Digital noise
Digital noise is the visible grain and random color specks in an image that degrade fine detail and the cleanness of tones.
Also known as: image noise, sensor noise, grain
Digital noise is the visible grain and random color specks that appear in an image, degrading fine detail and the cleanness of tones. It occurs mainly at high ISO (the sensor amplifies both the signal and the background noise) and in very long exposures (the sensor heats up and generates "thermal noise").
Raising the ISO does not add noise: it reveals the noise that was already there because few photons reached the sensor.
There are two main types: luminance noise (black-and-white grain, relatively tolerable) and color noise (red, green and blue specks, more annoying). Both can be reduced in editing, though at the cost of losing some detail.
Why does high ISO produce noise?
ISO does not collect more light: it amplifies the signal that already reached the sensor, and it amplifies the noise along with it. What creates the grain is the scarcity of photons. The signal grows with the number of photons captured and the noise grows with their square root, so losing four stops of light (from f/2.8 and 1/60 s to f/5.6 and 1/250 s, for example) makes the signal-to-noise ratio four times worse.
The practical consequence reverses intuition: the best tool against noise is not lowering the ISO, it is giving the sensor more light. Opening the aperture one stop or doubling the exposure time cleans up the shot; lowering the ISO with the same light only leaves it dark, and brightening it in the raw converter brings back the same grain.
At very high ISOs a second component comes into play, the sensor's own read noise, and there the square-root rule no longer describes the full curve. Exactly how much depends on each sensor.
How high can I raise the ISO?
It depends on the size at which you will display the photo: the same file that looks dirty at 100% on screen looks clean printed at 8 × 12 in (20 × 30 cm). These are reference settings on full frame; on APS-C or Micro Four Thirds you will get somewhat more noise at the same ISO because of the smaller sensor size.
Scene: Indoors in daylight next to a window; Aperture: f/2.8; Shutter speed: 1/125 s; ISO: 400; What you accept: Invisible noise
Scene: Concert in a small venue; Aperture: f/1.8; Shutter speed: 1/250 s; ISO: 3200; What you accept: Fine grain in the shadows
Scene: Street at night with streetlights; Aperture: f/2; Shutter speed: 1/60 s; ISO: 6400; What you accept: Color specks in the blacks
Scene: Milky Way at 24 mm; Aperture: f/2.8; Shutter speed: 20 s; ISO: 3200; What you accept: Visible grain; several shots are stacked
Scene: Church interior with a tripod; Aperture: f/8; Shutter speed: 4 s; ISO: 100; What you accept: Zero noise, and people walking through disappear
The 20 seconds for the Milky Way come from the 500 rule (500 ÷ 24 mm ≈ 20 s), an approximation that falls short on high-resolution sensors. The NPF rule calculator gives the real limit for your camera before the stars come out elongated.
Can noise be removed in editing?
It is reduced, not removed. The AI-based noise reducers in today's raw converters reconstruct plausible detail from the grain, and they work well up to a point; past that point they invent plastic textures on skin and erase faint stars.
Order matters: treat color noise first, since it can be removed at almost no cost, and only then luminance noise, which does take detail with it. A little luminance grain reads as photographic texture; a smoothed-out face reads as an error.
Common mistakes
Underexposing to "save" ISO. Shooting two stops under at ISO 800 and lifting in the raw converter gives noise similar to shooting at ISO 3200 with the same aperture and shutter speed: the sensor receives the same light in both cases, and that is what decides it. Underexposing does not buy cleanliness.
Leaving noise reduction on for RAW. On many cameras the setting only affects the JPEG, so you judge on screen a cleanness the file does not have. Evaluate the RAW in the raw converter.
Ignoring thermal noise in summer. A 5-minute exposure with a hot camera generates fixed hot pixels that are not random grain and do not disappear by averaging. This is where the camera's long-exposure noise reduction does help.
For in-camera and software reduction techniques, see night noise reduction: camera and software, the guide to smart ISO and the term dynamic range, which explains why shadows are always the noisiest area.