ISO is not a light switch: how much gain a dark sky actually needs

The ISO slider is the easiest control to reach in a manual camera app, and in the dark it is also the most seductive. Thirty seconds of exposure will not pull the Milky Way out of the preview, so up goes the gain, the screen brightens, and the problem looks solved. Then you get home to a stack full of bloated white stars sitting on a background like wet sand. That slider has never let a single extra photon through the lens. Knowing what it does move is the difference between a night you used and a night you merely amplified.

The photosite counts, the gain multiplies

A photosite works in a strict order. It collects photons for the length of the exposure, turns them into electrons, and piles that charge into a well. How many electrons end up in the well depends on exactly three things: how much light that patch of sky is throwing at you, how wide the lens is open, and how long the shutter stays open. Gain shows up after the collecting is done, at the moment the charge has to be read out. It amplifies a voltage on its way to the analogue-to-digital converter.

Which is why the word «sensitivity» is a hangover from film. The standard that governs all of this, ISO 12232:2019, does not define one sensitivity and stop there: it defines four quantities (exposure index, ISO speed ratings, standard output sensitivity, recommended exposure index), and that alone tells you something. In a digital camera the relationship between the light arriving and the number written into the file is a convention chosen by the manufacturer, built out of variable gain plus processing applied after capture. On film, ISO was a chemical property of the emulsion. On a phone it is a lookup table.

The noise no amount of gain will fix

There is a limit that sits upstream of the sensor, upstream of the lens, upstream of everything you own. Photons arrive at random intervals, so their arrival is a Poisson process: the noise on a count of N electrons is the square root of N. That is the same arithmetic the Space Telescope Science Institute puts at the heart of its WFPC2 Instrument Handbook when it works out exposure times for Hubble, and it applies unchanged to your phone on a forty-euro tripod.

You can feel the consequences straight away. A pixel that has collected 100 electrons carries 10 electrons of intrinsic noise: a signal-to-noise ratio of 10, and no electronics on Earth will improve it. Amplify by eight and you get 800 electrons of signal with 80 of noise. Still 10. The picture on the screen is brighter and the information inside it has not budged. Raising ISO shows you sooner what you already caught. It does not catch you any more.

Where the analogue stops and the pretending starts

Here is the part that matters in the field. Real analogue gain, the kind applied to the voltage before conversion, only exists up to a point on the scale. Android’s Camera2 API documentation states this outright through the SENSOR_MAX_ANALOG_SENSITIVITY characteristic: below that value the device guarantees it is using analogue amplification only, and above it the remainder comes from the image signal processor, which is to say a multiplication performed on numbers that have already been digitised.

You could perform that multiplication yourself afterwards, with any slider you like. Doing it in-camera costs you something: it inflates the high values until they clip and then writes them into the file that way, and on phones that apply tone mapping before saving, the damage is permanent. If your capture app exposes that figure, look at it once and remember it, because it marks the ceiling beyond which you are paying in dynamic range for something you do not receive in signal. If the app hides it, there is an empirical route: climb in steps and find the point where the sky background gets brighter while the faint stars stop gaining any contrast.

The one kind of noise gain really does touch

There is an honest exception, otherwise every camera would run at fixed gain. Read noise, the noise the electronics introduce at the moment they measure the charge, behaves differently from Poisson noise: raise the conversion gain and the useful signal is amplified ahead of the noisy stage, so the input-referred read noise drops. That is exactly why modern sensors ship with dual conversion gain and a high-gain mode intended for dark scenes.

The price is written on the same line. Raising the conversion gain shortens the well. One well-documented pixel, the 0.8-micron dual-conversion-gain design for a 64-megapixel sensor presented at IEDM in 2019, quotes a full-well capacity of 12,000 electrons. At that level Poisson noise is roughly 110 electrons, so the signal-to-noise ratio is about 110: the theoretical best that pixel can deliver in a single frame. Switch to high-gain mode and the well shortens, and so does that ceiling. You gain in the shadows and give it back in the highlights. Bright stars live in the highlights.

The practical rule, in three moves

Exposure time gets decided first, and what decides it is the rotation of the Earth by way of the limit beyond which stars start to trail. That is a geometric constraint. Everything else queues up behind it.

With the time fixed, gain has exactly one job: lifting the sky background off the bottom of the scale. If the sky signal stays pinned at zero, the few electrons you collected get crushed into the converter’s lowest levels and quantisation eats part of them. Judge it on the histogram rather than the preview: the sky peak should come away from the left edge and settle around a tenth of the way up. Every ISO stop past that is dynamic range given away for free.

Third move, and the one that gets forgotten. Keep ISO identical across the whole session, calibration frames included: read noise and offset both shift with gain, so a dark shot at a different ISO subtracts something your lights never contained. The same holds in reverse. If you raise the gain halfway through the night because the Moon has set, you now own two sets, and two sets is how they have to be treated.

Three hundred frames beat any slider

The closing arithmetic never changes. Stack N aligned frames and the signal grows with N while random noise grows with the square root of N, so a hundred integrated exposures are worth ten times the signal-to-noise ratio of one. No position of the ISO slider does anything of the kind, for the plain reason that gain multiplies the noise it is trying to bury along with everything else. Total time on target is the only currency the sky accepts.

Which yields a rule that sounds lazy and works better than the alternatives: pick the lowest gain that keeps your sky background clear of zero, then leave it alone and go to work on the frame count. Three hundred frames at moderate gain give you stars with readable colour, a background you can stretch, and calibration frames that still mean something. Thirty frames at maximum ISO gave you everything they had while you were still looking at the screen.

The sky is stingy with photons, but it is even-handed about it. They go to whoever stays out longest.