not in the source.
Diffusion-based upscalers often treat the enlarged image like an img2img pass — add a little noise, then denoise it again at low strength, adding only detail that naturally fits the structure already there.
"Filling in detail" also means inventing something that was not in the original. That is fine where merely plausible texture is enough, but risky where an original must be respected — text or faces — since invented detail there can read as distortion rather than sharpening.
The demo below is a simulation, not a real upscaler — it draws the same seeded scene coarsely at low resolution on one side and with added fine strokes and grain at higher resolution on the other, to contrast the two.
When to use
Use it when an already-made image needs to go to a larger resolution. For areas where accuracy matters — text, faces — it is safer to check the result visually.