Syqon StarLess vs PI StarXterminator first impressions comparison

My results from SyQon on Siril were very good as well. The artefacts from Star removal are much better as compared to StarNet++ Does some amount of denoising as well which makes the image and nebulosity much clear.

Yet to try on multiple images but sofar results look promising

Good effort from Zenith for Siril users

Georg N. Nyman · Feb 11, 2026, 03:38 PM

andrea tasselli · Feb 11, 2026, 01:48 PM

Willem Jan Drijfhout:
NXT works exceptionally well

No, not really...

NXT is OK, but imo not really excellent - from what I found out is that you have to be very careful regarding the amount of reduction because it creates artefacts and artificial structures quite easily. From what I tried is that noise reduction results are better when I applied a subtle reduction twice than one only with a larger amount set. Maybe this makes sense…. I am sure, Andrea has an opinion on this as well!

Are you using AI v2 or v3?

Hi everyone, and thanks for the feedback and positive comments. I’d like to clarify that the models released for Siril and SAS are not full versions but lightweight derivatives of our commercial model Axiom. They were specifically adapted for the open-source ecosystem (a deliberate commitment on our part) and optimized for the computational constraints and processing pipelines typical of those environments. Axiom, which Georg is currently testing, uses a larger and more refined architecture and produces noticeably cleaner and more stable results; I hope he’ll be able to publish a review of it soon.

A technical note regarding noise reduction: the model does not introduce synthetic artefacts or “reconstruct” nonexistent details. What it performs is a statistical separation between signal and noise based on learned inference. Structures that might appear anomalous are in fact low-SNR signal components that were previously buried in the background noise. In other words, the model does not invent information; it improves the estimability of the latent signal.

From an algorithmic standpoint, this is not a conventional denoising filter (linear or nonlinear), but a probabilistic regression process trained on real astronomical data distributions. This allows noise suppression that remains consistent with signal statistics, avoiding the artefacts typically associated with aggressive deterministic methods.

Both the open models and Axiom are designed to operate as the first step of the processing pipeline, immediately after stacking and before any further editing. At that stage the data is still linear and statistically intact, so improving the signal-to-noise ratio is more effective and does not get amplified or distorted by subsequent transformations.

Below is also a comment from one of the main Siril developers:

Actually at first I thought the elongated faint fuzz around the lower satellite galaxy and the darker region around it was possibly an artefact, but I compared it with https://arstechnica.com/space/2023/12/daily-telescope-one-of-the-most-stunning-andromeda-photos-ive-ever-seen/ and realised that the star removal (and the noise reducing effect!) is now that good that I'm just seeing more faint detail in the image than previously. The difference in bgnoise stats is remarkable:

```

00:18:20: Running command: stat

00:18:21: Statistics: processing...

00:18:21: Red layer: Mean: 701.7, Median: 647.3, Sigma: 682.9, Min: 575.6, Max: 65528.9, bgnoise: 9.7

Green layer: Mean: 708.1, Median: 647.3, Sigma: 939.8, Min: 580.4, Max: 118628.8, bgnoise: 9.5

Blue layer: Mean: 709.7, Median: 650.4, Sigma: 958.3, Min: 553.0, Max: 95911.3, bgnoise: 15.5

00:18:21: Statistics succeeded.

00:18:21: Execution time: 302.67 ms

00:18:24: Redo: SyQon Starless - star removal

00:18:28: Running command: stat

00:18:28: Statistics: processing...

00:18:28: Red layer: Mean: 682.2, Median: 640.8, Sigma: 248.2, Min: 615.5, Max: 65535.3, bgnoise: 1.3

Green layer: Mean: 684.2, Median: 640.6, Sigma: 330.5, Min: 612.4, Max: 65535.4, bgnoise: 1.5

Blue layer: Mean: 683.7, Median: 640.9, Sigma: 245.5, Min: 604.9, Max: 59284.8, bgnoise: 1.7

```

_________

Axiom was not trained simply on paired “input” and “target” files to learn what stars look like and how to remove them. In addition to supervised input–target training, it was also trained on statistical and scientifically characterized datasets, allowing the model to learn the underlying distributions of astronomical signal and noise rather than just memorizing transformations.

Yes, it´s me - Georg, as mentioned by Mike (Syqon). I shall write a review on the new version - after some additional explanations by Mike, I get on very well with this software and the results, I can see sofar are very good, really good! It might take another few days to compile my findings before I can post them.

SyQon Astro:

Fabrice Lamidey · Feb 16, 2026 at 03:26 AM

I found v3 introduces artifacts that v2 does not, on top of being much slower.

There is no v3, where did you download it?

andrea tasselli · Feb 11, 2026, 01:48 PM

No, not really...

Depends on your standards. As someone who has been doing imaging for over 30 years now, overall it works vastly better and vastly easier than the older non-AI methods (and I have used pretty much all of those at one time or another).

So for newbies (defined as someone that has been doing this for less than 5 years or so) looking for perfection, it is less than ideal. Looked at with historical perspective, it is very nice indeed.