289 replies to this topic

[rcroman]

In keeping with forum rules, please submit any questions and/or report any issues with NoiseXTerminator on this thread.

 

The beta test announcement can be found here: https://www.cloudyni...trophotography/

 

[pyrasanth]

I tested this tonight on a very noisy image & it cleaned up superbly. I have all your tools so I will be adding this to my arsenal of essential tools.

 

It is far easier to use than demure denoise and it perhaps produces result which are comparable or better. It was really nice to see the noise completely vanish and using this in conjunction with normalised scale gradient we are on our way to a far beter class of images.

 

What I like is the elegant simplicity- no need to go through that process of measuring darks & flats to get the best values for demure- yours just works out of the box and its very fast as well.

 

There is nothing not to like across your excellent suite of tools so thank you for your contribution to our cause.

 

Best wishes

Mark

Edited by pyrasanth, 30 April 2022 - 03:48 PM.

[rcroman]

I tested this tonight on a very noisy image & it cleaned up superbly....

Thanks for your kind words, Mark... glad it's working well for you.

[pyrasanth]

Thanks for your kind words, Mark... glad it's working well for yThanks Russel,

Thanks Russel,

 

The samples posted below illustrate the point- these were just straight out of the box settings on an old processed image:

 

Original

 

 

Clean

 

[Marcelofig]

Wow, I just tried it on my Mac mini M1, it took about 1 second.

 

Very clean results, amazing !!!

[pyrasanth]

There is a typo on your web page for the trial- the text refers to Starxterminator & not noisexterminator but its okay as I have that as well

[ChrisWhite]

Looking forward to trying this RC.

Do you remember the passionate discussion about topaz denoise ai?

Would you be able to comment on how this ai noise reduction program avoids the ethical boundary pushing that previous ai noise reduction algorithms engage in? What makes this less controversial? I know you mentioned briefly about not generating details (data) that isnt there, but what else can you disclose without revealing trade secrets.

Thank you for the work you do.

Edited by ChrisWhite, 30 April 2022 - 05:28 PM.

[rockstarbill]

Tested this out and it is very effective. Nice work!

[rcroman]

Looking forward to trying this RC.

Do you remember the passionate discussion about topaz denoise ai?

Would you be able to comment on how this ai noise reduction program avoids the ethical boundary pushing that previous ai noise reduction algorithms engage in? What makes this less controversial? I know you mentioned briefly about not generating details (data) that isnt there, but what else can you disclose without revealing trade secrets.

Thank you for the work you do.

Hi Chris... yep, I remember that discussion well. It's still a hot topic, especially in our field where there is a bias toward being "documentary" and not misrepresenting objects we shoot. It's even hotter now that newer AI algorithms are on the scene with some strong capabilities/proclivities for inventing information.

 

Technically the way NoiseXTerminator tries to walk this tightrope is that its neural network is not "generative." Generative networks invent new information by some process or other, often using a combination of features they have been trained on plus an element of randomness. Google "deep dream" to see what's out there. Fantastic, but definitely fictitious.

 

There is not a sharp line here... it's not purely black and white. If an algorithm can "undo" a gaussian blur or up-scale an image and fill in detail, then strictly speaking it has to be inventing information. I think of it as a question of how aggressively this is done. The software you mentioned can be *very* aggressive. It does a reasonable job on general photos, but easily goes off into the weeds with astrophotography and tends to create detail for the sake of appearing "sharp" rather than aiming for accuracy. My goal in training NoiseXTerminator's network was to aim closer to a "maximum likelihood" result rather than intentionally trying to get the network to be inventive. That and teaching it not to mangle the bloody star profiles. 

[ChrisWhite]

Hi Chris... yep, I remember that discussion well. It's still a hot topic, especially in our field where there is a bias toward being "documentary" and not misrepresenting objects we shoot. It's even hotter now that newer AI algorithms are on the scene with some strong capabilities/proclivities for inventing information.

Technically the way NoiseXTerminator tries to walk this tightrope is that its neural network is not "generative." Generative networks invent new information by some process or other, often using a combination of features they have been trained on plus an element of randomness. Google "deep dream" to see what's out there. Fantastic, but definitely fictitious.

There is not a sharp line here... it's not purely black and white. If an algorithm can "undo" a gaussian blur or up-scale an image and fill in detail, then strictly speaking it has to be inventing information. I think of it as a question of how aggressively this is done. The software you mentioned can be *very* aggressive. It does a reasonable job on general photos, but easily goes off into the weeds with astrophotography and tends to create detail for the sake of appearing "sharp" rather than aiming for accuracy. My goal in training NoiseXTerminator's network was to aim closer to a "maximum likelihood" result rather than intentionally trying to get the network to be inventive. That and teaching it not to mangle the bloody star profiles.

Thanks for the additional information. I'm looking forward to giving it a test run.

[rockstarbill]

Some examples, first a test on a non-linear finished image. I cropped the image down to just the Galaxy in the center and a few fuzzies and some stars.

 

Before:

 

After:

 

I used a Denoise of 0.65 (down from the default of 0.90) and left detail at the default of 0.15. I felt this cleaned up the image nicely and still left some residual noise in the image so it did not look plastic-like or too phony as some NR can do to images. No mask was used at all.

 

[rockstarbill]

Next was a linear image I took from a stack of test frames (5 of them) and is only 45 mins of HA data taken with the Epsilon E160 and the QHY600. One callout, there are some residual artifacts in this image that are not created by NoiseX at all. I did not have darks, frames were not dithered, as they were intended as test images only. Nonetheless a good test of how well this tool can do NR on what should be pretty noisy data.

 

Before:

 

After:

 

I tried to pick a preview that had a good spread of nebulous regions, stars, and even included a star cluster in the image. From my eye it did an outstanding job on the data, and the silly artifacts that were left in my image were also untouched in the process.  

 

[licho52]

My verdict following initial testing: take my money (just not too much of it!)

 

My future purchase is also based on the continuing development of StarXterminator which has proved to be a great value and in many cases gives me the best results of all other star removal routines.

 

EDIT: I might add that it seems to be using CUDA and produces results within a few seconds on my RTX3070.

Edited by licho52, 30 April 2022 - 08:18 PM.

[rockstarbill]

I took this to a much noisier narrowband image that I happen to like quite a bit.

 

Revision J of the following image is without NoiseX and Revision K is with NoiseX. I did increase the strength to 0.75, but no other changes were made.

 

https://www.astrobin.com/380061

Edited by rockstarbill, 30 April 2022 - 08:33 PM.

[ChrisWhite]

I took this to a much noisier narrowband image that I happen to like quite a bit.

 

Revision J of the following image is without NoiseX and Revision K is with NoiseX. I did increase the strength to 0.75, but no other changes were made.

 

https://www.astrobin.com/380061

Very nice result! 

[rockstarbill]

Same with this data:

 

https://www.astrobin.com/khu01o

 

Revision M without, Revision N with. Left the strength at 0.75 this time, and it is probably right at the edge. Curious what others think. 

[rcroman]

My verdict following initial testing: take my money (just not too much of it!)

 

My future purchase is also based on the continuing development of StarXterminator which has proved to be a great value and in many cases gives me the best results of all other star removal routines.

 

EDIT: I might add that it seems to be using CUDA and produces results within a few seconds on my RTX3070.

Thanks for the positive feedback... this has been a ton of work, and very different from StarXTerminator. Which, BTW, is still in active development. Now that my GPUs are freed up from training NXT, SXT will be getting another round of training.

Good to know re CUDA support... I had not tested that yet. I assume you dropped a different TensorFlow library into your PixInsight installation?

[arbit]

Just tested on a wip OSC image of Trifid.

The results are very good. I am currently using Jon Ristas method via Ez Noise Reduction as the first NR step.

This was much faster (couple of seconds), no masks needed, and the parameters are intuitive. The sharpness slider is also very useful- it helped get the balance right very very quickly.

The only thing I felt were a bit off were the star profiles when using a higher strength.

One suggestion is to have test runs (with previews) with different settings all displayed at the same time as that makes it easy to compare different settings.

Some more experimentation is called for, but certainly looks great so far and definitely a purchase.

Edit: Should mention that my setup is also using CUDA, using a different tensorflow lib.

Sent from my SM-N960F using Tapatalk

Edited by arbit, 01 May 2022 - 12:20 AM.

[HubSky]

I use your StarXterminator and like it very much, and I use it the vast majority of the time, though occasionally I may like the results better with Starnet2.  So, it's still nice to sometimes have more than one tool to do the job. 

 

And I appreciate the legitimacy concerns using Topaz Sharpen AI and Topaz Gigapixel AI.  However, I don't have those concerns with Topaz Denoise and use it on every photo.  So having said that, if I already use Topaz Denoise AI, is there something fundamentally different with NoiseXterminator?  Does it work is some fashion that makes it better than Denoise AI or maybe makes it more "legitimate"? 

[rockstarbill]

I use your StarXterminator and like it very much, and I use it the vast majority of the time, though occasionally I may like the results better with Starnet2. So, it's still nice to sometimes have more than one tool to do the job.

And I appreciate the legitimacy concerns using Topaz Sharpen AI and Topaz Gigapixel AI. However, I don't have those concerns with Topaz Denoise and use it on every photo. So having said that, if I already use Topaz Denoise AI, is there something fundamentally different with NoiseXterminator? Does it work is some fashion that makes it better than Denoise AI or maybe makes it more "legitimate"?

It was trained on deep sky images, so you don't have to go and take out your stars, then move the data from the original processing app into Topaz, run the Denoise, then back into the original app to put your stars back in. I can Denoise the data directly in PI without all of that added fuss, and I don't have to take the stars out to do it.

My other concern with topaz is that the finished image is too clean, looks very fake, and usually looks better if one adds noise back in with an Astro processing app. I'm not personally down with that, and I'm lazy.

[Arie]

I just took a trial license and I like what I see.

Have been using Topaz for a long time already.

I, as many others, had to learn to get the dose right.

Just a wiff suffices. Only to dust it up a bit at the very end.
Its like how people handle pain killers.

If you take the right dose, it cures your problem.

Take an overdose, out of addiction, and it will kill you.

StarX can produce a plastic coated image too . . . .

 

I wish that the PI developers enable the graphics card for those calculationswhere it makes a difference. We have seen that the Cuda trick works very well with Starnet. 
Topaz gives you the option.

Edited by Arie, 01 May 2022 - 09:37 AM.

[licho52]

Exterminator modules work with CUDA.  It's true that PI is a bit slow on making the support work "out of the box" and on other modules.  It was announced a couple of years ago but no signs of being implemented yet.

[rockstarbill]

Exterminator modules work with CUDA. It's true that PI is a bit slow on making the support work "out of the box" and on other modules. It was announced a couple of years ago but no signs of being implemented yet.

What's the setup process?

[Arie]

https://www.darkskie...t-starnet-cuda/

 

https://www.williaml...om/starnet-cuda

Edited by Arie, 01 May 2022 - 01:30 PM.

[rockstarbill]

Ahhh yes I recall going through this a long time ago with the v1 Starnet.