Both approaches turn a small image into a larger one, but they do fundamentally different things — and the difference matters legally as well as visually.

Traditional Interpolation

Nearest-neighbour, bilinear, bicubic and Lanczos all work the same way in principle: to create a new pixel, they average the surrounding original pixels according to a formula.

  • Nearest-neighbour copies the closest pixel — produces blocky, hard edges. Correct for pixel art, where you *want* hard squares.
  • Bilinear averages four neighbours — fast, soft.
  • Bicubic uses sixteen neighbours with a smoother curve — the general-purpose default.
  • Lanczos uses a wider window and preserves more apparent sharpness, sometimes with slight ringing at edges.

The essential property: no new information is created. Enlarging 2× with bicubic gives you a bigger, softer version of the same picture. Nothing is fabricated, which makes interpolation predictable and safe.

AI Upscaling (Super-Resolution)

A model trained on millions of image pairs predicts what a high-resolution version *would plausibly look like*. Where interpolation averages, the model synthesizes: it adds edges, textures and detail that were never in the source.

The results can be remarkable — sharp edges, plausible texture, no softness. But the detail is invented, not recovered. It's a statistically likely guess about what might have been there.

Where AI Upscaling Genuinely Helps

  • Moderate enlargement (up to about 2–4×) of photographs for print or display.
  • Old scans and low-resolution archive images for personal viewing.
  • Textures and backgrounds where plausible detail is more useful than accurate detail.
  • Recovering compressed images — many models also reduce JPEG artifacts effectively.

Where It Goes Wrong

  • Faces. Models fill in features from training data, which can subtly change identity — a well-known research finding. An upscaled face of a specific person may not be that person's actual features. Never use for identification.
  • Text. Small text becomes *sharp but wrong*, with letters invented into plausible shapes. This is worse than blurry text, because it looks authoritative.
  • Fine patterns: fabric weave, hair, foliage may be replaced with generic learned texture.
  • Medical, scientific, forensic and evidentiary images — inventing detail invalidates them entirely.
  • Extreme enlargement. Beyond about 4×, output is largely model imagination.

The Rule That Covers Both

Ask: does anyone make a decision based on the detail?

  • Family photo enlarged for a wall → AI upscaling is great.
  • Product photo where the buyer inspects texture → be careful; misrepresenting a product has consequences.
  • Anything identifying a person, reading a number, or documenting a fact → never.

And the practical alternative that beats both: go back and find a larger original. Re-scan the negative, re-export from the RAW, request the source file. No upscaler beats real pixels, and the search usually takes less time than the processing.