My Seestar S50 quietly stopped applying its color gains

Some words first. A sub is one exposure as the Seestar saves it, before stacking. Its sensor sits under a Bayer matrix, a checkerboard of tiny color filters: every 2×2 block of pixels has one red, two green and one blue. Each pixel records a whole number, its brightness in ADU (analog-to-digital units, the camera's raw counting unit), from 0 up to 65,535, the largest number the file can hold.

Normally the Seestar applies color gains before it saves a sub: it multiplies the colors by different factors so that ordinary starlight comes out roughly white. One visible side effect, which I wrote about in why a Seestar's bright stars turn green or purple: green's values are scaled down against red and blue, so when a bright star's core fills up, green stops well below red and blue.

On my S50 that stopped. On the nights of 8 and 9 October, green's brightest pixel never went past 57,209 ADU in any of the 39 subs I tested, while red and blue reached about 65,000 (measured). From the night of 10 October on, wherever a star filled its green pixels, green stopped at the same value as red and blue, about 65,100 (measured). The Seestar's sensor counts in whole steps and the camera stores each step as 16 ADU, so an untouched sub's values come in steps of 16 (stated, from my earlier calibration of this camera). A color gain multiplies the values and breaks those steps. The unbalanced subs have the steps again; the balanced ones don't (measured, next chart).

The next chart measures those steps two ways: a comb strength, which says how strongly the pixel values repeat every 16 ADU (a histogram of them looks like the teeth of a comb), and the share of pixels sitting on the 16-step grid.

The S50's sensor delivers 12-bit codes, which the camera stores multiplied by 16 in a 16-bit FITS file, so an untouched plane only takes values 16 ADU apart (stated, from my earlier calibration of this camera; the unbalanced subs show exactly that spacing, measured). A per-plane gain applied after that, followed by rounding, moves values off the 16-step lattice. Two independent per-sub statistics test for the lattice:

• Comb strength. Per Bayer plane, the histogram of integer codes from the sky median + 30 to + 3,000 ADU (one bin per code), detrended by a 31-code running mean, Hann-windowed and Fourier-transformed. The statistic is the largest power within ±2 frequency bins of 1/16 cycles per code, divided by the median power over periods of 8 to 30 codes. Without a lattice it read 0 to 15 here; with one, tens to tens of thousands.
• Grid share. The fraction of a plane's pixels on its most common residue modulo 16. No lattice gives 1/16 = 0.0625; a perfect lattice gives 1. Dark subtraction and hot-pixel repair move some pixels off it, so unbalanced subs read 0.37 to 0.80, not 1.

I measured the comb on the first, middle and last sub of every group (night, target, filter, exposure): 39 subs from the 2026-10-08 and 10-09 nights, 33 from 10-10 to 10-12, and one dawn sub of the 10-13 night. The grid share comes from 13 subs spread over the 10-08 to 10-10 nights and 17 subs of the 10-13 night, mostly different subs from the comb's. Both statistics put the change at the same night boundary, in every plane (measured). These are samples: the 10-08 to 10-12 nights hold 7,237 subs (stated, from my frame catalog), and the comb was read on 72 of them.

Balanced subs: comb power 0 to 15 in every plane (39 subs, 156 plane values). Unbalanced: 37 to 3,938 on the 31 dark-sky subs of the 10-10 to 10-12 nights, and 27,835 to 93,979 on the three last-of-the-night subs taken in the dawn sky (10-10, 10-11 and 10-13 nights); the best period in the 8 to 30 band is 16.1 to 16.2 codes on all 33 of the 10-10 to 10-12 subs (measured; on the 10-13 dawn sub I read only the power at period 16). Grid share: 0.064 to 0.172 balanced (M 31, M 35, 161P and IC 1805 subs), 0.539 to 0.788 on the 10-10 WASP-12 field and 0.369 to 0.802 on the 10-13 night, lowest at dawn. A pooled version (20 subs per set, 600 to 8,000 ADU) agrees: the 10-08 M 31 and 10-09 IC 1805 sets show no period stronger than 17 times the band median in any plane, while the 10-10 WASP-12 field shows the 16.1-code period at 2,800 to 3,200 times it in every plane.

The power of the unbalanced comb grows with the number of pixels in the window (dense star fields, a bright dawn sky; inferred), so it tells the two kinds of sub apart but doesn't measure how much of a sub sits on the grid. The weakest unbalanced value, 37, is blue on the very first unbalanced sub. The 37 to 15 gap is wide enough to classify every sub here, but it was read off these subs, not fixed in advance.

One thing I expected and didn't see: a gain followed by rounding should leave a comb of its own at about 16 times the gain (about 12 to 14 ADU for green, depending on the filter; inferred from the ratios below). No period stands out in the balanced subs at all. Why that comb is washed out isn't established here.

Two details. The top isn't 65,535 but a little under it, about 64,950 to 65,320, different in every sub, most likely because the Seestar subtracts its dark frame after the sensor has filled (inferred; see the next section). And on a few subs taken at dawn or early in the evening, the top is the full 65,535 (measured).

The color balancing also depended on the filter. With the IR-cut filter green stopped at about 54,950 to 57,200; with the light-pollution filter, at about 50,400 to 52,100 (measured, on the subs where a green star filled).

Balanced subs: the sub's top (red and blue together) was 64,948 to 65,137 ADU, or 65,535 on 5 subs (the last M 35 sub, at dawn, and four 161P subs early in the evening); green's maximum was 33,257 to 57,209 (29,080 to 57,209 over all 39 balanced subs, the lower ones where no green pixel filled). By filter, green's maximum on those 33 subs was 54,950 to 57,209 with the IR-cut filter (18 subs) and 50,365 to 52,120 on 13 of the 15 light-pollution-filter subs (the other two, 33,257 and 43,125, sit far below that band, so no green pixel filled; inferred). Where green filled, its top over the sub's top was 0.845 to 0.879 with the IR-cut filter and 0.775 to 0.802 with the light-pollution filter (measured), which suggests a green gain of about that size against red and blue, set per filter (inferred).

Unbalanced subs: on every one of the 41 subs where a plane filled, green's maximum equals the sub's top, 65,065 to 65,318, or 65,535 on the two dawn subs (measured). On 11 of them red peaked at only 21,976 to 39,828, so this is green reaching the shared limit, not every plane filling at once.

App sub against raw Alpaca frames. The app sub at 01:30:27 UTC on 10-11 against five raw Alpaca frames of the same field from 01:33 UTC, both 10 s with the IR-cut filter. Star light above sky, app over raw, at each plane's 99.9th percentile: G1 0.962, R 1.005, B 1.028, G2 0.984 (measured). At the 99.99th percentile, where few pixels and the seeing dominate, 0.848 to 1.073. Either way no plane carries a factor like the balanced subs' 0.78 to 0.88.

Dark and hot pixels. The raw frames' 577 hot pixels (median excess 2,112 ADU over the raw sky) read a median 0.00 of that excess in the app sub. With a hot-pixel map from 20 s raw darks (152 pixels over 800 ADU, median excess 1,032 ADU), the app subs read −0.13 to 0.01 of the dark's excess on three balanced subs and 0.03 to 0.14 on two unbalanced ones (measured): still mostly removed. The app's top (65,091) sits 413 ADU under the raw frame's (65,504); a dark subtracted after the sensor fills would do that (inferred).

Flat field. The unbalanced subs keep the 16-step lattice, so no per-pixel non-integer factor is applied to them (inferred from the grid share). Whether the balanced subs were flat-fielded isn't tested here.

• Saturation thresholds must be per sub. A table of saturation levels looked up by telescope and firmware version would give an unbalanced sub of this telescope the balanced green ceiling (the bottom of green's full band, about 49,600 to 50,000 with the light-pollution filter and 53,500 to 53,700 with the IR-cut filter; stated, from a separate measurement of where green fills on these nights), so green pixels from there to about 65,100 that aren't full would be flagged as full (lost data; inferred). A table built the other way round would pass full green pixels on balanced subs as unsaturated (biased photometry; inferred). Decide from each sub's own pixels which kind it is, just as its top has to be read per sub.
• Don't pool across the change. IC 1805 (10-09 balanced, 10-10 unbalanced) and NGC 6992 (10-09 balanced, 10-12 unbalanced) were shot on both sides; a stack or a per-plane noise model that pools them gets two green ceilings and two color scalings. Per-plane conversion gains (ADU per electron) measured on one side shouldn't be reused on the other (inferred).
• Per-night, per-plane zero points absorb it. In a variable-star measurement on M 31 frames (the 10-08 and 10-10 nights) calibrated against Gaia per night and per plane, the leftover red-to-green scale after the zero points was 0.992 on the balanced night and 1.022 on the unbalanced one, and the plane shares of the combined flux moved by under one percentage point (measured). The step this predicts in the variable's flux between the two nights, with the variable at a template's per-plane flux and no data on the variable itself, is 0.1 % of a night's 1σ noise (computed; not an observed step).
• My own earlier wording was too broad. My write-up of this telescope's first two nights says the camera rescales red and blue before saving. That holds for those two nights (8 and 9 October), not for every night of this telescope.

1. Open a raw sub (not a stack, not debayered) in a program that can show the raw pixel values or a histogram you can zoom to single ADU.
2. Zoomed in on the faint part of the histogram, an unbalanced sub looks like a comb: tall bars every 16 ADU with low ones between. A balanced sub is a smooth hill (measured, on mine).
3. Or find a star bright enough to fill even its green pixels and read its core's values per color. If green stops at the same value as red and blue (on my S50, about 65,100 to 65,300), the gains were off. If green stops thousands of ADU lower, they were on (measured, on mine; other models and firmware may stop at other values).
4. Check one sub from each session, especially after a firmware update, a power cycle or any session where other software controlled the telescope.

The grid share is a few lines of Python. It assumes the S50's GRBG Bayer pattern (BAYERPAT); swap the offsets for another pattern.

import numpy as np
from astropy.io import fits

a = fits.getdata("sub.fit").astype(np.int64)
for name, (y, x) in {"G1": (0, 0), "R": (0, 1), "B": (1, 0), "G2": (1, 1)}.items():
    p = a[y::2, x::2].ravel()
    share = np.bincount(p % 16, minlength=16).max() / p.size
    # 1/16 = 0.0625 means no grid; mine read 0.064-0.172 balanced, 0.37-0.80 unbalanced
    print(name, round(share, 3))

On a sub that is mostly dark sky with few stars, the grid share is the quicker test; the comb strength separated the two states more widely here (0 to 15 against 37 and up) and holds up in bright dawn skies, where the grid share drops to about 0.37.