Model-based upscaling produces enlargements that look sharper than the source could support. Understanding where that sharpness comes from determines when the result can be trusted.
What the older methods did
Traditional enlargement interpolates between existing pixels, estimating intermediate values from neighbours.
This cannot add information, so the result is larger and correspondingly softer, which is an honest representation of what was captured.
The softness is why interpolation fell out of favour, since users wanted enlargements that looked like higher-resolution originals.
What the model does instead
A model trained on pairs of low and high resolution images learns what fine detail typically accompanies a given coarse pattern.
Applied to a new image, it generates the detail that would plausibly have been there, based on everything similar it has seen.
The output is sharp because the detail is real detail, and it is invented because it came from the training distribution rather than from the photograph.
Why faces are the clearest example
Facial detail is heavily represented in training data, so models reconstruct eyes, skin texture and hair confidently from very little information.
The reconstruction is a plausible face rather than the face that was photographed, and at low source resolutions it can resemble a different person entirely.
This is the failure mode that matters legally, since an upscaled surveillance frame looks like evidence and is closer to an illustration.
Where invented detail is entirely acceptable
Preparing an image for large-format print, upscaling old footage for modern displays and enlarging a crop for a background element all tolerate invented detail.
In each case the purpose is a pleasing image at a required size, and nobody will compare the result against ground truth.
Restoration of degraded archival material sits in the same category provided the enhancement is disclosed, since a viewer should know which parts of a historical image were reconstructed.
How to tell what was added
Running the same upscale several times and comparing the results reveals what varies, and anything that changes between runs was generated rather than recovered.
Stable areas correspond to information genuinely present in the source, so the comparison functions as a rough map of what is real.
Making that check routine before an enlargement is used for anything consequential costs a few minutes and prevents the specific error of treating a model's best guess as a photograph.