The movie misconception

In crime shows a pixelated surveillance frame gets “enhanced” at the press of a button, and suddenly the license plate is readable. In reality that's impossible — and the reason is called interpolation. Understand what it does and you'll understand why a small image, when enlarged, doesn't get sharper but only bigger and softer.

What interpolation does

When you enlarge an image, new pixels appear for which there's no original data. Where there was one pixel, there should now be four — but from where? Interpolation computes these new pixels from their neighbors, for example as an average of the surrounding colors. The result is a smooth, soft transition. The key point: they're estimates, not real detail. Interpolation invents nothing; it fills the gaps with the most likely value.

Why no sharpness appears

An image can only hold the detail captured when it was taken. If it isn't there (because the original was small or soft), no calculation can bring it back — you can't create information from nothing. Enlarging stretches the existing, limited information over more pixels; the image grows, the detail stays the same, spread over more area — so relatively softer. That's why the rule is iron: shrinking is near-lossless, enlarging is not.

The methods: bilinear, bicubic, Lanczos

There are several interpolation methods, differing in how many neighboring pixels they consider and how they weight them:

  • Nearest neighbor: simply takes the closest pixel — hard, blocky, but pixel-accurate (good for pixel art).
  • Bilinear: average of the four neighbors — soft, simple.
  • Bicubic: considers 16 neighbors — sharper, often the default.
  • Lanczos: more involved, gives especially sharp results when shrinking.

Important: when shrinking, the method choice makes a real quality difference (which is why good scaling tools use Lanczos or bicubic). When enlarging, they only decide how the unavoidable detail loss looks — softer or a little edgier.

Where AI makes a difference

Here's the one route that goes beyond classic interpolation: AI upscaling. Instead of just computing in-between values, an AI generates plausible detail from knowledge of millions of similar images — it “knows” what a brick wall or an eye looks like and adds matching structure. It can look strikingly good. But it stays invention, not reconstruction: the detail is probable, not guaranteed real — the license plate is guessed, not revealed.

The practical advice

Enlarge only in a pinch, and with realistic expectations. Better always to get the original at a sufficient size. And when scaling, remember: shrinking is your friend (near-lossless, saves space), enlarging is the compromise. Where there are no pixels, no software conjures real ones — only AI guesses cleverly.

In short

  • Interpolation estimates new pixels from neighbors — no real detail.
  • You can't add what was never captured — enlarging goes soft.
  • Method matters when shrinking, less when enlarging.
  • AI upscaling invents plausible detail — impressive, but a guess.