Tutorial 18: A Seestar S50 session, from install to a finished picture¶
Goal: take 184 raw sub-frames from a Seestar S50 to a stacked, flattened, plate-solved, colour-calibrated, stretched and sharpened picture of the open cluster M 34, with every command run once and every number below pasted from that run. It is the whole chain of tutorials 3 to 9 and 16 on one fresh capture, so you can see what a normal session looks like before you meet a difficult one in tutorial 19.
Data: 184 sub-frames of M 34 (10 s, IR-cut filter, gain 80) from one night, 1080 by 1920
pixels, raw Bayer data in GRBG order, 2.375 arc seconds per pixel. They sit in a folder called
lights/, exactly as the Seestar wrote them: each .fit file has a .jpg and a _thn.jpg beside
it, and Valerian ignores the pictures. A Seestar writes no darks, flats or bias frames, so there is
nothing to calibrate and the chain starts at the stack.
Step 1: Check the install¶
valerian 0.4.0
commit: 69f25de3
built: 2026-10-25 20:06 UTC
profile: release
target: x86_64-unknown-linux-gnu
rust: rustc 1.99.1 (58cf8f9c3 2026-10-13)
wgpu: 30.0.1
catalogue: ~/valerian/gaia-dr3-g11.vcat (1247240 stars)
The catalogue: line must name a file: plate solving and colour calibration need it. If it says
not found, see tutorial 1 and the
troubleshooting page.
Step 2: Look at one frame¶
lights/Light_M 34_10.0s_IRCUT_20250925-020117.fit
size: 1080 x 1920 pixels, 1 plane (raw Bayer mosaic)
OBJECT M 34
IMAGETYP Light
FILTER IRCUT
EXPTIME 10
EXPOSURE 10
CCD-TEMP 27.0625
GAIN 80
XBINNING 1
BAYERPAT GRBG
DATE-OBS 2025-09-24T23:01:02.269863
INSTRUME Seestar S50
TELESCOP S50_xxxxxxxx
FOCALLEN 250
plane 0: min 497.0, sky 4904.0, noise 1576.01, max 65535.0 ADU
stars: 121 (median FWHM 3.60 px, ellipticity 0.556)
1 plane (raw Bayer mosaic)andBAYERPAT GRBG. The frame is the sensor's raw mosaic: one number per pixel, with the colour filters in a green-red / blue-green pattern.stackdemosaics it for you (the choices are in tutorial 3); you do not convert anything first.sky 4904. This is the very first frame of the night and the sky is still twilight-bright. The stack of Step 3 shows the same thing: a sky level of 4,550 for the first frames and 630 later on, and a tiny weight for the bright ones.TELESCOP. The frame carries a code that is different for every telescope; it is shown here asxxxxxxxx. Valerian removes it, and the observing site, from every file it writes (see Privacy).
Step 3: The first stack¶
compute: GPU: NVIDIA A100 80GB PCIe (Vulkan, DiscreteGpu)
frame 0: 336 stars, FWHM 3.76 px, ellipticity 0.526, sky 4549.8, noise 212.91
frame 1: 353 stars, FWHM 3.05 px, ellipticity 0.398, sky 4548.7, noise 213.79
...
frame 29: 1000 stars, FWHM 2.20 px, ellipticity 0.176, sky 632.6, noise 45.61
...
reference: frame 29 (FWHM 2.20 px)
combining 184 frames with LinearFit { low: 5.0, high: 2.5 }, weights NoiseFwhm on GPU: NVIDIA A100 80GB PCIe (Vulkan, DiscreteGpu), streamed in bands, budget 4.0 GiB (4.3 GiB staged)
GPU: 549 calls ran on the device, 3 fell back to the CPU (184 frames is more than the 128 the kernel holds)
# frame stars FWHM ellip sky noise flux x rms px weight reject status
0 .._IRCUT_20250925-020117.fit 336 3.76 0.526 4549.8 212.91 0.8975 0.369 0.04 80.50% stacked
5 .._IRCUT_20250925-020222.fit 565 2.58 0.278 1705.0 117.06 0.9741 0.311 0.22 54.41% stacked
6 .._IRCUT_20250925-020233.fit 1000 2.31 0.190 643.8 46.86 1.0074 0.285 1.61 5.38% stacked
29 .._IRCUT_20250925-020913.fit 1000 2.20 0.176 632.6 45.61 1.0000 0.000 1.90 6.74% reference
30 .._IRCUT_20250925-020925.fit 1000 2.84 0.294 632.9 45.68 0.9726 0.341 1.20 4.18% stacked
...
183 .._IRCUT_20250925-025950.fit 1000 2.44 0.194 629.6 45.36 1.0104 0.290 1.53 5.00% stacked
184 of 184 frames stacked with LinearFit { low: 5.0, high: 2.5 }, reference frame 29 (Light_M 34_10.0s_IRCUT_20250925-020913.fit), 35.1s
wrote stack.fits
wrote coverage.fits
wrote frames.txt
The table in frames.txt has all 184 rows; the lines above are the ones worth reading. This run took
33 s of wall clock and 2.1 GB of memory at its peak, on a shared 24-core machine with a load of 18 to 19;
your time depends on your computer (tutorial 19 times the stack on the
processor, on the graphics card and with different memory limits).
- Frames 0 to 5 are dusk frames. Their sky is 4,550 and 1,705 ADU where the rest is 630 to 645, so the
stacker gave them weights of 0.04 to 0.22 (the typical frame is about 1) and the rejection threw out
54 to 82 per cent of their pixels. Nothing was left out by hand: a bright sky is noise, and the weights
already say so. The frame limits (
--min-snr-weightand the others, in tutorial 12) leave such frames out altogether. - The reference is frame 29. It is the sharpest frame that the others register against (FWHM 2.20 pixels, and 1,000 stars, the most the detector keeps).
3 fell back to the CPU (184 frames is more than the 128 the kernel holds). The graphics card combines at most 128 frames at once. With more, it does what it can and the processor does the rest; the result is the same.
stack.fits
size: 1080 x 1920 pixels, 3 planes
OBJECT M 34
IMAGETYP Master Light
FILTER IRCUT
EXPTIME 10
EXPOSURE 10
CCD-TEMP 28.5
GAIN 80
XBINNING 1
DATE-OBS 2025-09-24T23:08:58.552260
INSTRUME Seestar S50
TELESCOP S50
FOCALLEN 250
NCOMBINE 184
EXPTOTAL 1840
plane 0: min 450.6, sky 610.4, noise 5.08, max 70969.2 ADU
plane 1: min 602.5, sky 665.4, noise 4.87, max 74961.2 ADU
plane 2: min 89.4, sky 612.8, noise 4.99, max 72013.4 ADU
stars: 1000 (median FWHM 2.77 px, ellipticity 0.104)
Three planes (red, green, blue), 184 frames and 1,840 s of light. The noise is 4.9 to 5.1 ADU in each plane where the reference frame has 45.6: about nine times less (the square root of 184 would give 13.6; the weights, the dusk frames and the resampling account for the difference). The stack is linear. Shown with a plain automatic stretch it is green, because the sensor has twice as many green pixels and the three planes have different sky levels.
Step 4: Take out the sky gradient¶
plane 0: 1135 samples of 1193 tiles (33 px), 24.5% masked, residual rms 0.35, gradient removed 3.9
plane 1: 1034 samples of 1093 tiles (33 px), 26.2% masked, residual rms 0.32, gradient removed 4.1
plane 2: 1045 samples of 1071 tiles (33 px), 22.6% masked, residual rms 0.33, gradient removed 2.4
wrote background.fits
wrote the background model model.fits
The model fits the clear sky between the stars (about a quarter of the tiles were masked as star or cluster) and takes it away: 3.9, 4.1 and 2.4 ADU of gradient across the picture, in a fit whose residual is 0.3 ADU. It ran in half a second. The models and the pictures are in tutorial 5.
Step 5: Plate solve¶
looking within 1.00 deg of RA 40.9125 Dec +42.9033, scale 2.03 to 2.75 arcsec/pixel
solved: centre RA 02h41m51.5s Dec +42d47m08.7s (40.46463, +42.78575 deg)
scale 2.375 arcsec/pixel, field 0.71 x 1.27 deg, rotation 72.09 deg east of north, mirrored
72 of 400 detected stars matched, rms 0.11 arcsec (0.05 px), plain TAN
wrote solved.fits
The hint (where the telescope said it was pointing, and its pixel scale) came from the header, so no
blind search was needed. 72 of the 400 brightest stars matched Gaia stars, to 0.11 arc seconds rms,
a twentieth of a pixel. The solution is written into solved.fits as a world coordinate system. See
tutorial 8.
Step 6: Colour calibration¶
26 Gaia stars measured, 22 used; aperture radius 5.6 px; neutral at BP-RP 0.82
red against green: slope -0.303 mag per BP-RP, scatter 0.012 mag, gain 1.3006 +/- 0.0031
blue against green: slope +0.451 mag per BP-RP, scatter 0.016 mag, gain 1.5844 +/- 0.0051
sky before 610.40 665.35 612.84, after 810.08 810.08 810.08 (offsets -16.17 -144.72 160.89)
wrote photocal-report/pcc.png
wrote photocal-report/pcc.csv
wrote photocal-report/pcc.json
wrote photocal-report/report.html
wrote the star table pcc_stars.csv
wrote pcc.fits
Valerian measured 26 stars that Gaia has colours for, used the 22 that were clean, and fitted how much brighter red and blue should be against green for a star of each colour. The gains, 1.30 for red and 1.58 for blue, are what turn the green stack into a colour-true one, and the sky of all three planes now sits at the same 810 ADU. The scatter, 0.012 and 0.016 magnitudes, is how far the 22 stars lie from the fitted line. How the fit works, and what its plots show, is tutorial 9. A field with far fewer stars needs two settings changed: that is the first thing in tutorial 19.
Step 7: Stretch¶
black 794.7897, white 37255.4414, Mtf { midtone: 0.003803931359042682 }
plane 0: median 810.3948 -> 0.1008
plane 1: median 810.3635 -> 0.1006
plane 2: median 810.4302 -> 0.1010
wrote stretched.fits
wrote preview stretched.png
The automatic stretch put the sky at a tenth of full brightness in all three planes. On the left, the
stack of Step 3 given the same kind of automatic stretch but none of Steps 4 to 6; on the right,
stretched.png:
Left: stack.fits stretched with valerian stretch. Right: stretched.png. The same 184 frames.
The stars are now the colours of their spectral types: the blue of the hot cluster members and the orange of the field giants. See tutorial 6 for the curves.
Step 8: Sharpen the stars¶
deconvolve runs on a linear image, so it takes pcc.fits, before the stretch:
1000 stars detected, 859 fit to measure a PSF (median FWHM 2.75 px)
deconvolving to round stars of FWHM 2.03 px
12 tile(s) of about 360 x 480 px; PSF FWHM 2.57 to 3.08 px, ellipticity 0.050 to 0.188
deconvolved 3 plane(s), 20 iterations, 12 tile(s), 14.0s
stars deconvolved to round, FWHM 2.03 px
wrote sharp.fits
wrote psf-tiles.csv
859 of the 1,000 stars were clean enough to measure a point spread function (PSF) in each of twelve tiles. The PSF is narrower in the middle (2.57 pixels) than in the corners (3.08) and more elongated towards the edge (ellipticity up to 0.19). Measure the stars before and after with the same tool:
998 usable stars: median FWHM 2.75 px, eccentricity 0.106
999 usable stars: median FWHM 2.09 px, eccentricity 0.063
(The first line of each; the zone tables and maps are in optics-before/ and optics-after/.) The
median star went from 2.75 to 2.09 pixels, 6.5 to 5.0 arc seconds, and from 0.106 to 0.063 in
eccentricity. The Seestar's own stack of the same 184 frames, measured the same way, is 5.4 arc seconds
and 0.042: the stars come out a little narrower than the Seestar's and not as round. The same stretch
of the centre of the cluster, magnified twofold, before (left) and after (right):
Look at the brightest stars on the right: they have a dark, cyan-tinged core. That is what
deconvolving bright, nearly saturated stars does, and the default 20 iterations show it most. With
--iterations 8 the median star is 2.20 pixels and 0.075 (instead of 2.09 and 0.063), the run takes
6.4 s instead of 14.0 s, and the cores are slightly shallower. Choose by eye on your own picture; if
the bright stars matter more than the faint ones, take the shorter run. Tutorial 16
has the settings.
The desktop app¶
The desktop app does Steps 3, 5, 6 and 7 and the measurement of Step 8 (the Optics tool of the Process
step) on the same frames. There is no panel for background or deconvolve yet, so those two stay on the
command line. The pictures are of the app working on the same 184 frames: it ran the stack itself.
Step 3, Stack, finished: "184 of 184 frames" in the step list, and the stack open in the Process step with its automatic stretch.
Plate solve: the same 72 stars, 0.11 arc seconds, and the same centre and rotation as valerian solve printed.
Colour: 22 of 26 stars, gains 1.3006 and 1.5844 as on the command line. The circled stars are the ones the fit used.
Optics: 998 stars, median FWHM 2.77 pixels and eccentricity 0.104, the same as valerian info and optics report for the stack. This is the measurement of Step 8; sharpening itself is command line only.
The reference for each panel is the app reference. The pictures were made by the
app's own test harness driving the real program on these frames (real_data_shots_seestar_m34).
What you have¶
| File | What it is |
|---|---|
stack.fits, coverage.fits, frames.txt |
the linear stack, the frames per pixel, the per-frame table |
background.fits, model.fits |
the stack with the gradient out, and the gradient |
solved.fits |
the same with its sky coordinates |
pcc.fits, pcc_stars.csv, photocal-report/ |
the colour-calibrated linear picture and the star fit |
stretched.fits, stretched.png |
the picture to look at |
sharp.fits, psf-tiles.csv |
the deconvolved linear picture and the PSF of each tile |
Troubleshooting¶
Every message you can meet in this chain, with its cause, is on the
troubleshooting page, by stage. The ones a Seestar user meets
most are the colour calibration of a sparse field (tutorial 19), a
mosaic folder (tutorial 17), and plate solving needs a demosaiced image, this one is a raw
colour mosaic, which solve and photocal print when given a raw sub-frame: stack first.
Where next¶
Tutorial 19 runs the chain on a faint galaxy, fixes the colour calibration of a sparse field, and times the stack on the graphics card, on the processor and with different memory limits.