New BlurXterminator (ML5 Model) accuracy - comparison with sky survey data - Forums

(This post is meant to be purely informational!)

RC Astro has just released a new version of BlurXterminator (the ML5 model), the showcase images may look exciting at first, but I think it’s important to take a closer look at what the tool is actually doing to your data:

Below are a few comparison between the examples posted in the RC Astro post about the new model being released (more specifically, the “Faint background galaxies near NGC 1097” images from the “Faint fuzzies” section) and a sky survey available at https://www.legacysurvey.org/viewer
The sample I’m going to show you can be found around these coordinates: Ra: 41.5485 Dec: -30.1851
You can locate it pretty easily by using the Jump to object field in the bottom left on the website, typing in “NGC 1097”, then locating this “Y shape” group of galaxies that can be seen roughly in the middle of the original image, going up at around the 1 o’clock position from NGC 1097.
📷 The location of the image relative to NGC 1097

The four images - the original, the ML4 (old blurX) sharpened image, the ML5 (new blurX) sharpened image, and a screenshot of the sky survey were all loaded into PixInsight. Then, a DynamicAlignment process was used with the Original image as the source and the Survey image as the target, thus matching the Survey screenshot to the same framing and pixel scale as the other images. They were arranged in a grid in PixInsight’s interface (Original in the top left, Survey in the top right, ML4 in the bottom left and ML5 in the bottom right), the views were “binded together” to display the same zoom/position on all four views.

📷 full framing, 1:1 scale📷 zoom on a galaxy group on the right📷 zoom on some galaxies on the left📷 zoom on the central bunch of galaxies📷 zoom on some galaxies in the top left📷 zoom on some galaxies in the bottom right

The aim of this post is purely informative, I will let you come to your own conclusions :)

In my opinion however, the results of the new ML5 model stray quite far from the real structures - the shapes of many of those galaxies are almost completely different from reality, the model appears to “make up” a lot that isn’t actually in the captured image and guess wrong.
The previous ML4 version, by my eye, appears to produce results somewhat closer to reality (particularly noticeable in the last two images), however that one too seems to not get stuff quite right sometimes - in my opinion it is far less egregious though and much more in the realm of being ‘fine’. It is important to keep in mind though that maybe it’s worth it to keep blurX settings a bit lighter the next time you process something :)
The only image to really stay properly true to the survey data is of course the original image - it is however quite lacking in sharpness and signal of course.

Personally I have a pretty open approach to art - if you like blasting your images with sharpening, “cooking” or “frying” them, or whatever else - whether you make ‘conventionally good’ images or ‘weird’ images, go ahead!
But, since this hobby is Astrophotography and our aim is usually capturing real images of the sky - the ‘realness’ of the captured data is quite important to a lot of imagers. I think a lot of people don’t realize how certain tools (and I don’t mean just blurX/RCAstro tools!) can put things quite far from both the actual image that they captured and what’s in the sky itself.
I feel like a lot of people tend to think that tools such as BlurXterminator will always keep true to their data and are simply straight-up improvements or fixes without compromises - and sometimes these tools do help a lot! But it’s important to keep in mind they can produce results that simply don’t match reality, particularly when pushed too hard (which I feel like a lot of people unintentionally do, unfortunately :/) and to look out for things that may be going wrong :)
Very often a more tasteful, ‘held back’ coat of BlurX can look better and be more true to reality than pushing for maximum detail :)

I hope you found this post helpful and informative!
Clear skies! ❤️

This has been debated many times since these tools have existed. I’ve seen very good (and some less good 🙂 ) points on each side of the argument, and I agree with some of them.

Whilst staying true to data is important, in my opinion, there is a cut-off line where that stops being true the moment you use an AI-based sharpening tool. By their nature, they will interpret the data shown and modify it. At large scale, where SNR is high and there is little room for interpretation, the tool will enhance what is there, but for background fuzzies, I accept that it will basically guess at the object.

It usually does a good job, and for me, personally, the fact that it’s showing “something” there in a (somewhat) vaguely correct shape is enough.

That being said, I do agree with you that the last two images (particularly the 2nd to last), look better with the previous version. But realistically, how often will you zoom in that far?

John Nedelcu · Aug 29, 2026, 12:51 AM
This has been debated many times since these tools have existed. I’ve seen very good (and some less good 🙂 ) points on each side of the argument, and I agree with some of them.

Whilst staying true to data is important, in my opinion, there is a cut-off line where that stops being true the moment you use an AI-based sharpening tool. By their nature, they will interpret the data shown and modify it. At large scale, where SNR is high and there is little room for interpretation, the tool will enhance what is there, but for background fuzzies, I accept that it will basically guess at the object.
It usually does a good job, and for me, personally, the fact that it’s showing “something” there in a (somewhat) vaguely correct shape is enough.

That being said, I do agree with you that the last two images (particularly the 2nd to last), look better with the previous version. But realistically, how often will you zoom in that far?

“Whilst staying true to data is important, in my opinion, there is a cut-off line where that stops being true the moment you use an AI-based sharpening tool” - while I agree to an extent, I think the line is rather fuzzy and not sharp (not pun in terms of the topic at hand intended 😅) where it’s one thing to have a light pass of an ML-based tool and another to crank it pretty hard, a light pass usually stays pretty close to reality while helping aesthetically :) stronger usage usually both makes the image far from the truth and doesn’t make it very pretty :/

and if you’re not gonna zoom in close enough to truly inspect that tiny detail, you might as well back the sharpening off a bit anyway :)

I guess it is also a fair point though that non-tiny structures are usually affected by this kind of stuff less (it is still somewhat common with denoising tools though, like DeepSNR for example), but this ML5 version also does some pretty strong altering to larger things, especially flares and flare-like stuff around stars seems to basically get replaced completely… I imagine there’s gonna be conflicting opinions on that one, personally I feel like it’s changing the underlying image too much though honestly :/

I checked out the demos on the RC astro site and for the correction of artifacts around bright stars there were several “new stars” that appear around the corrected bright stars using ML5 😬

When you are getting down into the weeds in terms of resolution and S/N things are going to get marginal. I think if you backed off a bit with these examples the differences wouldn’t be so noticeable. Giving the software so little to work with pretty much guarantees inacuracy.

Honesty, at the image scale presented I prefer the images before BX of any kind is applied.

I’ve always really enjoyed the challenges this hobby brings. As these AI based tools have trickled out I’ve noticed that it just doesn’t feel as fun anymore. This looks like another step in that direction. As the line keeps shifting further into the world of fantasy I feel more and more strongly about just not using these AI tools altogether.

Maybe this is like when I started getting into photography right around the film to digital transition. Plenty of people had no interest in switching to digital and bemoaned the changes to the art form. I thought they were just old and stuck in their ways. I embraced digital photography without skipping a beat.

And now, 25 years later, I find I only really enjoy practicing photography when I shoot on film.

John Nedelcu · Aug 29, 2026 at 12:51 AM
there is a cut-off line where that stops being true the moment you use an AI-based sharpening tool. By their nature, they will interpret the data shown and modify it. At large scale, where SNR is high and there is little room for interpretation, the tool will enhance what is there, but for background fuzzies, I accept that it will basically guess at the object.

BXT is not an AI based tool. It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution. The accuracy of the result is very dependent on the quality of the training data. That’s the data that is being used to minimize a least square fit between the real data and mathematically blurred version of the training data. The density of the training data and noise in the image are generally the major factors that limit the quality of the estimated output. To be clear, there is no “generative” component to the way that BXT works. Deconvolution is an ill-posed problem and no algorithm is perfectly accurate—and BXT is no exception. Setting the sharpening factor to a very high level increases the possibility of artifacts and errors in the result. However, used properly BXT is one of the most effective deconvolution methods devised for astronomical images.

- John

Dart Frog · Aug 29, 2026 at 12:30 AM
The aim of this post is purely informative, I will let you come to your own conclusions :)

Thank you for this comparison, it is indeed very informative. Based on your data, I come to the opposite conclusion. In almost all cases where there is a difference between ML4 and ML5 and ML5 shows more/crisper detail, this detail is backed up by the survey data. Proving the point that BXT is not ‘making up’ data, just mathematically improving data.

The discussion about how much you want to let maths do the work, what is aesthetically more pleasing etc., is a very different one and applies equally to ML5, ML4 or any other deconvolution model. This not only depends on the model, but just as much on how you use it, settings chosen, etc.

I tried out ML5 on some of my masters and I would agree that the deconvolution was (slightly) improved in low SNR areas like the background containing galaxies.

On the other hand even with the lowest setting of 0.0 the correction of star aberrations and diffraction spikes has a massive impact on the image. The stars are always “perfect”, which is odd and takes away the “personality” of your image that was taken with your equipment. Without differences the images may feel interchangeable and I do not know what to think of it. But that is only my opinion.

John Hayes · Aug 29, 2026, 05:31 AM
BXT is not an AI based tool. It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution. The accuracy of the result is very dependent on the quality of the training data.

yes it is. and if it isnt things like chatGPT arent AI. it all runs on neural nets. they are called neural nets because they are based on neurons. the things that make creatures intelligent. hence artificial intelligence

John Hayes · Aug 29, 2026 at 05:31 AM
It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution.

John, you are much more knowledgeable about deconvolution techniques than I’ll ever be, but is this not precisely what AI-based tools do? Non-analytic solutions, to me, read like they are not using rigid math formulas, but instead, learning from data to make predictions or estimates. This refers to probabilistic reasoning, which is how AI handles complex, real-world problems.

When an LLM generates text, it is not copying and pasting from a database or following a rigid, step-by-step mathematical formula. Instead, it relies entirely on neural net estimation to process your prompt, calculate the mathematical probabilities of what words should follow, and output a non-analytic, best guess solution one token (a word or fragment of a word) at a time.

- Pedro

Tobiasz · Aug 29, 2026 at 10:11 AM
I tried out ML5 on some of my masters and I would agree that the deconvolution was (slightly) improved in low SNR areas like the background containing galaxies.
On the other hand even with the lowest setting of 0.0 the correction of star abberations and diffraction spikes has a massive impact on the image. The stars are always “perfect”, which is odd and takes away the “personality” of your image that was taken with your equipment. Without differences the images may feel interchangeable and I do not know what to think of it. But that is only my opinion.

I agree 100%. When I compared 4 and 5 on a bright star from my newt, it looks perfect with no turned down edge flare whatsoever. ( image below at 4:1 magnification) But it looks a bit too clinical imo, equating in my mind to having noise reduction at 100%, looking fake. In that case, I may be inclined to use both 4 and 5 and combine them to bring down the flare but keep some personality and drama. He’s probably shaking his head right now thinking he can never make everyone happy, but hopefully he’ll provide a setting to control the level of flare correction. 📷 image.png

Heck, let’s ask the source……
CGPT sez:

“Is [BlurExterminator] actually performing deconvolution?

This depends on how strictly you define the term.

From a signal-processing perspective:

No. It is not explicitly solving the inverse convolution problem with a measured PSF.

From a practical imaging perspective:

Yes. It is learning an approximation to the inverse mapping that deconvolution seeks to compute.

Many researchers refer to this as learned deconvolution or implicit deconvolution.

Why can it outperform classical deconvolution?

This is perhaps the most interesting part.

A classical Richardson-Lucy algorithm knows only the PSF.

BlurXTerminator has effectively learned priors about astronomical images, such as:

stars are nearly point sources
galaxies have characteristic spiral structures
dust lanes are continuous
diffraction spikes have recognizable shapes
nebular filaments tend to continue smoothly
noise has statistical properties different from real structures

These learned priors let it make educated guesses in situations where classical deconvolution cannot distinguish between noise and signal.

The tradeoff is important:

Classical deconvolution is constrained by a physical imaging model and is less likely to invent plausible-but-incorrect structures if its assumptions hold.
A learned model can often produce cleaner, sharper images—especially when the PSF is imperfectly known—but it may also "hallucinate" fine detail that looks astrophysically plausible without being strictly supported by the recorded photons. Good AI restoration models are trained to minimize this risk, but it cannot be eliminated entirely.

For astrophotographers, this means BlurXTerminator is best thought of as a learned inverse imaging operator: it uses a neural network to approximate the inverse of atmospheric and optical blur directly, rather than searching over many outputs from a traditional deconvolution algorithm or selecting the best one from a set of classical solutions.”

That’s a summation to a longer response. Happy to provide the full response and prompt if anyone wants.

John, CGPT’s response doesn’t seem to stray far from yours, but am curious if/where you think it does.

Cheers,
Scott

John Hayes · Aug 29, 2026 at 05:31 AM
John Nedelcu · Aug 29, 2026 at 12:51 AM
there is a cut-off line where that stops being true the moment you use an AI-based sharpening tool. By their nature, they will interpret the data shown and modify it. At large scale, where SNR is high and there is little room for interpretation, the tool will enhance what is there, but for background fuzzies, I accept that it will basically guess at the object.
BXT is not an AI based tool. It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution. The accuracy of the result is very dependent on the quality of the training data. That’s the data that is being used to minimize a least square fit between the real data and mathematically blurred version of the training data. The density of the training data and noise in the image are generally the major factors that limit the quality of the estimated output. To be clear, there is no “generative” component to the way that BXT works. Deconvolution is an ill-posed problem and no algorithm is perfectly accurate—and BXT is no exception. Setting the sharpening factor to a very high level increases the possibility of artifacts and errors in the result. However, used properly BXT is one of the most effective deconvolution methods devised for astronomical images.
- John

"BXT is not an AI based tool. It uses the process of neural net estimation." Those are the same thing. RC Astro's own product page calls BXT an ML-powered deconvolution tool, and the model in question is literally named ML5.

Per RC's own website, BXT was trained by convolving an ideal image with a PSF and adding noise to make the input, with the target being that same ideal image convolved with a smaller PSF. I'd argue that's not meaningfully different from teaching a human artist to paint a sharper version of a scene by showing them thousands of blurred/sharp pairs until they get good at it. In both cases you get a plausible reconstruction, and in neither case can you point at a given pixel and say whether it came from the input or from priors picked up along the way. Learning a good prior doesn't change that.

You're right that deconvolution is ill-posed and that artifacts scale with the sharpening setting. But those points don't rebut a side-by-side against Legacy Survey data. If the structure isn't in the sky, what we call the process that produced it is beside the point.

I think this comparison points at the broader problem with BXT and other AI-tools like it. Push analytical decon past what your data and your optics support, and it screams at you with ringing, halos, worms in the background, etc. You back off because you can see you've gone too far. When you overcook an image with BXT, you don't get garbage, you get things that look exactly like faint galaxies and small stars, because that is precisely what it was trained to produce. The loud feedback loop that used to tell you to stop is gone.

I'm not trying to start the astrophotography vs. art debate. Plenty of people process for aesthetics and that's a legitimate thing to do. But I think a lot of us are in this hobby because we like being able to say the detail in our image came from the actual sky, on our actual night, through our actual scope. A tool that quietly fills in what the data doesn't support, in a form we can't visually distinguish from real detail, makes that much harder to claim in good faith.

Rick Krejci · Aug 29, 2026, 12:31 PM
Tobiasz · Aug 29, 2026 at 10:11 AM
I tried out ML5 on some of my masters and I would agree that the deconvolution was (slightly) improved in low SNR areas like the background containing galaxies.
On the other hand even with the lowest setting of 0.0 the correction of star abberations and diffraction spikes has a massive impact on the image. The stars are always “perfect”, which is odd and takes away the “personality” of your image that was taken with your equipment. Without differences the images may feel interchangeable and I do not know what to think of it. But that is only my opinion.
I agree 100%. When I compared 4 and 5 on a bright star from my newt, it looks perfect with no turned down edge flair whatsoever. ( image below at 4:1 magnification) But it looks a bit too clinical imo, equating in my mind to having noise reduction at 100%, looking fake. In that case, I may be inclined to use both 4 and 5 and combine them to bring down the flair but keep some personality and drama. He’s probably shaking his head right now thinking he can never make everyone happy, but hopefully he’ll provide a setting to control the level of flair correction. 📷 image.png

I think it would make sense to give the customers the option of how much they want their stars to be corrected. NoiseX developed in a similar way. It started with just one slider and now it has many different settings and options.

I reprocessed my M101 with ML4 and ML5 and posted both versions on my profile. Both got a BlurX treatment with 0.3 stellar and 0.5 non-stellar and I must admit that ML5 did a great job in sharpening while keeping a natural smoothness. ML4 was more aggressive and created much more artifacts like mottling in the background noise and “worming” in the high SNR areas. Stars are perfect in the ML5 version, as expected.

I don’t mind the stars if I could disable the correction with setting 0.0, but it does not happen because I think the process is not part of the stellar sharpening. I guess it’s happening when BlurX tries to correct mechanical errors like defocus, guide errors, field curvature and other aberrations.

Other than that it may be a worthwile upgrade, but on the other hand I could apply ML4 with a less aggressive setting and save 50$.

Rick Krejci · Aug 29, 2026, 12:31 PM
Tobiasz · Aug 29, 2026 at 10:11 AM
I tried out ML5 on some of my masters and I would agree that the deconvolution was (slightly) improved in low SNR areas like the background containing galaxies.
On the other hand even with the lowest setting of 0.0 the correction of star abberations and diffraction spikes has a massive impact on the image. The stars are always “perfect”, which is odd and takes away the “personality” of your image that was taken with your equipment. Without differences the images may feel interchangeable and I do not know what to think of it. But that is only my opinion.
I agree 100%. When I compared 4 and 5 on a bright star from my newt, it looks perfect with no turned down edge flair whatsoever. ( image below at 4:1 magnification) But it looks a bit too clinical imo, equating in my mind to having noise reduction at 100%, looking fake. In that case, I may be inclined to use both 4 and 5 and combine them to bring down the flair but keep some personality and drama. He’s probably shaking his head right now thinking he can never make everyone happy, but hopefully he’ll provide a setting to control the level of flair correction. 📷 image.png

The next logical step is to have a future version that can remove the spikes entirely and reconstruct the image as if it were shot with a perfect refractor. The price for the update will be in the “perfect refractor class” range 🤣

I am afraid that everyone starts with good intentions to solve such ill-posed problems using novel techniques, but at the end of the day, what they want is to minimize their loss function. Who knows, maybe the “PixInsight way” of “documentary astrophotography” will eventually win.

Being new to astrophotography, these conversations are important in maturation of my image workflow. A couple of months ago I started to have this exact conversation contained in my head over blurX v4 as it looked to be overfitting the data aka creating image artifacts. Also, the better quality data with lots of subs and good SNR, the worse the artifacts. I stopped using blurX v4 and used v2 and just a light pass if at all.

Conversely, I had some previous captured noisy data and blurX v4 did ‘magic’. I was able to make a pleasing image on low quality data. A couple of passes to scrutinize for artifacts and I didn’t find any! Caveat, I’m maturing my abilities. So, I could have missed some things still basking in blurX ‘magic’.

The crux of this discussion is how wickedly tough it is to balance representation versus art with every new ‘tool’ in astrophotography. I’m not an advocate for hardening your workflow against new software or upgrades. Just the opposite. I look forward to using blurX v5 soon and have the good/bad signal to experiment.

Very appreciative all those have participated. I will be holding these trade offs in my head to guide me through editing.

I am keen to find my own path in this maze to call my own.

Andrew Lu · Aug 29, 2026, 01:48 PM
I think this comparison points at the broader problem with BXT and other AI-tools like it. Push analytical decon past what your data and your optics support, and it screams at you with ringing, halos, worms in the background, etc. You back off because you can see you've gone too far. When you overcook an image with BXT, you don't get garbage, you get things that look exactly like faint galaxies and small stars, because that is precisely what it was trained to produce. The loud feedback loop that used to tell you to stop is gone.

Not sure I agree with this. I haven’t yet seen real looking results from BX that aren’t supported by the data and/or other images (professional and amateur) of greater native resolution while I have seen results, when pushed too far, that are clearly neither real, nor real looking.

I think accepting a learned estimation of what your data would look like deconvolved vs a mathematical derivation with the same goal is simply a matter of personal choice. Either way, your data is being altered and the alterations are based on the data presented. And either way, it’s ultimately up to the user to determine if the alterations are valid.

Honestly, a lot of these discussions sound not so much about how deconvolution/sharpening is done, but how much is done…..it’s ok to alter your data towards some more optimal form, but too much is gauche…. If an algorithm was developed that could take Seestar data and turn out Hubble detail, would that be any more acceptable?

Cheers,
Scott

Here’s another side by side with different data. The subject is NGC6888 captured in Ha with a 150mm Newt, 585 mono sensor. The process was crop, bx correct only, bx default correction, statistical stretch.

The image order left to right is:

Original - BX v4 - BX v5

📷 comp2.jpgcomp2.jpg

📷 comp1.jpgcomp1.jpg

As everyone else has found, the major change is to the stars. With this data set the stars have a bit less halo and are slightly tighter. Nebulosity is very slightly tighter also. Interestingly, V5 preserved the original diffraction effect around the bright star and tightened up the diffraction spikes. My overall impression with this data is the difference is fairly subtle when the image is viewed at reasonable screen magnifications.

Looking at some of the results so far, imo it is almost like the V5 results are too good. That original data has no right looking so sharp, small and perfect. I usually use very light settings with BXT and my original data is usually reasonably good to begin with, so the difference BXT gives me is not extreme. I was happy enough when it was performing only a little better than traditional deconvolution and I could use it as an easy button in my workflow. To each their own I guess.

Ped · Aug 29, 2026 at 10:51 AM
John Hayes · Aug 29, 2026 at 05:31 AM
It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution.
John, you are much more knowledgeable about deconvolution techniques than I’ll ever be, but is this not precisely what AI-based tools do? Non-analytic solutions, to me, read like they are not using rigid math formulas, but instead, learning from data to make predictions or estimates. This refers to probabilistic reasoning, which is how AI handles complex, real-world problems.
When an LLM generates text, it is not copying and pasting from a database or following a rigid, step-by-step mathematical formula. Instead, it relies entirely on neural net estimation to process your prompt, calculate the mathematical probabilities of what words should follow, and output a non-analytic, best guess solution one token (a word or fragment of a word) at a time.
- Pedro

Judging from the responses here and the widespread use of the term “AI” to refer to neural net analysis, there is pretty widespread misunderstanding of what “AI” means. Yes, artificial intelligence incorporates neural networks, but just as individual neurons in the brain are not intelligent, neither are individual neural networks. LLMs generate new and generally unique output in response to an input that can appear to be reasoning, which is why they are called “generative”. The neural net in BXT is doing no such thing. It is not generating something new and it is not using what we call “AI” in LLMs to generate output.

I’m not privy to the exact details of the inner workings of BXT but I can tell you in principle what it is doing and it operates a lot more like a look up table than a LLM. It starts with a large array of “very-sharp” image snippets from astronomical images (such as HST data). The size of the snippets might be in the range of 64 × 64 up to maybe 256×256 pixels. These snippets are then convolved with a range of PSF functions that include the effects of optical aberrations, atmospheric seeing, guide errors, and other calculable optical effects to create a blurred snippet. The size of this array of small images might include anywhere from 300,000 to 1,000,000 entries to cover small differences between all of the parameters, which are generally randomly generated over a selected range of parameters. The job of the neural net training is to find a match between the patch-wise image data and the training data that minimize the rms difference between the two data sets. At that point, the sharpened result is given by the original “very-sharp” image snippet that was used to create that “best-fit” blurred snippet. The clever thing here is that the NN works backwards to a known result! The output parameters selected by the user determine how this data is used in the final result with respect to the amount of sharpening the user wants to apply. A neural net is a very efficient way to solve the extremely large number of simultaneous equations needed to find the best fit solution.

It is the “best-fit” characteristic of matching the data that leads to referring to this method as a “estimation method.” It is “estimating” a mathematically rigorous best fit from a large collection of possible solutions. When it is properly implemented, the differences between adjacent choices can be very small! As I pointed out earlier, the ultimate limitations of this approach lie in the density of the training data and in the SNR of the original data.

NN Estimation is VERY different than analytic solutions, which aim to “undo” the effects of convolution and which often require iterative methods. Analytic solutions include Unsharp Masking, Richardson–Lucy, Wiener filter Inverse filtering, and Multiscale sharpening, which are all fairly sensitive to noise and prone to ringing and other sharpening artifacts.

BXT is not an “AI” solution and it would be better if it were more correctly called a “Neural Net” or a “Neural Net Estimation Method”.

- John

Die Launische Diva · Aug 29, 2026, 03:52 PM
The next logical step is to have a future version that can remove the spikes entirely and reconstruct the image as if it were shot with a perfect refractor. The price for the update will be in the “perfect refractor class” range 🤣

Ha, ML5 already does that sometimes. Here are a few crops from the same (!) image, pre and post correct-only ML5.

In this part of the image, it almost completely removes the spikes:

In this part of the image, it shifts the spikes around:

In this region, it slightly dims the spikes…

While making them brighter in another.

Here it does the same but only for the horizontal spikes:

And finally it leaves them more or less alone here:

One image, six different fun effects on the spikes…

Here I used it on an SHO image from a newtonian with (admittedly sligthly misaligned) double veins. This means the scope produces spikes that are dotted, especially in narrowband. For some spikes especially on the edges, ML5 just straight up changed my double-veined spikes into single-veined spikes, shown best in the Oiii channel:

Here on an RGB image it changed the spikes completely again - different interference frequency and different colors:

To me all this makes it seem like what ML5 is doing is closer to full star replacement than just correction.

To me, this crosses the line of acceptability.

The stars are transformed to the point where you almost cannot tell they came from the original dataset (especially when you consider the halo replacement), and I think that’s the point at which I would not consider an image produced with Blurx ML5 to still be “mine”, nor an accurate representation of what the telescope captured originally. Since every single comparison here is a correct-only run, the weakest setting, this cannot be avoided by adjusting the parameters, as is the case with oversharpening. One of if not the most beautiful parts of astrophotography to me is that we can capture pictures that are simultaneously pretty while also being fairly faithful representations of real astronomical objects. There’s room for creativity and artistry but it’s always grounded in reality, in the data that is projected on our cameras by our telescopes. For me the amount of “correction” that ML5 is too much and I would consider the resulting stars to be almost synthetic, both in how they look and how they compare to the original data.

Scott Badger · Aug 29, 2026 at 12:44 PM
Heck, let’s ask the source……
CGPT sez:

Well…CGPT says even more. It is after all, “generative”, so we’ll see different answers when we ask similar questions.

================ FROM CGPT ====================

… And this is the REALLY important part

Croman deliberately avoids training BXT to produce images that merely look good.

This is crucial. A neural network could be trained using something like a perceptual or SSIM-type loss: "Does the result look sharper and more visually pleasing?" That would be dangerous for astronomical deconvolution because the network could learn to insert plausible-looking structure.

Instead, BXT uses reconstruction losses based on the actual known ground truth. RC Astro explicitly says that perceptual, structural-similarity, and adversarial losses are not used for ML5; the losses are reconstruction terms that penalize departures from the known target. That's a very important distinction.

Imagine two possible outputs:

A. Looks spectacularly sharp.

B. Is numerically much closer to the actual original image.

BXT's training objective is designed to favor B. That's why calling it an "AI sharpening filter" is actually quite misleading.

===================================

This output contains some AI slop. BXT isn’t trained merely to merely “favor B”. It is trained using minimum rms departures which is the ONLY method used. There is no capacity to produce “spectacularly sharp” results. Again, the NN method used in BXT is not AI.

- John

Mikołaj Wadowski · Aug 29, 2026, 07:22 PM
Die Launische Diva · Aug 29, 2026, 03:52 PM
The next logical step is to have a future version that can remove the spikes entirely and reconstruct the image as if it were shot with a perfect refractor. The price for the update will be in the “perfect refractor class” range 🤣
Ha, ML5 already does that sometimes. Here are a few crops from the same (!) image, pre and post correct-only ML5.
In this part of the image, it almost completely removes the spikes:
In this part of the image, it shifts the spikes around:
In this region, it slightly dims the spikes…
While making them brighter in another.
Here it does the same but only for the horizontal spikes:
And finally it leaves them more or less alone here:
One image, six different fun effects on the spikes…
Here I used it on an SHO image from a newtonian with (admittedly sligthly misaligned) double veins. This means the scope produces spikes that are dotted, especially in narrowband. For some spikes especially on the edges, ML5 just straight up changed my double-veined spikes into single-veined spikes, shown best in the Oiii channel:
Here on an RGB image it changed the spikes completely again - different interference frequency and different colors:
To me all this makes it seem like what ML5 is doing is closer to full star replacement than just correction.
To me, this crosses the line of acceptability.
The stars are transformed to the point where you almost cannot tell they came from the original dataset (especially when you consider the halo replacement), and I think that’s the point at which I would not consider an image produced with Blurx ML5 to still be “mine”, nor an accurate representation of what the telescope captured originally. Since every single comparison here is a correct-only run, the weakest setting, this cannot be avoided by adjusting the parameters, as is the case with oversharpening. One of if not the most beautiful parts of astrophotography to me is that we can capture pictures that are simultaneously pretty while also being fairly faithful representations of real astronomical objects. There’s room for creativity and artistry but it’s always grounded in reality, in the data that is projected on our cameras by our telescopes. For me the amount of “correction” that ML5 is too much and I would consider the resulting stars to be almost synthetic, both in how they look and how they compare to the original data.

Well, first, thank you for taking the time to make all these comparisons. I consider myself open-minded and welcoming to new tools, but after seeing your examples, this crosses the line, especially because the resulting correction can be random and depends on the neighbors of a given star ☹️

John Hayes · Aug 29, 2026, 07:07 PM
Ped · Aug 29, 2026 at 10:51 AM
John Hayes · Aug 29, 2026 at 05:31 AM
It uses the process of neural net estimation to arrive at a non-analytic, “best guess” solution.
John, you are much more knowledgeable about deconvolution techniques than I’ll ever be, but is this not precisely what AI-based tools do? Non-analytic solutions, to me, read like they are not using rigid math formulas, but instead, learning from data to make predictions or estimates. This refers to probabilistic reasoning, which is how AI handles complex, real-world problems.
When an LLM generates text, it is not copying and pasting from a database or following a rigid, step-by-step mathematical formula. Instead, it relies entirely on neural net estimation to process your prompt, calculate the mathematical probabilities of what words should follow, and output a non-analytic, best guess solution one token (a word or fragment of a word) at a time.
- Pedro
Judging from the responses here and the widespread use of the term “AI” to refer to neural net analysis, there is pretty widespread misunderstanding of what “AI” means. Yes, artificial intelligence incorporates neural networks, but just as individual neurons in the brain are not intelligent, neither are individual neural networks. LLMs generate new and generally unique output in response to an input that can appear to be reasoning, which is why they are called “generative”. The neural net in BXT is doing no such thing. It is not generating something new and it is not using what we call “AI” in LLMs to generate output.
I’m not privy to the exact details of the inner workings of BXT but I can tell you in principle what it is doing and it operates a lot more like a look up table than a LLM. It starts with a large array of “very-sharp” image snippets from astronomical images (such as HST data). The size of the snippets might be in the range of 64 × 64 up to maybe 256×256 pixels. These snippets are then convolved with a range of PSF functions that include the effects of optical aberrations, atmospheric seeing, guide errors, and other calculable optical effects to create a blurred snippet. The size of this array of small images might include anywhere from 300,000 to 1,000,000 entries to cover small differences between all of the parameters, which are generally randomly generated over a selected range of parameters. The job of the neural net training is to find a match between the patch-wise image data and the training data that minimize the rms difference between the two data sets. At that point, the sharpened result is given by the original “very-sharp” image snippet that was used to create that “best-fit” blurred snippet. The clever thing here is that the NN works backwards to a known result! The output parameters selected by the user determine how this data is used in the final result with respect to the amount of sharpening the user wants to apply. A neural net is a very efficient way to solve the extremely large number of simultaneous equations needed to find the best fit solution.
It is the “best-fit” characteristic of matching the data that leads to referring to this method as a “estimation method.” It is “estimating” a mathematically rigorous best fit from a large collection of possible solutions. When it is properly implemented, the differences between adjacent choices can be very small! As I pointed out earlier, the ultimate limitations of this approach lie in the density of the training data and in the SNR of the original data.
NN Estimation is VERY different than analytic solutions, which aim to “undo” the effects of convolution and which often require iterative methods. Analytic solutions include Unsharp Masking, Richardson–Lucy, Wiener filter Inverse filtering, and Multiscale sharpening, which are all fairly sensitive to noise and prone to ringing and other sharpening artifacts.
BXT is not an “AI” solution and it would be better if it were more correctly called a “Neural Net” or a “Neural Net Estimation Method”.
- John

You’re conflicting the umbrella term AI with generative AI.

The fist sentence from rcastro’s page on blurx reads: “BlurXTerminator is an ML-powered deconvolution tool designed specifically for astronomical images.”. Machine learning is a subset of artificial intelligence. Therefore blurx is AI. Even RC calls it that:

You are correct, it is not generative, but you’re the only one in this thread who used that term anyway.