Deep-sky images vol. 3: a refined technique for higher quality results

It’s been a bit over a year since my last astrophotography post highlighting galaxies and a bit over two years doing serious astrophotography. Even though my rig has been nearly static, with the exception of adding an Antlia ALP-T 5nm Ha+Oiii filter to retire the Svbony and its impressive Oiii halos and replacing my flexible spider vane with a rigid one including better adjustment screws for the secondary mount, I have learned a lot about how to handle the raw data to produce more impressive results. Now it’s your turn, or you can skip to the end and just enjoy the images.

A refined deep-sky workflow

Data acquisition

At the core of my routine is still acquiring dithered and calibrated data of the sky with my telescope and an OSC camera, optionally using an Ha+Oiii dualband filter. I now use a ZWO ASI2600MC-Pro camera exclusively, which has some implications for calibration: there’s (almost) no dark current (sometimes called “amp glow”) so dark calibration is entirely unnecessary. That leaves just bias and flats, which I take each session with the same gain and offset as the target with the camera always cooled to -10C so that discrepancy is never a concern.

Recently I’ve defaulted to a gain of 100 (minimum for high conversion efficiency to achieve lowest read noise + highest dynamic range) and offset of 50 for broadband targets, adjusting exposure time as necessary to fill the available dynamic range. When using a dualband filter, I’ll increase the gain to 300 to keep exposure times in the several minute range for full use of the dynamic range.

I use the INDI and KStars/Ekos suite for all my telescope control and data acquisition needs. It can do everything from goto to polar alignment to guiding to scheduling acquisitions and is Linux native.

Remember: good data is taken in a dark location. Both the sky and the surroundings (street lights, etc) matter, and any skimping here will show up in the final stacked image and must be handled in post-processing.

Pre-processing

For pre-processing (calibrating and stacking) the raw data I strongly prefer Siril’s workflow to any competition. The reason for this is multifaceted: it’s FOSS, I understand how to use it, and it implements a drizzle integration algorithm for raw (undebayered) OSC images. I also find in my earlier data, which is plagued by gradients from street lights, that the Radial Basis Function (RBF) background extraction technique in Siril does the best of the non-AI algorithms.

My Siril workflow goes roughly like this, which would work for any zero dark current OSC camera:

1. Create a sequence of bias frames
2. Average stack with rejection to create the bias master
   • No normalization, 3 sigma winsorized clipping
3. Create a sequence of flat frames
4. Calibrate flats with dark subtraction
   • Using the bias master from #2 as the dark master and no bias or flat master set
5. Average stack with rejection to create the flat master
   • Multiplicative normalization, 3 sigma winsorized clipping
6. Create a sequence of light frames
7. Calibrate with flat division and dark subtraction to produce calibrated lights
   • Using the bias master from #2 as the dark master, flat master from #5, and no bias master set
8. Background subtract 1D poly on sequence with default sample points
   • Performing this before stacking can better handle per-frame gradients
9. Perform 2-pass global registration (first pass only)
10. Perform sub-frame selection on wFWHM, roundness, background, etc
11. Apply 2-pass global registration (second pass only) 1x drizzle
    • Try a larger drizzle scale if feeling spicy – 1.5x is not insane for OSC
12. Average stack with rejection to create the integrated stack
    • Additive normalization with scaling, 3 sigma winsorized clipping
    • Tweak rejection sigma (2.0 - 5.0) depending on data amount and artifacts (satellites etc)
13. Background subtract 1D/2D poly or RBF to remove remaining sky-glow
    • Careful with the sample points this time to avoid stars / features
14. Save final image and proceed to post-processing

Post-processing

Taking good data isn’t even half the battle, as it turns out. People have a lot of aesthetic expectations about what a “good astrophotograph” looks like, and those are not usually driven by scientific accuracy or true-to-life appearance, but rather the false color highly processed images released by space telescopes like Hubble or the James Webb or other hobbyists chasing the same look. The solution, of course, is to highly process the data in similar ways, and choosing your own line in the sand for what is “too much” in terms of allowable image manipulations.

There are some major manipulations to consider: color calibration, noise reduction, deconvolution, masking, gradient reduction, nonlinear stretching, and other standard photo adjustments. Of these, only stretching (similar to an exposure adjustment in photo processing) is truly required to get a viewable image; linear data is simply not meant for direct viewing. Color calibration, photo adjustments, are not technically required but very important in achieving a pleasing final result, and generally accepted. That just leaves noise reduction, deconvolution, gradient reduction, and masking as points of tension.

Noise reduction

Noise reduction is as critical here as any other low-light situation, though there are different techniques for the extreme low-light night sky than for a dim room. Historically, many heuristics have been tried from median filters to edge-dependent blurring algorithms, and they all are well understood and have drawbacks. The cutting edge, and top performers, are neural networks trained on reconstructing the latent image most likely to have produced noisy data. This family has several contenders: GraXpert, NoiseXterminator, and DeepSNR of varying degrees of openness and cost. I found GraXpert and NoiseXterminator to perform similarly, perhaps with a slight edge to NoiseXterminator, which I did purchase to evaluate. Overall, DeepSNR consistently produced the best results out of the three. Props to GraXpert for being FOSS and DeepSNR for being free. Images below were produced with one of the three – I’ll try to note which.

Overall, the lack of high-performing algorithms without neural-networks generating pixel-level corrections to the image leads me to use DeepSNR as the noise reducer of choice.

Gradient reduction

I was going to leave this step out, since my pre-processing includes gradient removal with Siril using mathematically well-defined background models, but shout-out to GraXpert for having a quite good neural-network based gradient removal tool in addition to its noise reduction. Some images below leveraged this, and I am following the development to include this in the Siril UI as a background removal option. I should note that GraXpert (FOSS) should not be confused with GradientXterminator (proprietary) which I did not evaluate due to its cost.

Deconvolution

Deconvolution is the astrophotography word for “sharpening” and realistically that’s how its meant these days as opposed to a true deconvolution. The true deconvolution techniques utilized a point spread function based on stars in the image to estimate the imperfections in the optics to in turn estimate the latent image most likely to have produced this one without those imperfections, thereby sharpening the image. Now practitioners apply sharpening techniques at the point in the process at which one would previously have deconvolved, and there are again many options out there. GraXpert again has some neural-network latent image reconstruction functionality in this space, making its inclusion in Siril and future refinement a very promising development for FOSS, but I was not impressed with the degree to which it made images look artificial. Indeed, this is a problem that exists even with the deterministic non-AI techniques: it is very easy to over-sharpen an image.

RC Astro’s BlurXterminator shines here for offering a “correct only” option that tones down the sharpening operation into something that could more believably be described as an ideal deconvolution with a position dependent point spread function. I find the resulting image to look like data taken with excellent collimation, zero tilt, and a perfectly coma-free flat field instead of something run through a sharpening algorithm, but make no mistake this too is a neural-network based reconstruction technique.

BlurX “correct only” is my line in the sand as far as substantial manipulation goes, and I now use it on most images to clean them up.

Masking

Slightly less egregious than neural-network sharpening is the idea of using masks to process different regions of an image differently, i.e. (and most common in practice) separating stars from the rest of the image. The motivation for this is good: the stars are very bright and the rest of the image is typically very dim, so it better leverages the output dynamic range to stretch the stars separately. It is possible to use range masks or generate masks with Gaia data to literally mask out stars, but again a neural-network latent image reconstructor from RC Astro, StarXterminator, is the only game in town.

There is a free alternative, StarNet (StarNet2? StarNet++? inconsistent branding…), but I found the results from StarNet to be particularly poor on Newtonian diffraction spikes, which StarXterminator identified and removed reliably. If you’re using a refractor instead of a Newtonian, give StarNet a try, but otherwise my recommendation would be StarXterminator despite the cost.

PixInsight

Historically I have used Siril for these steps, but in chasing the dragon of noise reduction techniques, I eventually landed in the market-cornering insanely-priced monopoly that is PixInsight for astro image processing. It’s a de-facto monopoly because its often the only/best way to access tools meant to be used on astro data: they’re distributed by third parties as Pix plugins. If you’re going to bite the bullet and pay for it, you might as well use it for every post-processing step, since it is a nicely integrated environment for standard techniques even if its most powerful features are 3rd-party (sometimes paid) plugins. Fingers crossed Siril gets there, but until then, PixInsight it is.

A post-processing workflow is highly varied and includes many more options than pre-processing, but most recent projects have followed the following procedure:

0. ScreenTransferFunction (STF) the linear image into something that looks decent
1. Use the ImageSolver script to get an astronomical solution for future steps
2. Apply Spectro-Photometric Color Calibration (SPCC) to the image.
   • This uses reference information for the expected color of the stars to calibrate the image
   • Many other options exist, but this one is data-driven (choose camera settings) and deterministic
3. Remove noise with DeepSNR or other technique
   • Generally strength of 0.7-0.95 produce decent results, depending on starting noise levels
4. Deconvolve / sharpen / correct image with BlurXterminator “correct only”
   • If feeling spicy, try the heavier sharpening options
5. Split stars into separate image with StarXterminator
   • Remaining steps apply to both variants, but won’t use the same settings
6. Adjust the STF to get decent star and starless starting points
7. HistogramTransform the linear images into nonlinear form
8. CurvesTransform to highlight detail, hide flaws, boost saturation, etc
   • GeneralizedHyperbolicTransform could be used here, especially on the star image
9. Use BackgroundNeutralization primarily to get background levels at 0.05-0.08
   • AutomaticBackgroundExtraction might be a good idea as well, if stretching revealed more gradient
   • Note: this is not needed on the star-having image, but greatly impacts the starless
10. Recombine the star and starless images into the final result with PixelMath
    • ~((~starless)*(~stars)) for the math: invert both, multiply together, invert result

Now, on to the results.

Deep-sky images taken or processed in 2025

These are grouped by type of object for some insight into the nature of what’s captured. Some were processed with GraXpert noise reduction only, others were processed with NoiseX and BlurX (correct), and yet others with DeepSNR and BlurX (correct), as noted in the captions. All were pre-processed in Siril and post-processed in PixInsight, following what was described above. I have added exposure details where readily available, and included any interesting processing quirks as well. All were captured from my backyard in Bortle 7/8 conditions.

Many of these images were featured in the last two astro blog posts, but have now been reprocessed and worth a comparison. Others are entirely new, being processed for the first time with the more advanced tools.

Supernova Remnants

Nebula generally are going to be glowing gas, usually red (Ha) and blue (Oiii) and various other light atoms. In order to glow, something has to keep them hot, and usually that something has some bearing on the shape and age of the nebula as well. In this case, the glow is sustained by the shockwave of a star that previously went supernova and blew itself up, making them relative recent and ephemeral compared to other deep-sky objects. Being single star scale objects, the smallest will be O(1 ly - 10 ly), while being very energetic, the largest can be O(100 ly) before fading.

M1 Crab Nebula

IC443 Jellyfish Nebula

NGC6992 Eastern Veil Nebula

Cygnus Loop (widefield)

Planetary Nebula

Unlike supernova remnants where a star exploded, a planetary nebula is a slower more gradual death as the outer layers of a star are blown away by its own power. These gas expand into complex oddly symmetrical shapes with variously colored layers, and are heated by the solar wind and light of the dying star. Like supernova remnants, these are decently ephemeral, but may last much longer than a remnant. The small selection here is in the O(10k year) range of age. These are on the smaller side compared to remnants, as they tend to be far less energetic than supernova, but still can reach the O(10 ly) scale.

NGC7293 Helix Nebula

M27 Dumbbell Nebula

M76 Little Dumbbell Nebula

Reflection / Emission Nebula

Unlike planetary nebula and supernova remnants, nebula with a reflection component are always associated with living and typically young (for a star) stars. The starlight is reflected off local dust, hence the name. These nebula often have some amount of emission light as well, since dust usually contains some hydrogen, and hydrogen tends to glow red. Usually the star is actively blowing this dust away, so regions of low density will have been swept clean, while higher density regions are being swept. Proximity to a star is still important here, so objects (the illuminated parts, anyway) will tend to be on the O(1 ly - 10 ly) size, but since the existence is now contingent on star life, the ages and lifespans of objects easily exceeds O(10m years). I’ll lump dark nebula in here as well, as occlusion and reflection seem related to me.

M45 Pleiades

M78

B33 Horsehead, NGC2024 Flame, IC434, NGC2023

IC63 Cassiopeia’s Ghost

M42 Orion Nebula

Emission Nebula

Emission nebula may still be heated by stars, but they are characterized by primarily emission instead of reflected star light. Since the object is visible primarily through emission (much like older supernova remnants), dualband filtering helps to reduce the impact of a bright sky to see much fainter objects than would usually be possible. These nebula come in all sizes, but tend to be both larger and older than reflection nebula.

M16 Eagle Nebula ft. Pillars of Creation

NGC2359 Thor’s Helmet

C49 Rosette Nebula

C11 Bubble Nebula

NGC2264 Cone Nebula Region

IC405 Wizard Nebula

NGC6888 Crescent Nebula

Sadr Region ft. Crescent Nebula (widefield)

NGC7000 North American Nebula (widefield)

Galaxies

Nebula are great, but all of them reasonably accessible to amateur gear are within our own Milky Way galaxy and, even within that subset of the universe, not too far from us relatively, at no more than 10k light years distant in a galaxy hundreds of thousands of light years across. This is largely due to the galaxy being full of stars and dust, blocking the view of more distant things, i.e. we can’t see behind the Orion Molecular Cloud at all. If one looks through the breaks in the dust outside of our galaxy, many more galaxies are visible and easy to photograph.

Other galaxies range from “nearby” at 2.5 million light years to “intermediate” in the 20-50 million light year range, with even more distant galaxies in the >100 million light year range achievable with amateur equipment.

M31 Andromeda Galaxy

M33 Triangulum Galaxy

M100 Mirror Galaxy

M101 Pinwheel Galaxy

M77 Squid & NGC1055

NGC247 Needle’s Eye Galaxy

M74 Phantom Galaxy

NGC7331 & friends - Deer Lick Group

HCG92 - Stephan’s Quintet

M106 & friends

M51 Whirlpool Galaxy

NGC6946 Fireworks Galaxy

M66, M65 and NGC3628 - Leo Triplet