For an image, upsampling estimates values on a larger pixel grid. Simple methods interpolate neighboring values, while learned methods may use patterns from training to predict additional detail.
A larger image does not automatically contain more reliable information. Generated detail can look convincing without matching the original scene, and processing may introduce blur or artifacts.
ELI5
Think of enlarging a small picture by working out what to put in the extra spaces between its pixels.
For example, a photo can be expanded to twice its width, but fine lettering that was unreadable in the original may still be uncertain.
Does upsampling recover every lost detail?
No. It estimates new sample values, and information absent from the original cannot be guaranteed by enlargement.
Is upsampling only used for images?
No. The same general idea also applies to audio and other sampled signals.
