Starnet++ v2

[The Lazy Astronomer]

Just a heads up to anyone interested and who doesn't already know, I've just discovered Starnet++ v2 was just released last week. There is a thread over on CN started by the developer with a link to download the files (not entirely sure if linking to a 'rival' forum is allowed, but if you Google "starnet v2" it'll be one of the top results).

Initial images coming through on that thread seem to show good results vs starnet v1 - I'm currently downloading 👍

Edit: files removed from SourceForge, download available from http://starnetastro.com/

Edited February 10, 2022 by The Lazy Astronomer

[Laurieast]

Thanks, it will be interesting to to compare it to StarXTerminator 

 

Edited January 31, 2022 by Laurieast

[AbsolutelyN]

Thanks for the heads up. Installed this morning and below is a quick test on an old image I had already run starnet v1 with.  

Thanks 

 

 

[scitmon]

Wow looks much better

[Dustspeakers]

Has anyone been able to download the command line version for Mac - not for PixInsight. The source forge version doesn't seen to include any binaries like rgb_starnet++ that was in version 1?

[SyedT]

Thanks for starting the thread!

For those with 64-bit Windows and NVIDIA GPUs, this webpage takes you through using the GPU for Starnet when using Pixinsight, which massively accelerates the process: https://darkskies.space/pixinsight-starnet-cuda/

Starnet V2 is a huge improvement on V1, both from a star removal and speed perspective. In PI, V1 took 1 min 55 seconds and V2 took just 55 seconds via the GPU-accelerated process described above. It's much better at identifying and removing larger stars and also leaves fewer and less obvious removal artefacts. 

Original:

 

V1:

 

V2:

Edited January 31, 2022 by SyedT

[Doversole]

It seems Starnet++ version 2 is really good at removing stars on a stretched image (is that what we call non linear?).

But ideally you would want to separate the stars from the DSO early on to process both images differently.

Trying to strech the result of a starless image I see some star artifcats appearing, not sure that is the case with StarXterminator.

[ONIKKINEN]

Did some tests with the new starnet compared to StarXterminator, i dont have the original starnet so wont be testing with that. This particular image is very noisy when stretched this far and so shows remaining artifacts very easily. Very pleased with how the new starnet reduced artifacts! Both run on the same 16bit stretched file.

StarXterminator:

Starnet V2:

Its not artifact free, but very close. The brightest stars still have a background patch of blue behind them, but much better than with StarXterminator.

[vlaiv]

7 hours ago, Doversole said:

t seems Starnet++ version 2 is really good at removing stars on a stretched image (is that what we call non linear?).

But ideally you would want to separate the stars from the DSO early on to process both images differently.

One way it can be handled is to do linear stretch. and save image as 16bit. 

(not sure if new version still requires 16bit images).

Another would be to do "reversible" stretch before creating starless image.

For linear stretch - just set white level somewhere just before target starts to saturate. Stars will saturate because of this - but that does not matter as they will be removed.

Reversible non linear stretch would consist out of three steps:

step 1 - same as above - lower white point almost down to saturation of target

step 2 - apply gamma of known power

step 3 - bring up black point just enough so you avoid clipping on left

in each step note down what you used. To get linear data again - do inverse steps in reverse order

- bring down black point by same amount

- do inverse gamma (with reciprocal of the power)

- move white point to original position

Now you can subtract such data from original image to have stars only image and you can also start stretching it again as you please.

 

[The Lazy Astronomer]

1 hour ago, vlaiv said:

One way it can be handled is to do linear stretch. and save image as 16bit. 

(not sure if new version still requires 16bit images).

Another would be to do "reversible" stretch before creating starless image.

For linear stretch - just set white level somewhere just before target starts to saturate. Stars will saturate because of this - but that does not matter as they will be removed.

Reversible non linear stretch would consist out of three steps:

step 1 - same as above - lower white point almost down to saturation of target

step 2 - apply gamma of known power

step 3 - bring up black point just enough so you avoid clipping on left

in each step note down what you used. To get linear data again - do inverse steps in reverse order

- bring down black point by same amount

- do inverse gamma (with reciprocal of the power)

- move white point to original position

Now you can subtract such data from original image to have stars only image and you can also start stretching it again as you please.

 

V2 does have an option to remove stars on "linear' data (inverted commas because it does it by doing similar to as you described above - stretching>removing stars>restoring to linear)

[ONIKKINEN]

22 hours ago, The Lazy Astronomer said:

V2 does have an option to remove stars on "linear' data (inverted commas because it does it by doing similar to as you described above - stretching>removing stars>restoring to linear)

Doesn't seem to work with 32-bit files though.

[The Lazy Astronomer]

2 minutes ago, ONIKKINEN said:

Doesn't seem to work with 32-bit files though.

Hmm. Does anyone know how it works within PI? I used it on an image that was working on in 64-bit (I know, I know, overkill) - does it change the bit depth of the image down to 16? I think PI still thought it was working in 64-bit, because I got a warning about how it couldn't save a 64-bit jpeg when I saved finished image out. 

[vlaiv]

1 hour ago, The Lazy Astronomer said:

Hmm. Does anyone know how it works within PI? I used it on an image that was working on in 64-bit (I know, I know, overkill) - does it change the bit depth of the image down to 16? I think PI still thought it was working in 64-bit, because I got a warning about how it couldn't save a 64-bit jpeg when I saved finished image out. 

Probably encodes image in 16 bit after stretch if it is still linear and then converts to 32bit before "unstretching" it.

One of good sides of applying gamma is handling of 16 bit data and introduced error.

Gamma is essentially power law which means that produced values are stretched at the low end and squeezed at the high end (compared to linear data)

If you stretch values a the lower end - and you round the values - you get less rounding error, and lower values will suffer less from rounding error in terms of SNR - because it's signal is already low. \

Overall it does not help much - but it does help a bit and every little bit counts.

[The Lazy Astronomer]

To add a note to this: I've been playing around with using it on linear data, and it seems to only work properly if you run it with stf active. I'm not really sure why...?

Also, the check box to create a star mask does not seem to generate satisfactory results when using linear data, so you have to do it manually by subtracting in pixelmath.

Those little niggles aside, it seems to work very well at removing the stars.

For background: I was trying to see if I could find an even easier (than EZ Decon) way to do deconvolution by removing stars while linear, doing deconvolution without the need for star masks or deringing, then adding stars back. The issues with SNv2 mentioned above took me an embarrassingly long time to figure out and I didn't quite get far enough along in the process to see if the process worked well or not - please someone tell me if you think what I'm trying to do is a stupid idea and won't work well because of some reason I've overlooked or haven't thought of!

[david_taurus83]

Wondering if anybody has downloaded this if they would kindly send me the zip folder for Windows 64 bit install? WeTransfer maybe? PM for email address.

It has been removed from sourceforge as the author has not made it open source. He did say on CN when he first released it there were going to be licence restrictions or something.

[david_taurus83]

Sorted now, thanks Martin! 🙂

[The Lazy Astronomer]

To save anyone else going through the CN topic, the files can now be downloaded from http://starnetastro.com/

[mike1485]

Experience seems to be that Starnet++ (even v2) works better if the image is stretched first, even if only mildly.  To this end people may like to know of the GHS stretching script for Pixinsight.  Using GHS your image can be stretched (remember the parameters you use) - you can then remove the stars with Starnet++ - and GHS then allows you to enter the same stretch parameters but check the "Invert Transformation" checkbox.  GHS will then apply an exact mathematical inversion of the original stretch to get you back where you started, but without stars.  I have just released version 2 of the GHS script if anyone is interested - the details are in this post: GHS Version 2 Release.

Edited March 7, 2022 by mike1485

[billhinge]

On 31/01/2022 at 15:04, SyedT said:

Thanks for starting the thread!

For those with 64-bit Windows and NVIDIA GPUs, this webpage takes you through using the GPU for Starnet when using Pixinsight, which massively accelerates the process: https://darkskies.space/pixinsight-starnet-cuda/

Starnet V2 is a huge improvement on V1, both from a star removal and speed perspective. In PI, V1 took 1 min 55 seconds and V2 took just 55 seconds via the GPU-accelerated process described above. It's much better at identifying and removing larger stars and also leaves fewer and less obvious removal artefacts. 

Original:

 

V1:

 

V2:

 

I was hoping to test starnetv2 against StarXterminator but starnetv2 seems to have disappeared from the internet!  - at least for mac users, the various download sites don't seem to work (yes I read the CN threads)

Anyway I wanted to try against a standard ref so found the above, this is the equivalent StarXterminator and NoiseXterminator with the same source

To me StarXterminator seems slightly better? (I'm thinking of buying, especially as starnetv2 doesn't play nice with Apple M1 whereas StarXterminator has neural engine support on the mac)

 

 

[david_taurus83]

8 hours ago, billhinge said:

 

I was hoping to test starnetv2 against StarXterminator but starnetv2 seems to have disappeared from the internet!  - at least for mac users, the various download sites don't seem to work (yes I read the CN threads)

Anyway I wanted to try against a standard ref so found the above, this is the equivalent StarXterminator and NoiseXterminator with the same source

To me StarXterminator seems slightly better? (I'm thinking of buying, especially as starnetv2 doesn't play nice with Apple M1 whereas StarXterminator has neural engine support on the mac)

 

 

Has StarXterminator sorted out the over-softening of the star areas? When I trialled it, it was very good and very quick but it over softened the area of where the stars were and it didn't look right over a noisy background. I've stuck with Starnet2. I have bought the excellent NoiseXterminator though.

[geeklee]

7 minutes ago, david_taurus83 said:

Has StarXterminator sorted out the over-softening of the star areas? When I trialled it, it was very good and very quick but it over softened the area of where the stars were and it didn't look right over a noisy background. I've stuck with Starnet2. I have bought the excellent NoiseXterminator though.

I think this was one of the earliest things that was "fixed".  I found it very noticeable and didn't like it much either. 

If you check https://www.rc-astro.com/resources/StarXTerminator/ and look down the "AI Version History" table near the bottom, it was mainly resolved in AI6 (Nov 2021).  It's not something I notice now.

EDIT: It's being frequently updated as well.

Edited September 16, 2022 by geeklee

[Laurieast]

8 hours ago, billhinge said:

I was hoping to test starnetv2 against StarXterminator but starnetv2 seems to have disappeared from the internet!  - at least for mac users, the various download sites don't seem to work

Is this it? 
https://www.starnetastro.com/download/

[billhinge]

54 minutes ago, Laurieast said:

Is this it? 
https://www.starnetastro.com/download/

How did you find that! I spent about 2 hours googling and reading CN threads last night. Thanks I have to integrate it onto a new M1 macbook now, unlike my last macbook it doesn't have cuda support

[Laurieast]

32 minutes ago, billhinge said:

How did you find that! I spent about 2 hours googling and reading CN threads last night. Thanks

Your welcome! All I did was:

 

Maybe it was the way I asked it? Who knows, good luck with it, I like it.

Edited September 16, 2022 by Laurieast

[The Lazy Astronomer]

12 hours ago, billhinge said:

 

I was hoping to test starnetv2 against StarXterminator but starnetv2 seems to have disappeared from the internet!  - at least for mac users, the various download sites don't seem to work (yes I read the CN threads)

Anyway I wanted to try against a standard ref so found the above, this is the equivalent StarXterminator and NoiseXterminator with the same source

To me StarXterminator seems slightly better? (I'm thinking of buying, especially as starnetv2 doesn't play nice with Apple M1 whereas StarXterminator has neural engine support on the mac)

 

 

I don't mean to embarrass you, but there was a link literally 2 posts above yours 🤣