I subtracted 98,258 catalog stars from a Seestar image of the North America Nebula. The nebula stayed.

The stack is four layers added together

Instead of erasing stars, the method models each part of the picture on its own and subtracts the parts that aren't nebula. Choose a layer to see it by itself, on the full frame.

Your stackWhat the camera recorded: everything at once. = Stars98,258 Gaia stars at their catalog positions, each drawn with this night's star shape, plus a model of the brightest stars' broad halos. Only each star's brightness is fitted. + Sky glowA smooth gradient across the frame: 17.4, 16.7 and 18.8 ADU from end to end in red, green and blue. Shown with its own stretch so you can see it. + NoiseThe grain that noise reduction took out. Noise is random, so this is not a model of it, just what came out. How much to take out was measured against a null stack, and what came out checks as noise. Shown with a strong stretch: gray is zero, full scale is ±2 sigma of the null. + NebulaWhat is left: the stack minus the stars, the sky glow and the grain.

Clouds came in, so the stack was rebuilt

The first version of this page used every sub, added up as they came. But the night was clear and steady for only the first 1.8 hours. Then cloud moved in and the stars dimmed by up to 0.91 mag, and the cloud wasn't gray: green lost about 20 % more than red and blue. Because the field turns on an alt-az mount, different parts of the frame were built from different mixes of clear and cloudy subs, so the nebula's brightness and color carried that mix.

So I measured each sub's transparency, color by color, from about 80 Gaia stars per sub with known spectra, scaled every sub to the clearest stretch of the night, and kept only the subs that lost less than 0.5 mag to cloud and had sharp stars. The rules were written down before the transparency was measured.

Star brightness scatter: each star's measured flux against what its Gaia XP spectrum predicts through this filter, before and after the correction. The new stack's numbers are in its own units, set by the clearest part of the night, so on the nebula they run about 1.26 times the first stack's. For display, every image on this page is scaled back to the first version's brightness, so the pictures are comparable; no number on the page is scaled.

The brightest stars' halos, modeled instead of left behind

↔

A bright star's light doesn't stop at its core. Through this lens a broad halo, with a core radius of about 90 pixels, holds 7 to 18 % of it depending on color, and green and blue add a scattered-light wing. The first star model stopped at 60 pixels, so the brightest stars left a ring and a teal glow, and a G 3.1 star just outside the frame lit up the left edge.

The new model is physical: the lens's own diffraction core, the scattered-light wing and the broad halo, fitted jointly over the 53 brightest stars, with the saturated ones' brightness set from Gaia. Around 57 Cygni the green and blue excess at 40 to 120 pixels fell from +2.1 and +2.7 ADU to +0.3; the off-frame star's glow is gone; and the light red oval around 57 Cygni stays, because IPHAS shows it too. The saturated cores are filled: red from IPHAS, scaled to ours, where IPHAS covers them, and green and blue following red.

Up close, the texture behind the stars survives

Because each star is subtracted rather than painted over, the filaments running behind it are kept. In a test where I planted fake stars on top of real nebula and took them out again, 98 to 99 % of the nebula's light under them came back. The right-hand sides here are also noise-reduced; the original sides are untouched.

↔

↔

Taking out the grain, and checking what came out

After the stars and the sky glow, a noise-reduction pass takes out the grain. It is a wavelet method: the image is split into detail scales from 1 to 32 pixels, and each detail is kept in proportion to how far it stands above the noise at that spot. Nothing is set by eye. The noise level comes from a null stack: the night split into two halves and subtracted, which cancels the sky and the nebula and leaves only noise. It is scaled pixel by pixel by how many subs covered that pixel, so the thin edges get a higher bar. The pass is deterministic: a second run gives a bit-identical result.

These checks were measured on the first version's stack; the same denoiser runs unchanged on the new one. There, the corners, where fewer subs overlap, ended up nearly as clean as the middle: run through the same pass, the null stack's red noise dropped to 0.21 to 0.26 ADU in the four corners and 0.20 ADU in the center.

A second night sees the same gas

Night 1 (18 September, 195 × 10 s, about 32 minutes) went through the same steps as the first version and was mapped onto night 2's pixels through both nights' plate solutions. The two overlap on 59 % of night 2's frame. This check was made on the first version's stacks and has not been redone on the cloud-corrected one.

Spearman rank correlations on the overlap, 12-pixel blocks, first-version stacks. The three Seestar panels share one stretch and are shown binned 3 × 3.

How the stars come out

1. Correct each sub for cloud. Scale every sub, color by color, to the clearest part of the night, and drop the cloudiest and softest subs.
2. Place every catalog star. Each Gaia DR3 star lands at its known position through the frame's plate solution: here 98,258 stars down to G = 17. Fainter stars are not modeled and stay in the image.
3. Measure the star shape from the same night. The point-spread function comes from this night's own stars, per color, in zones across the field.
4. Fit one number per star, then subtract. Position and shape are fixed, so each star gets only a brightness. Subtracting a model keeps the books balanced: light the star model doesn't claim stays in the image.
5. Take the background under each star from the whole session. It comes from a starless template of the session rather than a small ring around the star, so a star sitting on bright nebula doesn't take the nebula with it.
6. Model the brightest stars' halos. Diffraction core, scattered-light wing and broad halo, with brightness from Gaia, including bright stars just outside the frame. Their saturated cores are filled.
7. Model the sky glow. A smooth quadratic surface per color, fitted together with a scaled copy of the IPHAS image so the gradient comes out without eating the faint outer nebula.
8. Take out the grain. The wavelet noise reduction described above, with its noise level set by the null stack, not by eye.

What isn't solved

What's still visible

What I tried that didn't help

• Deconvolution. Stars came out about 10 % sharper, for 26 to 37 % more grain. At this signal-to-noise the extra detail reads as sharpened noise, so it failed the rule I set before trying it.
• A direction-aware noise reduction meant to keep thin filaments sharper. It failed its own pre-written acceptance rule, so the simpler one stays.
• Subtracting the faint Gaia stars too (G 17 to 21), predicted from their catalog brightness without fitting. Their brightness was over-predicted by 16 to 20 %, which dug shallow holes; a scaled version was only a marginal gain.
• Looking for a sensor pattern behind the leftover bands and fringe. Field rotation spreads any sensor-fixed light so thinly that it can't explain them.
• Other per-sub background models (a plane, a fit restricted to well-covered areas, a joint fit with a sky-glow term). None beat the one used here.

Compared with neural star removers

Tools like StarNet and StarXTerminator are one click, and today they handle bright stars, and anything that isn't in a star catalog, better than this does. What the catalog approach adds is different: it is deterministic, every star it removes is a catalog star at a known position, the light it takes out is accounted for star by star, and the result can be checked against an independent survey, as above. The two could also work together.

What your processing software could make of it

Post-processed for looks An example of what you could do with this data in your own software (Siril, PixInsight, Photoshop). Not a measurement.

Everything above is shown plainly, with one stretch, so it can be checked. Here is the same data processed the way an imager would for a picture: the noise-reduced starless nebula in an HOO palette with a careful stretch, and the real stars put back, smaller.

↔ Plain stretchHOO, light hand

The steps, in imager terms

1. Background neutralization on the darkest dust, so the dark nebula comes out neutral.
2. Dual-band channel extraction: Hα from red, [O III] from green and blue. [O III] is boosted to 0.9 of Hα in the bright nebula.
3. HOO palette: Hα to red, [O III] to green and blue, so the Cygnus Wall goes red and the [O III] glow goes teal.
4. Color-preserving arcsinh stretch, then curves with a gentle S.
5. Gentle large-scale local contrast and light sharpening, soft-limited so nothing halos.
6. Chroma-only noise blur and saturation with a highlight roll-off; no nebula pixel clips.
7. Stars: a separate gentle stretch (star reduction), screened back at 92 %.

Cosmetic repair, for this picture only: small star-removal pits were filled up to the local median, never above it. The bright-star halos needed no cosmetic repair this time, because the halo model above already handled them. Apart from switching off the tool's own halo repair, the render settings are the same as in the first version; the layers were first scaled to the first version's brightness. No step paints, clones or uses a trained model, and the same input always gives the same output.

People often push this target much further: Foraxx dynamic palettes, synthetic SHO, CLAHE or HDR-style local contrast. I kept to a palette whose colors map to real emission lines (red is Hα, teal is [O III]) and to large-scale contrast only. Heavy local contrast on 1.5 hours of 30 mm data turns faint nebula into a cotton-ball texture that isn't in the sky.

As capturedThe stack at the page stretch, stars and all. Starless, denoisedStars, sky glow and grain out, same stretch. This is the data. Post-processedHOO, light hand, real stars reduced and screened back. This is taste.

The two heavier looks come from a separate editing pass on the first version's stack, with a slightly different framing. Neither is used anywhere else on this page.

The same idea on a galaxy: let the sky turn

On an alt-az mount like the Seestar's, the sky slowly rotates on the sensor through the night. Over one night of 291 subs on M 33 it turned 140°. Light that turns with the sky is real; light that stays put on the sensor, or changes from sub to sub, is not. One joint fit sorts the night into those layers.

The stackEverything together, including the bright wedges that don't turn with the sky. Fixed to the skyThe galaxy layer, with the stars fitted out only roughly so far: 16 to 21 % of a typical star's light (G 10.8 to 13) is still in it, hence the faint blotches. Split offLight fixed to the sensor, and light that changed from sub to sub, removed from the galaxy layer.

The same subtract-what's-known idea is behind two other workbench results that will get their own write-ups. Legacy Survey images of the Leo Triplet, pushed through a model of the Seestar's optics, registered to the Seestar frames to 0.12 px and matched M 66's spiral-arm phase to 1.7°. And a frame-by-frame model of community M 31 frames shows Hubble's Cepheid V1 at 3.2 sigma at its catalog position; whether it can be seen pulsing is not confirmed.

Fixed in advance, and measured

The post-processed look is neither fixed nor measured: its palette, stretch, contrast and star settings are taste, chosen to look good on this frame, and none of its numbers should be read as a result.

No fresh data was looked at for this page. A confirmation needs a night, and ideally a camera, that the method has never seen, chosen and named before the check is run.

Data and credits

My data: a Seestar S30 Pro with its dual-band light-pollution filter on its alt-az mount. NGC 7000 on 18 and 19 September 2026 (195 and 752 subs of 10 s; 532 of the 752 in the final stack) and one night on M 33 (291 subs). Orientation: north up, tilted 33.4°; 3.66″ per pixel. Every comparison uses one stretch for both sides unless a caption says otherwise. Correlations are Spearman rank correlations on 12- or 16-pixel blocks.

Star positions, magnitudes and spectra: ESA Gaia DR3. Hα reference: IPHAS DR2, the INT Photometric Hα Survey of the Northern Galactic Plane, taken with the 2.5 m Isaac Newton Telescope on La Palma; I built the continuum-subtracted image and projected it into the Seestar frame. Second reference: Pan-STARRS1 r.