Why Deconvolution Deserves a Spot in Your Workflow
Every image you capture has been softened. The atmosphere blurs it, your optics blur it, and even a perfect focus and a perfectly tracked mount cannot escape the fact that stars in your frames are not points. They are tiny disks, and the size of those disks sets the ceiling on how much detail your final image can hold.
Deconvolution is the processing step that pushes back against that softening. Applied carefully in Siril, it tightens star sizes, recovers fine structure in galaxies and nebulae, and gives your image a crispness that no amount of sharpening sliders can fake. Applied carelessly, it creates ringing artifacts, crunchy stars, and haloed edges that scream overprocessed.
This tutorial walks through how Siril’s deconvolution tool works, the settings that matter, and the warning signs that tell you to back off.
What Deconvolution Actually Does
Blur in an astronomical image can be described mathematically by something called the point spread function, or PSF. The PSF describes how a single point of light, a star, gets spread into a disk by the atmosphere and your telescope. A tight PSF means sharp stars. A fat PSF means mush.
Deconvolution runs that process in reverse. You give Siril an estimate of your PSF, and the algorithm, Siril uses a Richardson-Lucy implementation, iteratively works out what the image would have looked like before that blur was applied. Globular clusters resolve into individual stars. Galaxy dust lanes gain definition. Faint halo detail around planetary nebulae separates from the background.
The key word is iteratively. Each pass sharpens a little more, and each pass also amplifies noise a little more. The entire skill of deconvolution is finding the stopping point where you have gained detail without destroying the image.
The Golden Rule: Linear Data Only
This is the mistake that ruins most first attempts at deconvolution. The algorithm assumes the pixel values in your image have a linear relationship to the amount of light that hit your sensor. That is true of your stacked master image straight out of registration and integration, and it stops being true the moment you apply any non-linear stretch.
If you run deconvolution after stretching your histogram, the math no longer describes reality. The result is usually harsh, ringing stars and a plasticky look. So the correct place in your workflow is:
- Stack your light frames as described in the Siril stacking workflow.
- Remove gradients and calibrate color while still linear.
- Run deconvolution on the calibrated, still linear stack.
- Only then apply your histogram stretch and finish the image.
If your stars turn into donuts with dark centers or bright rings around their edges, the first thing to check is whether the data was still linear when deconvolution ran.
Finding Your PSF Sigma
Siril needs your PSF described as a Gaussian with a sigma value in pixels. You have two ways to get it.
The easy way is the autodetect button in the deconvolution tool. Siril analyzes stars across the image, fits PSFs to them, and uses the median values. On a well-stacked image with round stars, this works well and is the right starting point.
The manual way gives you more control. Run the PSF analysis on a handful of isolated, unsaturated stars spread across the frame, and look at the reported sigma values. Averaging them yourself lets you sanity check the autodetect result, and on frames with elongated stars from guiding error, manual inspection tells you the deconvolution may fight a losing battle. Deconvolution sharpens what is there. It cannot round stars that were never round to begin with, which is one more reason good focus and solid guiding come first.
One refinement worth knowing: Siril supports a beta parameter that lets the PSF model have a Moffat-like shape rather than a pure Gaussian. A pure Gaussian falls off too quickly to perfectly describe real star profiles, so a beta value around 2 is a common choice for deep sky images. If your stars look right but halos persist, experimenting with beta is the next knob to turn.
Iterations: Where Restraint Wins
The number of iterations is the setting people push too far. A reasonable range for typical deep sky data is somewhere between five and twenty iterations. The exact number depends on your signal to noise ratio, which is one more argument for generous integration time, because stacking more exposures gives deconvolution more signal to work with.
Work in small steps. Run five iterations, look at the result, run five more, look again. Zoom in on stars, not on the galaxy core. Stars tell the truth first.
Here is what to watch for:
- Ringing: bright rings or dark halos appearing around stars. Stop and reduce.
- Donut stars: dark centers in formerly solid stars. You have gone well past the limit.
- Crunchy background: noise in empty sky becoming speckled and mottled. Reduce iterations or improve your data first.
- Plastic galaxy cores: the core losing all gradation and turning into a flat white blob with hard edges.
A Real Example: M106
The galaxy image below is a good candidate for this kind of processing. Messier 106 is an intermediate spiral about 22 to 25 million light years away in Canes Venatici, with an active Seyfert nucleus and structure in its arms that rewards careful sharpening.
This frame came from 76 exposures of 180 seconds each across two nights, captured with NINA and stacked in Siril. The imaging train was an Orion 8 inch 1000mm f4.9 Newtonian on an Orion Atlas EQ-G mount, with a full spectrum modified Canon T3i at prime focus and a Meade 800mm guide scope reduced to f5 riding alongside with a DSI Pro guide camera. That is roughly 3.8 hours of integration, enough signal for deconvolution to work with cleanly. On the stretched final image, deconvolution of the linear stack is what separates the faint outer arms from the sky background and keeps the star fields crisp across the frame.
You can see the full resolution version in the gallery.
Quick Settings Checklist
If you want a starting point to experiment with on your own linear stack, try this:
- Sigma: autodetect, then verify against a few manual PSF measurements
- Beta: 2.0 as a starting value for deep sky targets
- Iterations: start at 5, increase in steps of 2 to 5, inspect stars at every step
- Data: strictly linear, after background extraction and color calibration but before any stretch
Every telescope, camera, and sky produces a different PSF, so treat these as a starting point rather than a recipe. The habit that matters is checking star shapes at high magnification after every run.
The Bottom Line
Deconvolution is one of the highest impact steps you can add to a Siril workflow, and also one of the easiest to overdo. Keep your data linear, let Siril measure your PSF, ramp up iterations slowly, and let the stars be the judge. Done right, your galaxies gain arm structure, your clusters resolve to the core, and the image keeps the natural look that separates disciplined processing from crunch.
Clear skies, and may your sigma be small.
