First nights from a remote Seestar S50 at a dark-sky observatory. The square stars were our stacker's fault, not the telescope's.

Comet 161P/Hartley-IRAS: hold the comet still, not the stars

On the evening of 9 October the Outpost followed comet 161P/Hartley-IRAS for about an hour, starting in evening twilight. A comet moves against the stars: this one moved about 368″ per hour, so over the hour it slid 157 stack pixels across the frame. Stack on the stars and the comet smears. So we stacked a second time on the comet itself, shifting every sub by where JPL Horizons, NASA's ephemeris service, says the comet was at that moment, and gave zero weight to each sub's pixels that sat on a star, so the stars don't smear into lines instead.

What it shows: a compact, round coma (the comet's cloud of gas and dust) with a brighter condensation at its center, and no tail visible at this depth, which is what we expected for a small, faint comet in a 50 mm telescope. Two checks: the coma's peak sits within 1 pixel of where the ephemeris puts it, and its peak above the sky is more than twice what it is in the star-held stack (about 69 against 30 counts, in green), because its light is no longer spread along a smear. Leaving the star pixels out did not change the coma (peak within 1 %). Near the comet, each stack pixel still has at least 32 and typically 67 of the 185 subs behind it (the drizzle drops are small, so even without the mask it is about 70).

What went wrong: square stars, and they were ours

The first stacks from these nights had square star cores, dark rings and black holes in the brightest stars. That looked like a bad camera or bad optics. The squares, rings and holes were neither: they were our own stacking code.

Old stacker

New stacker

A Seestar's sensor has a color filter over each pixel, in a repeating 2 × 2 pattern of red, two greens and blue. Our old stacker separated the four colors into four quarter-size images and resampled each one with a smooth curve. Star images here are only about two of those quarter-size pixels wide, too sharp for that curve: it overshoots and leaves dark rings, and its square shape stays in the star. On an alt-az mount the field turns and blurs the squares away. Here it never did, because the mount runs in equatorial mode and the field held the same angle all night. The black holes came from a third choice: the old stacker threw away every saturated (overexposed) pixel.

The new stacker drops each colored pixel where it actually sits on the sky, as a small round drop (a method called Bayer drizzle), keeps saturated pixels and flags them, and rejects trailed subs. The result: round cores, no holes, no rings, and star brightnesses within 1.5 % of the old stacks. Part of the gain in width comes from dropping the trailed subs: in M 35 the new method alone took stars from 6.77″ to 5.46″, and rejecting trailed subs to 5.20″.

How much light it collects: a healthy unit

To compare telescopes we use the photometric zero point: how bright a star must be, in Gaia catalog magnitudes, to give one count per second. Measured on every sub against Gaia stars and compared with the same measurement on 200 subs of Sh2-129 from a community S50 (another owner's unit), it says how much starlight the whole system, including the air above each site, turns into signal. The community frames were taken through the light-pollution filter, so only our sets through that filter compare like for like. We compare green only: red reads 1.1 to 1.3 and blue 0.8 to 0.9, because the camera rescales red and blue before saving (see inside the subs).

An earlier, quicker check had put the Outpost at about half the community unit (0.54), which would have meant dew, focus or a filter problem. It came from the first 13 minutes of the first night, the only light-pollution-filter data on hand then, when the zero point was wandering by ±0.14 mag, for a reason we haven't pinned down (haze or late twilight are candidates). The rest of both nights reads 0.85 to 0.93. Two units are not a fleet: this says the Outpost is not the weak unit the early check suggested, not how S50s compare in general.

10 s or 30 s subs? Little difference per hour

A longer sub helps when the camera's read noise (the noise added each time a frame is read out) matters, because an hour of short subs pays it more often. We measured this S50's read noise at 1.0 electron, which is very low. Through the light-pollution filter the noise in a 10 s sub is mostly read noise and the warm sensor's own dark current, not sky; dark current grows with exposure just as sky does, so 10 s subs give up little. From the noise budget we estimate about 0.15 mag per hour against 30 s (0.1 to 0.2 across these sets), which is smaller than this comparison across different fields and nights can resolve. (The only same-field 10 s set, 10 minutes of Sh2-129 in late twilight, reached 22.5 per hour.)

Faintest smooth glow detected at 3 sigma over 10″ boxes, scaled to one hour of exposure, in magnitudes per square arcsecond. It comes from each stack's two independent halves, in an instrumental blue-green band tied to Gaia, so it compares our own sets with each other; it is not a standard sky brightness. Measured on the first stacks, with one method for all three. The fields and nights differ, so this is a like-for-like check of sub length only within that limit. For stars, the new stacks reach a signal-to-noise ratio of 5 at Gaia G 18.9 (M 31), 18.1 (M 35) and 17.4 (IC 1805, 10 s subs through the filter), measured at the stars' catalog positions.

For an S50 owner: on a dark site, 10 s subs cost little depth per hour and lose less to a gust of wind or a tracking hiccup.

Trailed subs, mostly on the first night

A trailed sub has stars stretched into short dashes because the telescope moved during the exposure. The new stacker measures every sub's star shape at full resolution and rejects any sub whose stars are more than 1.4 times longer than wide. Most trailed subs were stretched along the sensor's rows (76 to 96 % of them within 20° of the rows), which points at the mount's tracking or wind rather than the optics.

Also rejected as trailed: M 35 64 of 234 and Sh2-129 87 of 316 on 8 October, the comet field 46 of 261 on 9 October. A handful more were rejected by the older, coarser shape rule. Separately, the optics stretch stars slightly on the sensor's right side (about 1.4 to 1.55 times longer than wide there), and with the field fixed on the sensor that stays in the stack.

Inside the subs: the Seestar processes them before saving

A Seestar sub is not a raw camera frame, and that matters only if you want to measure with it. From the Seestar subs alone, with our own dark and bias frames from the same camera to compare, we infer three things. The first two look like deliberate processing; the third may be the sensor itself rather than processing.

1. Dark-subtracted. The camera's own dark frame is already taken off, so the sky sits below the camera's fixed offset (its bias, the level a zero-light pixel reads). That is why the absolute sky brightness cannot be read from these subs: the zero point of the dark subtraction moves from session to session.
2. Red and blue rescaled. At the same signal, the red and blue pixels' noise is 1.5 to 2.5 times the green pixels', which is what a color-balancing scale factor does.
3. Light shifted between even and odd rows. A star's brightness in each green pixel set swings by about ±0.18 mag with the star's exact vertical position on the sensor. The light moves between rows, so what one green set loses the other gains, and adding the two cancels it: bright stars then repeat to about 0.03 mag.

The first two are reasonable choices for a smart telescope whose job is a clean, color-balanced live picture. For measurement it means: add both green pixel sets for star brightnesses, don't trust a single color set in a single sub to better than about 8 to 20 %, and don't read the sky's absolute brightness from the subs. On 10 October the Outpost was set to take raw frames through the camera's Alpaca interface (a standard remote-control interface) beside its own subs, field by field; they are not yet landed or analyzed, so the three findings above still wait for that direct check.

What is still open

How the pictures were made

One recipe for every target, in a script committed with this page. It only changes how the data are displayed; the stacks themselves are untouched.

1. Take the stack's red, green (the mean of the two green sets) and blue images as stacked. Flip so north is up and east is left.
2. Trim to the rectangle where nearly every pixel (97 % per edge row and column) has at least a quarter of the full sub coverage, then center-crop to a 9:16 shape. The few thinly covered pixels left inside are painted at the sky level.
3. For each color, subtract one constant: the sky level, taken as the 10th percentile of 32-pixel block medians.
4. Balance the colors on the stars: one gain each for red and blue so that the median unsaturated star comes out neutral. That makes the field's average star white, not a Sun-like star.
5. Stretch every color with the same inverse hyperbolic sine curve: soft at 4 times the stack's own noise (measured from its two independent halves), white at the 99.95th percentile, and the sky lifted a little above black so its noise is not clipped.
6. Resize to 900 × 1600 pixels; save as JPEG at quality 85 with no metadata.

The comet-held stack has no stars to balance or set white on, so it uses the star-held stack's color gains and stretch: the two comet views share one stretch. No noise reduction, no sharpening, no star reduction, no filling of star cores, no local contrast. The close-ups are the same pixels enlarged without smoothing. No flat frames were applied to the stacks (none exist for this unit yet).

What we expected, and what we found

Written down before any new pixel was opened:

• Missed Throughput. Expected 0.55 ± 0.1 of the community S50, as the early check suggested. Found 0.85 to 0.93; the early number was the first 13 minutes.
• Not testable Sky. Expected 1.2 to 1.6 mag darker than the community site. The in-camera dark subtraction hides the absolute sky level.
• Missed Sub length. Expected 30 s subs to go 0.2 to 0.4 mag deeper per hour than 10 s. The comparison found no difference it can resolve, and the noise budget puts it at about 0.15 mag; the read noise is lower than we had assumed.
• Held, then more Sharpness. Expected star widths of 4.8 to 6.2″ in the subs. The subs' cores were that sharp or sharper; the old stacks were not, which led to the stacker fix.
• Missed Star shape. Expected few elongated subs except in wind. On the first night 23 to 37 % of each target's subs were trailed.
• Held Comet. Expected a soft coma and no visible tail in a comet-held stack built from the ephemeris. That is what came out.

Fixed in advance, and measured

Chosen on these nights, not in advance: the new stacker's settings, including the 1.4 trailing limit, and the picture recipe, which was set by looking at these stacks. No night other than 8 and 9 October was used for any claim on this page, apart from bias and dark frames taken with the roof closed on 10 October.

Data

Our own Seestar S50 at a remote dark-sky observatory, on its built-in mount in equatorial mode. Night of 8 October 2026: Sh2-129 (light-pollution filter), M 31, M 33 and M 35 (UV/IR-cut filter). Night of 9 October: comet 161P (UV/IR-cut filter), NGC 6992 and IC 1805 (light-pollution filter). 4,202 subs of 2 to 30 s, 2.375″ per stack pixel. The M 31 frames are held back for a separate, pre-registered test on Hubble's Cepheid variable V1, so nothing here measures variable stars in them. Comparison: 200 light-pollution-filter subs of Sh2-129 from one community S50; none of its pixels are shown.

Star positions and magnitudes: ESA Gaia DR3. Comet ephemeris: JPL Horizons, NASA Jet Propulsion Laboratory.