Applying BXT as final step in processing

I'm not sure I would call it "interesting", at least in any positive way, I'm afraid. The short answer is: don't. The longer answer is: you can apply unsharp mask to better effect (with provisions) than BXT outside where it is reasonable for it to be applied and still produce realistic results (to a degree).

As I expected, this is a bad idea. Thanks for the comments. This is why I didn't post the B revision as the final 

I have experimented with this quite a bit, using BXT as a sharpening tool on non-linear data.  On undersampled data it can sometimes give the illusion that theres more detail in there, but it's really hard to not make it look cooked, and as mentioned above, artifact-free.  At the time it was kind of a novel ideal (for me) and I thought it could be useful, but ultimately I decided that it was doing more harm than good, and no longer use the tool this way.  There are definitely better sharpening methods out there and to be honest, I dont generally sharpen either. I find that a really light application of LHT with something like the settings below can make the details and the image pop a little better.  Gotta be careful with any tool though, it's easy to overcook the image.

Tony Gondola:
I say don't worry about the "rules". Everyone's data and process is different. This hobby if full of rules that are just heard and passed down as gospel. Do the AB testing, if it benefits your image, do it. It doesn't matter what everyone else thinks.

One reason the “rules” exist (as does the RCC forum here) is that we tend to be pretty poor judges of our own images. Beginner to intermediate imagers tend to be especially bad at seeing things in the moment that they easily see after a few days and then being disappointed with the results they posted. It’s sort of like writing where there are “rules” of grammar and style, but great writers know when they can get away with not applying them for some specific purpose.

You could do a lot worse in processing your images than just following the “rules” until you get very good at it. Experimentation is great, but focus it on areas that are more creative and a less firm technical answer. Experiment with palettes and stretches, HDRMT, saturation, soft light, those kinds of things that are creative choices rather than things like “do deconvolution first” which is a “rule” because of good technical reasons.

Speaking for myself, I never use BXT on anything but linear images very early in the workflow. That is because, as others have noted, it is designed for linear data and Russ has been very clear on that point.

On the subject of "looking overcooked" that can happen with any deconvolution of any kind at any time. It often results in what is commonly called "threadlike detail".  OTOH, some objects, especially some planetary nebula, actually do have threadlike detail so one has to be careful in assuming that all such detail is a result of "BXT overcooking". My rule is to look at the unconvolved image after giving it a temporary stretch (just to examine it, you do the BXT on unstretched data, of course) and see if there is a hint of such structure. Using the BXT real time preview and blinking it on and off is helpful as well here to see what the enhancement is doing and where. If there is such real detail, then "looking overcooked" may not be actually overcooked and the enhanced structure may be real. A perfect example is Abell 30 that I processed yesterday, which is why this comes to mind. Pretty much all the images of that object show the detail to some degree, even including images from the Chandra X-ray telescope. That leads to another "reality check" and that is professional images of the same object, where available.

The biggest deconvolve error is to use non-astro AI enhancement programs on non-linear astro images. I have seen a small handful of images on AB that have done that and they look nothing short of bizarre. I will be kind and hope the imagers were just being "artistic". 

It used to “kinda” work as a sort of AI sharpening, albeit not as well as when done at a proper stage for deconvolution.

But in an updated model concern for maintaining any degree of performance in stretched images was set aside in favor of optimal performance as a deconvolution tool, and it now can be quite problematic when not used at a proper stage for deconvolution (resulting in strange rendering and hallucination of stars; odd handling of various details, etc.). And so it is recommended not to use it outside those conditions (generally and by the developer). Which is great in my book: it excels at doing what it is meant to do.

Edit: Typos.

It works very well on the great majority of my images. It's apparently popular with the Hubble team. And it's reasonably priced. I would not want to be without it.

It is a free world, so everyone can try whatever makes him happy  with his image. However , i think in our hobby we should pay some respect also to the nature and physics .  Deconvolution does not make sense if  any kind of  noise reduction has been done before .. Also the PSF should be derived from the image itself (stars are the best measure of it) , not determined arbitrary because image looks nicer/sharper.  So, BXT at the end of processing is perhaps not a good idea.  Deconvolution is a bit of a paradox, if one wants to do deconvolution accurately , it requires high (statistically) quality data . But when one has high quality data deconvolution is not that much needed anyway. 
So the art is to use it with enough caution 
CS

Russell Croman as made it very clear on the Astro Imaging Channel and a discussion with Adam Block that BlurXTerminator should only be applied to linear data. He explained it is not a shatpening tool.  I view it as a final step of the calibration process, and not a post processing tool. It is deconvolution and needs unproceesed linear stars to optain correct meassuerment of the convolution in order to apply deconvolution to the image.  It will work on non-linear data BUT can create artifacts  and false pixels.  You might as well use Topez DeNoise.

Lynn K.

Math is great but that tends to idealize everything. The real world is messy, at the end of the day we're talking about manipulating pixels. There are so many variables in processing, blending and so on that yes, you can follow general rules and it's a good base but sometimes the results of doing something different will surprise you.

Wei-Hao Wang:
It may work, or may not.  (Here, "work" means it gives you a sharpening effect without obvious artifacts.  It will never give you a result that's mathematically equivalent to deconvolution.)

Wei-Hao,
I'm going to have to disagree with you here.  Properly implemented, a neural net solution can indeed produce a result "that's mathematically equivalent to a deconvolution."  I believe that Russ has done an outstanding job of being careful to apply the best form of the PSF and he's made it shift variant over the field to handle field aberrations.  There are, in fact, a number of advantages to solving this problem using a NN.  For one thing, the solution is much less influenced by noise.  It is also less likey to introduce ringing artifacts.

The hazard of applying BXT to non-linear data is that it violates the initial conditions assumed for the calculation, which will cause BXT to generate artifacts often referred to in the AI world as "hallucination".   The image below is from my recent talk at NEAIC about galaxy imaging illustrates the problem.  When you use BXT on stretched data, small features are turned into stars, stars are amplified, and many subtle features are suppressed.  This isn't a problem specific to a neural net solution.  You'll also get undesirable artifacts from traditional iterative solutions to the deconvolution problem when you use those algorithms incorrectly as well.

Imagers can do whatever they want but if you want to do a true deconvolution to bring out more "real" detail in an image, BXT should only be applied to linear data...and it should only be applied once.  Using BXT on non-linear data is not a good procedure that would tend to infect the archives here with "imaginary images".

John