Resampling = averaging
Photoshop's Image Size, GIMP's Scale Image, Preview and Paint all resample: every new pixel is a weighted average of the originals [1]. Bicubic (Keys, 1981) uses the 16 nearest pixels [2]; Lanczos (Duchon, 1979) is a windowed-sinc filter that is sharper but can ring around edges [3]. None of them can create detail that wasn't captured.
GIMP's NoHalo and LoHalo
GIMP 3 offers None (nearest neighbour), Linear, Cubic, NoHalo and LoHalo [4]. NoHalo and LoHalo are designed to avoid the halos that sharper filters create, which makes them the safest manual choice for enlarging photos.
AI Super Resolution
Adobe's Super Resolution in Camera Raw uses a trained model to double the width and height of a photo [5] — a small, fixed step. Online AI upscalers apply the same idea with larger models: Real-ESRGAN, for example, was trained on blurred, noisy and JPEG-compressed images so it can add plausible edges and texture as it enlarges [6].
Why no method recovers everything
A sampled image can only hold detail up to half its sampling rate [7]. Beyond that, AI chooses a plausible version — which is why faces and text in an upscaled image should be checked before you rely on them.











































