What is Denoising? — one analogy

1The analogy

Feed an old document into a copier and the yellowed stains and specks on the paper get copied right along with the writing. Nothing about the original page changes, so the only place to clean up the stains is on the copy that comes out.

A density dial does that job. Turn it down and the faint stains fade away — but a pencil line written lightly fades with them. Turn it up and the faint writing comes back, but so do the stains. There's no setting that gets you both at once.

Someone who's done this before doesn't reach for the dial first. They look at the page and figure out whether it's an even yellow cast across the whole sheet or a crease line, and adjust to match. Denoising is that same read-and-adjust move.

2In detail

Spot the outlier and pull it toward its neighbors

Zoom way into a photo and an even surface like the sky turns out to have scattered points that are unusually bright or dark. Compared to the points around them, they stand alone — and that gap from the surroundings is exactly what marks them as flecks.

A simple fix blends in a few neighboring values to nudge the point back toward the average. Scattered points shrink noticeably this way. The catch is that eyebrows or loose threads — real detail that's also different from its surroundings — get smoothed away right along with the flecks.

So current methods tell speck from pattern by shape. Train on countless pairs of clean photos and photos with flecks laid on top, and the model picks up the trick of keeping edges intact while stripping out only the scattered points.

Find it, then subtract it

A trained model doesn't hand back a clean photo right away. First it works out where the flecks are in the current picture and how much of them there is, then subtracts that amount from the original picture. It's the same order as reading a page first and setting the density dial to match.

This order has an advantage. Drawing a clean result outright would mean rebuilding the whole photo from scratch, but pointing out only what's layered on top is a far easier problem. The real content stays put and only the layered part gets touched, which also cuts down on mishaps like a face turning into someone else's.

How much to scrub is always a trade-off

A person can set how hard the scrubbing runs. Set it light and flecks remain; set it hard and the picture turns smooth, but fine texture like hair or fabric weave disappears along with it. No setting removes every fleck while keeping every detail intact.

A heavily scrubbed photo can look tidy at a glance but somehow waxy. Places that should carry natural fine texture — skin, grass — turn slick instead. That giveaway shows up most clearly when the image gets enlarged or printed.

Knowing the kind of speck helps it get removed

Flecks come in different kinds. The grainy speckle from a low-light photo, the static laid over an old recording, and the banding from a dropped signal all look different from each other. Knowing which kind is mixed in lets the removal land far more precisely.

This comes in handy for generating pictures. Since the training process laid the blur on deliberately, it can also tell the model exactly how much was added, as a number. The model takes that figure along with the picture and uses it to gauge how much to remove at this particular step — like being told in advance how much landed on the page before reaching for the dial.

Repeat it and a picture that never existed appears

Removing flecks once leaves a photo a little cleaner. Feed that result back in, remove again, and repeat dozens of times — what happens? If it started as a photo, it keeps becoming a smoother photo. But if it started as a screen of meaningless blur, shapes that were never there gradually settle into place.

At every removal, the model nudges the screen a little toward "if there were a plausible picture here, it would look like this." Stack enough of these nudges and a blank screen becomes a finished picture. That's why the technique for cleaning up flecks and the technique for generating pictures share the same root.

3More precisely

Denoising treats an observed value as the sum of an original signal and a noise component, and works to estimate and subtract the noise. Training builds its pairs by deliberately adding known noise to clean material and having the model guess what was added. Because no one has to hand-label the correct answer, practice material is nearly unlimited.

The analogy breaks down in one place. A copier's density dial applies evenly across the whole sheet, but denoising judges each spot on its own — scrubbing hard in some places and leaving others untouched. And where the stains on paper are visible to the eye, the flecks in a photo sit in the very same spots as the real signal, so where the fleck ends and the content begins is never a clean line, and the value pointed to is always a guess. That's why a scrubbed result sometimes carries texture that was never there, while a very faint detail, like a fingerprint, is lost rather than recovered.

One more point: the result of scrubbing isn't identical to the original. It's a plausible fill-in, so anywhere the exact original matters — evidence photos, for instance — that has to be stated plainly.

4Try it yourself

Diffusion Explainer ailearn.space Watching what gets subtracted at each step builds a feel for how the removal works

5Common misconceptions

It's easy to think denoising brings back information that was lost, but actually it plausibly fills in detail buried under flecks rather than recovering the value that was actually there.
It's easy to think scrubbing harder always gives a better result, but actually fine texture disappears along with the flecks, leaving a waxy, smoothed-over look.
It's easy to think removing flecks and generating a picture are separate things, but actually today's picture generators are built out of repeating this same removal step over and over.

7One-line summary

In shortDenoising is the act of finding only the flecks layered onto a result and subtracting them, and repeating that same move all the way leads to generating a picture that never existed.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02