Noise

The first face: blur mixed in

Anything recorded picks up extra baggage along the way. A photo shot in the dark gets speckled with colored dots, and a voice recorded in a café picks up the clatter of dishes. Whatever comes along that has nothing to do with what you meant to capture — that's noise.

Noise has different personalities depending on the kind. Some, like photo speckle, jumps around at random from spot to spot. Others, like straight vertical banding, follow a rule. The kind that follows a rule is actually easier to remove, because it's easier to tell where the baggage is.

The trouble comes when the real pattern looks like the noise. Sand's natural bumpiness and photo grain can look nearly identical, so trying to erase the fleck can wipe out the sand along with it.

The second face: the starting point where a picture rises

Picture-generating AI doesn't start from a blank page — it starts from a single screen of blur. It lays down a pile of meaningless dots, checks in with the request, and clears the blur away a little. It looks at the result and clears a little more. Repeat this dozens of times and the blur resolves into a picture.

Because the starting blur is freshly drawn every time, the same request produces a different picture. Hold onto the number used to draw that blur and the same picture can be called back up again.

There's a reason the clearing doesn't happen all at once and gets split into steps instead. While things are still blurry, only the big shapes and layout get decided; as the blur thins out, fine branches and texture get settled. Clear it in a hurry and the details lock in before the big picture is settled.

Practicing adding it and clearing it away

This skill isn't free. During training, a large pile of clean photos gets collected, and noise gets laid onto every one of them. Several stages get made, from barely any noise at all up to so much that the original shape can't be told.

The model is then shown only the noisy photo and asked to guess what the layer that was just added looked like. Since the answer is already known, grading is easy. Repeat the correcting enough times and the model learns to point out what's baggage even in noise it's never seen before.

Adding the noise is a rule a person sets, so it's simple; only the removing side has to be learned. Not needing a person to hand-write every correct answer is a big part of why this approach is so widely used.

Sometimes it's shaken in on purpose

Training uses noise for another purpose too. Show the same photo over and over and the model just memorizes that one photo. So the photo gets tilted slightly or dusted with a bit of blur, showing something a little different each time. Learning from material that's shaken up a bit this way builds an eye that also works on photos it's never seen.

There's also a way to shake values slightly while learning is underway. A bit of jitter makes it easier to climb back out after falling into a shallow trap. Total quiet isn't always the better environment.