You wiped the lens twice and the speckles are still there—red, green and blue confetti scattered through the shadows of a photo taken at a dim dinner table. Nothing is wrong with the camera, and nothing is on it. The speckles are the visible signature of statistics: in near-darkness, each pixel receives so little light that the randomness of photons arriving becomes a large fraction of the signal, and the electronics that measure that trickle add their own uncertainty. The final insult is algorithmic. Most sensors see only one primary color per pixel and compute the rest from neighbors, so every tiny error is not merely shown—it is shared, and painted, in color. The same scarcity of light that shapes camera behavior also reshapes human vision after dark—it is the regime behind why blue LED signals look so piercing at night—so what follows about cameras has a cousin in your own eyes.
Where the confetti comes from
Too few photons, amplified and interpolated.
Shot noise from scarce light, read noise from the sensor, then demosaicing spreads each pixel's error into colored clusters.
Light itself is grainy
Light arrives in packets, and packets arrive at random intervals. Bright scenes flood every pixel with vastly more photons, so the randomness averages out and the image looks smooth. Dim scenes do not grant that luxury. When a shadow region delivers only dozens of photons per pixel, one pixel may catch a few more and its neighbor a few less, and that spread—shot noise—is a meaningful percentage of the whole measurement. Amplifying the signal afterwards, which is what a higher ISO setting does, raises the picture and the speckles together. You cannot amplify a signal whose noise was baked in at capture.
The sensor reads imperfectly, too
Reading a pixel out is an electrical measurement, and electrical measurements have a noise floor. Sensor engineers quantify this as read noise, one of the fundamental figures tracked in camera characterization—the classic photon transfer method treats it as a basic parameter of any imaging system. Read noise matters most exactly where the signal is weakest, because a fixed quantity of electrical uncertainty looms large against a trickle of light. That is why the darkest corners of a night photo degrade first: they face both kinds of noise at once, photon randomness below and read noise above, with little signal in between.
Why the noise is colored
The color part of the confetti comes from how cameras build color at all. A consumer sensor does not record red, green and blue at every pixel; it records one primary per pixel in a mosaic pattern and interpolates the missing two values from neighbors. Research on noise-resistant demosaicing exists precisely because this interpolation and noise interact: a single erroneous pixel is spread across the colors of the pixels around it, so a spark of raw noise becomes a small colored blob. Fine luminance grain is often tolerable to the eye, but these interpolated chroma blotches read as unreal, which is why noise-reduction tools hunt them specifically and why over-aggressive removal leaves surfaces looking plastic.
Both panels use the same base tone. The right panel adds scattered red, green and blue clusters of the kind demosaicing produces from raw pixel errors.
Where reduction happens changes what survives
The order of operations matters as much as the tools. Raw processing can, in principle, weigh the evidence before color interpolation has ever guessed, keeping honest uncertainty instead of baking it into colored blotches; by the time a heavily processed JPEG leaves the phone, decisions about what was noise and what was texture have already been made for you. This is why the same night scene can look different from two phones or two apps: sensors may genuinely differ in read noise and sensitivity, and pipelines also differ in where they spend their smoothing—the two sources of difference are not mutually exclusive. Understanding this turns gear arguments into a smaller question—first get more photons, then choose a pipeline that spends its guesses carefully.
Practical consequences follow directly. More light—longer exposure, wider aperture, added illumination—beats any ISO setting, because it addresses the photon scarcity that starts the chain. RAW files preserve the noise honestly and let reduction work before compression; heavy in-camera smoothing at extreme ISO has already guessed on your behalf. And comparing two dark scenes for color, like deciding whether a shadow tone has drifted toward blue, is exactly the judgment our eyes are worst at in dim conditions—the naming boundary you feel in Blue or Green? wobbles when the stimulus is faint and noisy, for cameras as for people.
A short checklist for the next dark room
Move the light first: one lamp rerouted toward the subject outperforms every menu setting. Stabilize the camera second—against a doorframe if no tripod—so a longer exposure is possible without motion smear. Prefer the widest aperture available and the lowest ISO that the situation allows, checking the result at full size rather than trusting the preview screen. If the photo must be pushed to extreme ISO anyway, prefer apps or modes that keep noise reduction modest, because plastic faces and smeared text age worse than honest grain. And when a treasured photo is already speckled, treat the colored blotches and the gray grain separately in editing: chroma noise can often be tamed without flattening the luminance texture that makes the picture feel real.