A finished astrophoto is several kinds of frame combined: lights of the sky, darks with the cap on at the same exposure and ISO, flats of an even surface at the same focus, and bias or dark-flats to calibrate the flats. Each removes one thing the sensor or lens added, and a star-trail night needs fewer of them than a deep-sky night does. This page says what each frame records, how to shoot it so that it actually works, how many to take and why, the arithmetic that combines them, and which of them your night needs at all. The numbers about the sensor are measured on real frames — twenty-five bias frames from the same camera as the pictures — and the exposure rules are computed from their published formulas, with the sources linked where each figure is stated.
A camera does not record the sky. It records the sky plus everything the sensor and the lens add on the way: a fixed offset every pixel reports even in total darkness, a slow accumulation of thermal signal that grows with exposure and temperature, a scatter of pixels that are always too bright, a falloff toward the corners because the lens delivers less light there, and the soft shadows of dust on the filter in front of the sensor. In a daytime photograph these are invisible under the signal. At night, when the signal is a few photons per pixel, they are the picture.
Calibration frames are photographs of those additions on their own, so they can be taken back out. A dark frame is the sensor’s contribution with no light at all. A flat frame is the lens’s contribution, lit evenly. A bias frame is the offset alone. Each one is shot so that it contains exactly one thing, and the arithmetic in section 8 removes each thing from every light frame before the frames are ever combined. The Calibration Lab on this site is where that arithmetic happens; its own page describes what it does with the frames once you have them. This page is about how to have them.
The light frames are the photographs of the sky — every one you meant to take. Everything about how to shoot them well for star trails is on the field guide; for pinpoint stars and the Milky Way, section 10 below gives the exposure rules and section 11 the lens. Three things matter here because the calibration frames have to match them: the exposure time, the ISO, and the physical state of the lens — focus, aperture, and how the camera is rotated. Write them down, or leave the camera untouched until the calibration frames are done. A dark shot at a different ISO, or a flat shot after the focus ring moved, is a frame that removes the wrong thing.
Shoot RAW. Every frame on this page is arithmetic on linear sensor counts, and a JPEG has already had a curve applied, noise reduction run, and the offset stripped. The tool reads camera RAW for exactly this reason; nothing below works on a JPEG.
A dark frame is an exposure with no light reaching the sensor: the lens cap on, the viewfinder covered, the same exposure time and ISO as the lights, at the same sensor temperature. What it records is what the sensor generates by itself in that time — the thermal signal that accumulates in every pixel, the hot pixels that accumulate it far faster than their neighbours, and on some cameras a glow near an amplifier that brightens one edge or corner. Subtracting a good dark from a light removes all three.
“CCD Dark Frame” by H. Raab (User:Vesta), Johannes-Kepler-Observatory, Linz, adapted (recompressed), via Wikimedia Commons, used under the CC BY-SA 3.0 licensing option. This adaptation is offered under the same licence.
The how-to literature treats darks as mandatory. One careful measurement says otherwise for a specific class of camera, and it is worth understanding because it is not a small effect. Subtracting a dark removes the sensor’s pattern — the hot pixels and the glow, which are the same in every frame — but it also adds the dark’s own random noise to every light frame. On a sensor whose thermal signal is already tiny, there is little pattern to remove and the random noise added is a net loss. Roger Clark measured the dark current of a Canon 7D Mark II at about 0.02 electrons per second and concluded that on such sensors subtracting a dark frame can add more noise than it removes. His argument is at clarkvision.com, and the stacking comparison that goes with it is here. No source in the how-to literature engages with the point; they simply do not mention it.
What to do with that: if your camera is a recent body with on-sensor dark-current suppression — most Canon and Sony bodies of the last decade — test it. Stack one night with a master dark and once without, and look at the background at 200 %. If the version without darks is cleaner, dither instead: nudge the framing slightly between frames so that the hot pixels land on different sky and a sigma-clipped stack rejects them. If the version with darks is cleaner, or you are stacking star trails, keep shooting darks. Section 9 explains why trails are the case where darks still earn their place.
“Dark frame subtraction” by Spigget, adapted (resized, recompressed), via Wikimedia Commons, licensed under CC BY-SA 4.0. This adaptation is offered under the same licence.
If your camera holds its sensor at a set temperature — a cooled astronomy camera does, a DSLR does not — you can shoot darks once and reuse them for months at that temperature, exposure and gain. Sources disagree on how long: some say every few weeks, some every three or four months, some every six. All agree a library is for cooled cameras only; for a DSLR or mirrorless body in the field, the sensor’s temperature is the night’s, and so are the darks.
A flat frame is a photograph of an evenly lit, featureless surface through the same lens at the same focus, aperture and orientation as the lights. It records what the lens does to even light: darker corners, because a wide lens delivers less light off-axis; and soft dark discs, the out-of-focus shadows of dust on the filter stack over the sensor. Dividing every light by a good flat makes the sky the same brightness at the corner as at the centre and erases the dust. For a deep-sky stack, where faint nebulosity is being recovered across the whole frame, this is the difference between a picture and a picture with a bright hole in the middle.
A bias frame is the shortest exposure the camera can make, with the cap on, at the lights’ ISO. In the few microseconds the shutter is open nothing thermal accumulates and no light arrives; what remains is the offset the read-out electronics add to every pixel, plus the read noise around it. Every other frame — light, dark, flat — contains this same offset. Bias is how the flats get it taken out; the darks, shot at the lights’ own exposure, already contain it and take it out of the lights on their own.
Because bias depends on the read-out and not on time or temperature, it is the one calibration frame that keeps. A master bias shot in July at ISO 320 calibrates flats in December at ISO 320. That is also why a bias frame is the right thing to measure a camera with, and why the twenty-five below are the evidence for the rest of this page.
One bias frame from this camera has a read noise of 3.14 counts; the median of twenty-five has 0.80 and the mean 0.65, against 0.63 predicted by the square root of twenty-five. Counts here are 14-bit sensor units after the black point of 512 is removed. The mean lands on the prediction because read noise is random and averages exactly as the textbook says; the median is a little noisier than the mean, which is the price of its refusing to be pulled by an outlier. Twenty-five frames took under a minute to shoot.
A dark-flat is a dark shot at the flat’s exposure and ISO: cap on, same fraction of a second the flats used. It exists because of a quirk of some CMOS sensors: at the very short exposures a bias frame uses, the sensor’s response is not quite linear, so a bias frame does not measure the offset that the flats actually contain. A dark-flat, being shot at the flats’ own exposure, contains exactly what the flats contain apart from the light, and calibrates them more faithfully. Several guides now recommend dark-flats over bias for any CMOS camera, particularly when the flats run longer than about a second.
Two cautions. First, the term is used two ways: the majority of sources mean what this section means; at least one widely read page uses “dark-flat” for ordinary darks matched to the lights, which is a different thing. Second, “CMOS needs dark-flats” is a rule of thumb, not a law. The one source read for this page that actually measured it, a thread of working photometrists, found the effect real and large on one camera and negligible on another at its short flat exposures. If you want to know about yours, shoot both a master bias and a master dark-flat once and compare them: if they agree to within a couple of counts, bias will do.
You never subtract a single dark or divide by a single flat. You combine many of each into a master, and the reason is the figure in section 5: each calibration frame carries its own random noise, and subtracting or dividing by one frame stamps that noise onto every light. Averaging N frames reduces the random part by √N while keeping the fixed part — the pattern you actually want to remove — intact. Ten frames cut the added noise to about a third; twenty-five to a fifth; a hundred to a tenth. One source quantifies the residual as roughly 10 % with three or four frames, 5 % with six to eight, and 1 % with twenty to thirty, which is where every recommendation of “twenty or more” comes from.
The combining method matters less than the count, but not nothing. A plain mean is the most efficient at reducing noise and the most vulnerable to a single odd frame. A median ignores odd frames and costs about a quarter more noise, as the numbers above show. A sigma-clipped mean does both: it throws out values that are far from their neighbours, then averages the rest. That is what the Calibration Lab does, and why it refuses to sigma-clip fewer than eight frames: below that, the statistics cannot tell an outlier from the noise.
“Dark Frame” by Rawastrodata, adapted (recompressed), via Wikimedia Commons, used under the CC BY-SA 3.0 licensing option. This adaptation is offered under the same licence.
For every light frame:
calibrated light = (light − master dark) ÷ (master flat − master bias)
with a master dark-flat in place of the master bias where section 6 applies, and the divisor normalised so that dividing does not change the overall brightness. The subtraction on the left removes what the sensor added in the dark; the subtraction on the right removes the same offset from the flat so that the flat is a picture of the lens alone; the division removes what the lens did to even light. Because the dark was shot at the lights’ exposure it already contains the bias, so the bias is not subtracted from the light separately — doing so would remove it twice. That is the way Siril’s documentation writes it, and the way every serious stacker computes it; the differences between programs are in how the masters are built and normalised, not in the equation.
The order is fixed and the calibration happens before any alignment or stacking. A frame that has been aligned has had its pixels moved; a flat is a statement about pixels in their sensor positions.
Everything above is written, in the literature, for deep-sky stacking: many frames averaged so that faint, extended structure emerges across the whole field. Star trails are combined the opposite way — the brightest value at each pixel from any frame, a lighten blend — and a single nightscape frame is not combined at all. No source read for this page addresses which calibration frames those two cases need. The arithmetic does.
What this site accepts: the stacker’s star-trail, nightscape and deep-sky modes take optional matching RAW darks, flats, bias and dark-flats alongside the lights, and the Calibration Lab builds reusable masters from RAW groups and refuses to mix frames whose exposure, ISO or dimensions disagree. It will tell you by name when a frame does not belong to the group.
The sky moves 15.041 arcseconds a second at the celestial equator. How many pixels that is depends on two numbers: how many arcseconds one pixel covers, which is the plate scale, and how long the shutter is open. Plate scale is 206,265 times the pixel pitch divided by the focal length. Every “rule” for the longest exposure before stars trail is an attempt to skip that arithmetic with one constant.
| Focal length | 500 rule | 600 rule | 400 rule | NPF rule at f/2.8 | Arcsec per pixel | Drift per second |
|---|---|---|---|---|---|---|
| 14 mm | 35.7 s | 42.9 s | 28.6 s | 19.8 s | 88.0″ | 0.17 px |
| 16 mm | 31.3 s | 37.5 s | 25.0 s | 17.3 s | 77.0″ | 0.20 px |
| 20 mm | 25.0 s | 30.0 s | 20.0 s | 13.9 s | 61.6″ | 0.24 px |
| 24 mm | 20.8 s | 25.0 s | 16.7 s | 11.5 s | 51.3″ | 0.29 px |
| 35 mm | 14.3 s | 17.1 s | 11.4 s | 7.9 s | 35.2″ | 0.43 px |
| 50 mm | 10.0 s | 12.0 s | 8.0 s | 5.5 s | 24.6″ | 0.61 px |
Scroll the table sideways for the remaining columns.
| Focal length | 500 rule | 600 rule | 400 rule | NPF rule at f/2.8 | Arcsec per pixel | Drift per second |
|---|---|---|---|---|---|---|
| 14 mm | 23.8 s | 28.6 s | 19.0 s | 15.4 s | 57.8″ | 0.26 px |
| 16 mm | 20.8 s | 25.0 s | 16.7 s | 13.5 s | 50.5″ | 0.30 px |
| 20 mm | 16.7 s | 20.0 s | 13.3 s | 10.8 s | 40.4″ | 0.37 px |
| 24 mm | 13.9 s | 16.7 s | 11.1 s | 9.0 s | 33.7″ | 0.45 px |
| 35 mm | 9.5 s | 11.4 s | 7.6 s | 6.2 s | 23.1″ | 0.65 px |
| 50 mm | 6.7 s | 8.0 s | 5.3 s | 4.3 s | 16.2″ | 0.93 px |
Scroll the table sideways for the remaining columns.
Computed from the rules’ published formulas: the 500, 600 and 400 rules as seconds = constant ÷ (focal length × crop factor), from Scantips, which also gives the sidereal rate and shows the 600 rule costs about a quarter of a stop of sharpness for a quarter of a stop more light; the NPF rule as seconds = (35 × f-number + 30 × pixel pitch in microns) ÷ focal length, from Frédéric Michaud’s formula as PetaPixel published it and as AstroBackyard works it; plate scale from Clarkvision. The inputs, outputs and formulas are in this site’s receipt for the table, and two cells were checked against the sources’ own worked examples.
The 500 rule gives 20.8 seconds at 24 millimetres on full frame; the NPF rule at f/2.8 and a 5.97-micron pixel gives 11.5. That gap is the whole argument about the rule: 500 was tuned for film grain, and a modern sensor resolves the drift the film never saw. Which one you follow depends on what you will do with the frame. At a 100 % crop or a large print, NPF is right and 500 leaves visible dashes. For a frame that will be viewed whole on a screen, 500 is a fine compromise, and for star trails you want the opposite of both — the 500-rule page makes that case, and why the dash length depends on where in the sky you point as well as how long you expose.
The frame at the top of this page is 24 mm at 20 seconds on a sensor with 50.79″ per pixel: the sky moved 5.9 pixels during the exposure. That is invisible at page size and a short dash at 100 %, which is exactly what the two rules disagree about.
“Sunflower field Milky Way” by Juliancolton, adapted (resized, recompressed), via Wikimedia Commons, licensed under CC BY-SA 4.0. This adaptation is offered under the same licence.
The exposure rules fix the shutter; the aperture is as wide as the lens allows; ISO is what is left. Two independent analyses, Clark’s and Lonely Speck’s, arrive at the same place: ISO does not make the sensor more sensitive, it sets how the counts are scaled before they are written, and the noise you see at high ISO is the shot noise of too few photons, not something the ISO dial added. On many cameras the dial barely matters above a certain value; on some, the high-ISO frame is genuinely cleaner in the dark. The practical advice across sources is a range of 1600 to 6400 and a target on the histogram rather than a number: the sky’s hump separated from the left edge, roughly a quarter to a third of the way across. The frame at the top of this page is ISO 5000.
Wide and fast: 14 to 24 mm at f/1.4 to f/2.8 is the consensus across every lens round-up read for this page, for one reason, light. An f/1.4 lens gathers four times the light of an f/2.8 lens in the same exposure, which at night is the difference between a Milky Way and a hint of one. The price is what fast wide lenses do to a star at the edge of the frame.
Stopping down helps, and by how much is the one lens question where the tutorials genuinely disagree. One widely cited test says coma only “sufficiently diminishes” by f/5.6 — two stops and more from wide open, at which point you have thrown away the light you bought the lens for. Others, testing specific good lenses, find one stop does it: a review of a 14 mm f/1.8 found coma gone by f/2.8, and a working nightscape photographer’s notes recommend a third to a full stop on the lenses he actually uses. The honest answer is that it is your lens’s number, not a rule’s: shoot the same star field at f/1.4, f/2 and f/2.8, look at a corner at 100 %, and stop where you stop minding. The page on coma explains what stacking can and cannot do about it, and the lens page shows what published measurements say about yours.
Focus on a bright star with the live view magnified, not on the infinity mark, which on most lenses is a region rather than a point and moves with temperature. Do it before the calibration frames, and do not touch it after.
Fifty-six sources were read for this page and its sister on shooting, and on most of it they agree. These are the places they do not, with both sides named, so that when you meet one of them in another guide you know which argument you are reading.
| Question | One side | The other | What decides it |
|---|---|---|---|
| Are darks necessary on a modern sensor? | Every how-to guide: yes, always, twenty or more | Clark, and a quoted measurement by Craig Stark on a low-dark-current CCD: subtraction can add more noise than it removes | Your sensor’s dark current; test with and without |
| Scale darks to a different exposure, or match exactly? | Match exactly, especially on CMOS: Siril, most guides, the AAVSO manual | Scaled darks plus bias tested identical on one CMOS camera in a 352-image photometric comparison | Camera-specific; matching is always safe |
| Bias or dark-flats for calibrating flats? | Bias is fine on CCDs with stable short exposures | Dark-flats on CMOS, especially for flats over a second | Measure both once; the one thread with data found it varied by camera |
| How bright a flat? | Around 51 %, give or take five | Anywhere from 30 to 70 %; “not above 80” | The overlap, 30 to 50 %, works everywhere; state it as a percentage |
| What does “dark-flat” mean? | A dark at the flat’s exposure, replacing bias (most sources) | An ordinary matched dark for the lights (one encyclopaedia article) | Use the majority meaning; this page does |
| Do star trails need calibration at all? | Optional, a refinement (the tools that stack them) | Recommended without qualification (guides written for deep sky) | Section 9: darks yes, flats rarely, bias only with flats |
Scroll the table sideways for the remaining columns.
| Question | One side | The other | What decides it |
|---|---|---|---|
| 500 rule or NPF? | NPF; 500 is “very inaccurate today” and gives twice the exposure | 500 is adequate for frames viewed whole; NPF was written for pixel-peeped nightscapes | How the frame will be viewed |
| The 600 rule | Wrong: you will get a small trail | A quantified compromise: a quarter-stop more light for a quarter-stop of softness | Whether you accept the trade knowingly |
| ISO: high, invariant, or low for colour? | High ISO adds no noise and on some bodies removes it | Keep under 1600 or lose star colour | Different goals: noise versus saturated star colour |
| Wide open or stopped down? | Manufacturers: shoot wide open, coma is controlled | Stop to f/5.6 before coma sufficiently diminishes (one test); one stop is enough on a good lens (another) | Your lens, at a corner, at 100 % |
Scroll the table sideways for the remaining columns.
Usually for a stack, not always. Darks remove hot pixels and thermal signal, and a lighten-blend star trail keeps every hot pixel perfectly, so they help there. On a sensor with on-chip dark-current suppression, subtracting darks can add more random noise than the pattern it removes; test your own camera with and without.
Twenty to thirty of each is where the returns flatten: the noise a master adds falls with the square root of the count, so ten frames cut it to a third, twenty-five to a fifth. Bias frames are cheap and fifty is common. Fewer than eight of anything is a poor master.
Yes, within a few degrees, and at the same exposure and ISO. Thermal signal rises steeply with sensor temperature, so a dark library only works for a cooled camera that holds a set point. Shoot darks in the field, in the same hour, with the cap on.
One hump roughly a third to half of the way across, never touching either edge. Sources put the target between 30 and 50 percent of full scale; a raw-count figure only means something at a stated bit depth. Keep the focus, aperture and camera rotation exactly as they were for the lights.
Yes, at the same ISO. Bias records the sensor’s read-out offset, which does not depend on temperature or exposure, so a master bias keeps for months. Darks do not travel like that.
A dark shot at the flat’s own exposure and ISO. Some CMOS sensors are not linear at the very short exposures a bias frame uses, so a dark-flat calibrates the flats more faithfully, especially for flats longer than about a second. It is camera-specific: test, rather than assume.
Darks help, flats rarely matter, bias is pointless without flats. A lighten blend keeps the brightest value at every pixel, and a hot pixel sits in the same place in every frame, so the finished trail picture is peppered with sharp fixed specks that no trail runs through; a master dark removes them. Vignetting hardly shows on bright point subjects and can be corrected afterwards.
Only if nothing has moved: same lens, same focus, same aperture, same camera rotation, no lens change in between. Any of those invalidates the flat. Sky flats at dawn straight after the session are the safe habit.
