starnet++ vs. starXterminator - Forums

Topic starter

Jun 13, 2023

Hi Community,

is there anything wrong with starXterminator ? Are there any hidden functions to get better results ? The whole world tells me that starXterminator is the best you can get for eliminating starts but I got several examples where starnet++ is much better then starXterminator. See here the faint nebulas in NGC6960. From my point of view, you need both programs and you have to try which one gives the best results for every picture.
Does anyone have the save experiences ?

starXterminator:

starnet++:

Helpful Engaging

Dave Ek

Jun 13, 2023

Generally, I feel like i get fewer star removal artifacts from StarXTerminator, but that will also depend on how challenging is the image. It’s certainly possible that Starnet++ does just as well (or better) on some images. You have nothing to lose by trying both and choosing whichever gives you the best results for that image. My impression is that StarXTerminator does a bit better job at filling in the space where a star was removed. I’m usually pretty happy with StarXTerminator’s result, but not always with Starnet++, mostly when there are lots of stars in the image.

Personally, I’m just extremely grateful to have such tools at my disposal for image processing.

Helpful Concise

Norman Fuerst

Jun 19, 2023

Which AI version do you use? I got better results when I switched from AI 11 lite to the normal AI 11 version, if your computer can handle it.
I think StarXTerminator leaves often cleaner backgrounds for my taste but it also depends on the object and I normally try both with all the available settings and sometimes I also use it a second time on the same image which can help to remove remaining big stars.
I also test all stride versions on starnet++ if the first result is not satisfactory.

My biggest problem is when I can't get rid of a big star with both tools and clone-stamping these out might be the only solution.
I wonder what other people do in these situation and if thats considered ethically correct :-)

Helpful Engaging Supportive

Zak Jones

Jun 20, 2023

I think StarXTerminator wins without a doubt.

I got it when it was first released in 2021, and the improvements that it has now compared to back when it first came out are mind blowing, mainly thanks to the hard work of Russ Croman in improving all of his AI astrophotography tools. I do acknowledge Nikita Misiura's awesome StarNet++ that really started the revolution of using AI in astrophotography, but when comparing StarNet++/StarNet 2 against StarXTerminator, in my view there is no comparison. I understand that some people might think otherwise, but I am just expressing my opinion that's all.

If StarXTerminator can't get rid of all the stars, Russ said to me that you can run it twice and it should clean up the remaining leftover stars. If any halos or artifacts remain after using StarXTerminator, I save my image as a 16-bit TIFF and import it into the latest version of Photoshop where I use the new Remove tool. It does an amazing job at removing halos and leftover artifacts, although sometimes it doesn't quite get it right. If this happens, all I do is re-draw the area where the artifact first was, and then it usually cleans the rest of it up nicely. Funny thing is it's powered by AI lol.

Zak

Helpful Engaging

Dale Penkala

Jun 20, 2023

Zak Jones:
I think StarXTerminator wins without a doubt.

I got it when it was first released in 2021, and the improvements that it has now compared to back when it first came out are mind blowing, mainly thanks to the hard work of Russ Croman in improving all of his AI astrophotography tools. I do acknowledge Nikita Misiura's awesome StarNet++ that really started the revolution of using AI in astrophotography, but when comparing StarNet++/StarNet 2 against StarXTerminator, in my view there is no comparison. I understand that some people might think otherwise, but I am just expressing my opinion that's all.

If StarXTerminator can't get rid of all the stars, Russ said to me that you can run it twice and it should clean up the remaining leftover stars. If any halos or artifacts remain after using StarXTerminator, I save my image as a 16-bit TIFF and import it into the latest version of Photoshop where I use the new Remove tool. It does an amazing job at removing halos and leftover artifacts, although sometimes it doesn't quite get it right. If this happens, all I do is re-draw the area where the artifact first was, and then it usually cleans the rest of it up nicely. Funny thing is it's powered by AI lol.

Zak

I tend to agree with @Zak Jones I really prefer SX over SN but again others have different views. I think it does depend on the image itself but for me and my experience has been SX is the best overall for my images.

Dale

Werner Stumpferl

Topic starter

Jun 20, 2023

Regarding stars, SXT gives mostly much better results, especially with big stars and halos, but I was surprised about (maybe only in this case) that it was so bad with the faint Nebula. Result for me is that I will use both together and if necessary I will combine both results.

Werner

gfunkernator

Jun 20, 2023

As much as I love how much faster SN is than SX, SX does a much better job removing stars for me. SN has always left artifacts. I use the SX Photoshop plugin, so no CUDA acceleration, stuck with CPU, which can take up to 15min working with 2x Drizzled stacks and large tiles.

Zak Jones

Jun 20, 2023

As much as I love how much faster SN is than SX, SX does a much better job removing stars for me. SN has always left artifacts. I use the SX Photoshop plugin, so no CUDA acceleration, stuck with CPU, which can take up to 15min working with 2x Drizzled stacks and large tiles.

I actually managed to get GPU acceleration in Photoshop for StarXTerminator and NoiseXTerminator thanks to Russ providing a guide on his website.

All I had to do is replace the tensorflow file with the one I used with my PixInsight tensorflow file in the respective file directories. Not sure if people need to do the same thing as I did when setting up GPU acceleration for PixInsight though. Russ didn't specify on his website.

When I did this, it was game changing. SXT and NXT run so much faster with GPU acceleration in Photoshop now. Even though I use PixInsight for most of my main workflow, when I need to use SXT or NXT in Photoshop I will without the need to worry about slow waiting times.

Here's the link to Russ' website where he talks about enabling GPU acceleration for PixInsight and Photoshop if anyone's interested: https://www.rc-astro.com/gpu-acceleration-for-ai-powered-tools/

Zak

Helpful Engaging

gfunkernator

Jun 20, 2023

@Zak Jones Well slap me in the mouth with that! I didn't know that was possible. I guess I could just copy the tensorflow.dll that starnet++ uses? I'll have to read through the tutorial. Thanks man!

andrea tasselli

Frequent contributor

Jun 20, 2023

Curiously GPU acceleration for SX doesn't work for me despite following all the steps. May it be because of Win 11?

gfunkernator

Jun 20, 2023 · Edited: Jun 20, 2023

For the StarXterminator Photoshop plugin, assuming you've set up CUDA for Starnet++, just copy the tensorflow.dll from your Starnet++ folder to C:\Program Files\Common Files\RC-Astro\StarXTerminator. Don't overwrite the existing tensorflow.dll, just rename it to tensorflow.dll.old before you copy the Starnet++ tensorflow.dll. I just tested it on a cropped TIFF with large tiles, 5 seconds it was done. Amazing!

EDIT:
I just did another test with an uncropped, 6720x4480 2x drizzled image, a little over 2GB, large tile overlap, took about 53 seconds, whereas on the CPU alone, took 10min.

Helpful Engaging Supportive

andrea tasselli

Frequent contributor

Jun 20, 2023

For the StarXterminator Photoshop plugin, assuming you've set up CUDA for Starnet++, just copy the tensorflow.dll from your Starnet++ folder to C:\Program Files\Common Files\RC-Astro\StarXTerminator. Don't overwrite the existing tensorflow.dll, just rename it to tensorflow.dll.old before you copy the Starnet++ tensorflow.dll. I just tested it on a cropped TIFF with large tiles, 5 seconds it was done. Amazing!

EDIT:
I just did another test with an uncropped, 6720x4480 2x drizzled image, a little over 2GB, large tile overlap, took about 53 seconds, whereas on the CPU alone, took 10min.

I was talking about PI. GPU does move one tiny bit. CUDA 12.1 installed...

gfunkernator

Jun 20, 2023 · Edited: Jun 20, 2023

andrea tasselli:
Curiously GPU acceleration for SX doesn't work for me despite following all the steps. May it be because of Win 11?

Which guide did you follow? I used this tutorial and I am on Win 11. I installed CUDA 11.7. The environment variables and placing the files correctly are crucial to getting tensorflow to use your GPU.

EDIT: Sorry I posted before I saw your response. I tried CUDA 12 and it didn't work, had to revert to 11.7.

@Zak Jones Thanks for steering me the right direction! I never thought to just look at the SX folder to just replace the tensorflow.dll. Now I'm going to be more aware of any other apps that use tensorflow and try using the CUDA version.

andrea tasselli

Frequent contributor

Jun 20, 2023

Which guide did you follow? I used this tutorial and I am on Win 11. I installed CUDA 11.7. The environment variables and placing the files correctly are crucial to getting tensorflow to use your GPU.

EDIT: Sorry I posted before I saw your response. I tried CUDA 12 and it didn't work, had to revert to 11.7.

@Zak Jones Thanks for steering me the right direction! I never thought to just look at the SX folder to just replace the tensorflow.dll. Now I'm going to be more aware of any other apps that use tensorflow and try using the CUDA version.

Darn! I'll have to remove the 12.1 CUDA and put 11.8 in its place. And yes, the tutorial is the same. Followed it to the letter.

Dan H. M.

Jun 21, 2023

I initially thought SXT was the clear winner but on an image I'm working on now only SN2 works to remove all stars from the linear integration. I do still think SXT does a better job when it works, but, of course, it has to actually remove all stars from the image to be useful.

Wei-Hao Wang

Jun 21, 2023

I prefer SN2 very much over SXT. I downloaded SXT for trial, and found that it does not just remove stars. It also removes some faint features around stars. If you simply look at the starless image, you won't feel it. But if you look at the stars that it remove (the difference between the original and the starless one) by stretching it, you will see those features that are not supposed to be removed. If a starless image is what you want as an end product, this doesn't matter. But if you want to separate the stars and nebulas to process them differently and then combine, this will be a very big problem.

So at the end, I purchased BXT and NXT, but not SXT. The above was my experience quite some months ago. Many SXT has improved since then.

Helpful Insightful Respectful Engaging

Ashraf AbuSara

Jun 21, 2023

andrea tasselli:
Curiously GPU acceleration for SX doesn't work for me despite following all the steps. May it be because of Win 11?

I assume it works for me on Win 11 because it takes about 30s in most instances for me with my RTX4090 while using 2x drizzle on a 2600mm pro frame.

Zak Jones

Jun 22, 2023

@Zak Jones Well slap me in the mouth with that! I didn't know that was possible. I guess I could just copy the tensorflow.dll that starnet++ uses? I'll have to read through the tutorial. Thanks man!

Neither did I lol.

Good on Russ for pointing this out in the first place! I'm glad people are getting benefits from using GPU in Photoshop now.

All good! That's why I posted it for, so everyone knows about this now!

Zak

Zak Jones

Jun 22, 2023

andrea tasselli:
Which guide did you follow? I used this tutorial and I am on Win 11. I installed CUDA 11.7. The environment variables and placing the files correctly are crucial to getting tensorflow to use your GPU.

EDIT: Sorry I posted before I saw your response. I tried CUDA 12 and it didn't work, had to revert to 11.7.

@Zak Jones Thanks for steering me the right direction! I never thought to just look at the SX folder to just replace the tensorflow.dll. Now I'm going to be more aware of any other apps that use tensorflow and try using the CUDA version.

Darn! I'll have to remove the 12.1 CUDA and put 11.8 in its place. And yes, the tutorial is the same. Followed it to the letter.

Yeah unfortunately it didn't work for me either, which is a bummer...

I had to do the same: uninstall 12.1 and re-install 11.8, which works now like a charm.

Hopefully in the future there will be a compatibility update so we can use a newer version of CUDA to further increase our GPU speeds when processing with SXT, NXT etc in both PixInsight and Photoshop.

Zak

Supportive

Zak Jones

Jun 22, 2023

andrea tasselli:
Curiously GPU acceleration for SX doesn't work for me despite following all the steps. May it be because of Win 11?

Which guide did you follow? I used this tutorial and I am on Win 11. I installed CUDA 11.7. The environment variables and placing the files correctly are crucial to getting tensorflow to use your GPU.

EDIT: Sorry I posted before I saw your response. I tried CUDA 12 and it didn't work, had to revert to 11.7.

@Zak Jones Thanks for steering me the right direction! I never thought to just look at the SX folder to just replace the tensorflow.dll. Now I'm going to be more aware of any other apps that use tensorflow and try using the CUDA version.

All good! Glad I could help out!

It's interesting because I didn't even know about that too until I saw Russ' guide on GPU acceleration on his website. Yes, I would see if any other apps use the GPU now thanks to the CUDA installations etc.

Zak