Sensor Physics Quiz: ISO Invariance and Dynamic Range

In practice, a sensor is called ISO invariant above a certain ISO when shooting at that ISO and brightening in editing gives nearly the same noise as raising ISO in camera. Below that point, raising ISO in camera still gives cleaner results.

Read noise is a fixed amount of electronic noise added each time the sensor is read. In bright areas, it is tiny compared to the signal, but in dark shadows it is a large share of what was recorded. Lifting those shadows a lot brings it into view.

Full well capacity is the maximum number of electrons a photosite can accumulate before it clips. When capacity is exceeded, all excess photons are lost, producing featureless white with no recoverable highlight detail.

Raising ISO amplifies the sensor's signal, so bright areas reach the maximum recordable value with less light. Each stop of ISO removes about a stop of highlight headroom, so highlights clip sooner.

Shot noise comes from the random way photons arrive, so it varies from shot to shot. Its size grows with the square root of the signal: four times the light gives only twice the noise, so brighter exposures have a better signal to noise ratio. Fixed pattern noise is a separate type.

Dual conversion gain sensors can switch each pixel between two gain modes. Low gain preserves full well capacity and highlight headroom at low ISO; high gain cuts read noise at higher ISO. The camera switches at a set ISO to get the better balance at each setting.

Banding is a pattern of faint stripes caused by read noise and fixed pattern noise from the sensor's electronics. It is hidden in a normal exposure, but lifting dark shadows by several stops amplifies it until the lines appear.

RAW files store 12 or 14 bits per channel and leave white balance, contrast and color unapplied. JPEG has only 8 bits and those choices are already fixed, so it breaks down faster when exposure or color is pushed hard.

Clipped highlights have no detail left to recover, while dark shadows still hold data that can be lifted, at the cost of some noise. In a high contrast scene the safer choice is to expose so the highlights are kept, then brighten the shadows in editing.

Brighter exposures give each photosite more photon events, raising the signal level relative to the random electronic noise floor. The higher the signal-to-noise ratio, the less visible the noise is in the final image.

Shot noise, read noise and fixed pattern noise are all real sources of noise in digital images. Diffraction is an optical effect that softens detail at small apertures, not a type of noise.

When an analog signal is sampled and converted to digital values with a limited number of bits, tonal steps can only be represented at fixed intervals. The rounding error between the true value and the nearest available step is quantization error, which produces visible banding in smooth gradients.

Exposing to the right gives the sensor as much light as possible without clipping. More light means a stronger signal compared with the noise, so once the image is brought back to normal brightness, shadows are cleaner. It does not affect autofocus, distortion or white balance.

Clipping happens when a pixel reaches the sensor's maximum value, so all brighter tones record as pure white, or when it reaches the minimum, where darker tones all record as pure black. Detail in those areas is lost.

Shadow noise is mostly set by how much light the sensor collected. Giving the sensor more photons, with more exposure but no clipped highlights, raises the signal compared with the noise. Underexposing and lifting five stops later amplifies the noise along with the image.