Four pixels into one: why twelve megapixels beat fifty after dark

There is a switch in the camera app that promises four times as much. Left alone, the phone hands you twelve megapixels; tap «50 MP» and it hands you fifty, in files four times the size. In daylight the choice barely matters. At night it decides the photograph, and usually for the worse, because the sensor has not gained any pixels. It has stopped adding them together, and that sum was the best favour it was doing you.

Four photosites under one colour filter

The Bayer mosaic from the textbooks gives every photosite its own colour filter and alternates them on a fixed pattern. The ultra high resolution sensors in phones changed the deal: the filters come in two by two squares, four photosites sharing a colour. Samsung calls its version Tetrapixel, and the public sheet for the ISOCELL JN1 spells it out plainly enough to quote. Fifty megapixels of 0.64 micrometre pixels inside a 1/2.76 inch format. Four neighbouring pixels working as a single 1.28 micrometre pixel and delivering a 12.5 megapixel image with less noise. Full resolution, the same sheet adds, comes from a re-mosaic algorithm that rearranges the pixels, and it does that when light is plentiful.

That last detail deserves a second reading. The native mode of these sensors is the summed one; full resolution is a software reconstruction of a Bayer pattern that does not physically exist under the filters. The reconstruction does fine work on a daylight scene stuffed with signal. On a night sky it works from thin data, and what it invents is not detail. It is a guess.

The noise arithmetic, done honestly

Four photosites becoming one collect four times the photons. Counting noise grows as the square root, so the signal to noise ratio doubles: two full stops, free, without touching either the shutter or the gain. It is the same arithmetic that makes stacking work, applied inside the chip instead of inside the software.

The obvious objection: shoot at fifty megapixels and downsize to twelve and a half at home, in your own time. Where photon noise rules the scene, that objection holds and the two routes come out level. The sums change once read noise joins in, the noise the electronics add at the moment they measure the charge. If the four photosites are summed as charge before they are read, you pay that noise once against a quadrupled signal. Read them separately and average afterwards and you pay it four times, then halve it. The first route is twice as good on read noise as the second, and a night sky is precisely where read noise matters, because the signal in any one pixel is pitiful.

Precision about what is actually known: whether a given phone sums charge inside the pixel or averages already digitised values inside the image processor is something manufacturers rarely state. The difference is real, it is measurable, and it is one of the few things you can settle yourself with two sets of frames. Until you have, the summed mode is the one the sensor was designed to use in the dark.

The lens has no such detail to give

Say you want to pay for that extra detail anyway. It still has to come from somewhere, and a phone lens has a physical limit no sensor can talk its way around. An f/1.8 optic, the usual figure for a main camera, produces an Airy disc roughly 2.4 micrometres across at 550 nanometres, the wavelength the eye handles best and a sensible one for arithmetic. On a 0.64 micrometre grid that disc smears over nearly four pixels of diameter. On the summed 1.28 micrometre grid it covers under two, which is about the right sampling: enough not to throw information away, not so much that you are describing empty space to four decimal places.

Diffraction is only the start. Add the aberrations at the edge of the field, generous on any phone wide angle, and add motion. On a fixed tripod stars begin to stretch within seconds, and their image is already wider than the theoretical disc before the sensor ever samples it. How long that trail takes to appear depends on declination and focal length, and those numbers live in another piece. Full resolution is promising to resolve structures finer than your optics and your mechanics are handing over.

What the API declares, and what the app passes on

Since Android 12 the platform treats these sensors as a category of their own: there is a maximum resolution mode alongside the default one, and the camera documentation carries a flag, SENSOR_RAW_BINNING_FACTOR_USED, that says whether the RAW buffer you received was binned. It exists because it is needed. Without it, an app has no way of knowing what it is holding.

The practical fallout is concrete. A full resolution RAW may be the sensor read out in maximum resolution mode, or it may be the output of the re-mosaic, and the two behave differently once you stack them. It runs the other way too: some phones save a «50 MP» DNG whose fine detail looks exactly like an upsampled 12.5. If your capture app exposes the flag, look at it. If it does not, fall back on the old habit of checking what is really inside a file rather than trusting the label on it.

The bill comes due downstream

Four times the pixels means four times the data to write, read, align and sum. Across a three hundred frame session that is not a theoretical difference: it is the free space on the phone, it is the minutes the processor spends heating up while you would rather be shooting, and it is sensor temperature climbing at exactly the moment you are trying to keep thermal noise down. Alignment charges a surcharge of its own, since it works by recognising stars, and faint stars on a full resolution frame sit much closer to the background.

When full resolution earns its keep

Two cases, and it is fair to name them. The Moon sends photons in embarrassing quantities and lives in a regime where noise decides nothing and resolution decides everything: if your phone genuinely reads the sensor in maximum resolution mode, one more crater may survive, and that is worth testing. The Sun through a certified filter belongs in the same box for the same reasons. Everywhere else, deep sky, conjunctions, wide fields, nightscapes, the calculation flips and megapixels turn into a cost.

One evening, well spent, settles the argument. Same target, same tripod, two runs of fifty frames close together, one in normal mode and one at full resolution. Downsize the second run to the size of the first, then look at three things: the grain of the sky background at matched display scale, the full width at half maximum of one isolated star, and how many faint stars survive the stretch. The verdict applies to your phone and to nobody else’s, which is exactly the kind of truth that is useful in a dark field.

The big number on the spec sheet was written for the shop floor. At night, the sensor already knows what to do.