One picture is worth ten thousand words.
I have uploaded a starless version of M81 + M82. The image has been calculated by the new version 2 of StarNet.
https://www.astrobin.com/miuye8/
Comparison with the normal image:
https://www.astrobin.com/zvha0y/
Clear skies, AstroEdy
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
I have also done some comparisons between Starnet2 and StarXTerminator. The results are mixed but I would lean towards Starnet2 being slightly better for now.
1st example: Top of the Elephant's Trunk. Starnet2 (SN, left) removes more of the very dim red star residuals than StarXTerminator (SX)
2nd example: The Wizard's nose. SN (left) again removes stars slightly better but also creates some artefacts. See the little rings slightly above and to the right of the "nose". What SN does much better is filling the residual areas with artificial noise which leads to a cleaner look. Nice!
3rd example: Rim of the Crescent. Here again SN (left) does remove stars slightly better than SX.
But, 4th example: Core of the Crescent. SN creates some crosshatch-like artefacts in the large area where the bright star was removed. SX just shows a blob.
Where SN is clearly better than SX is in the area of creating star masks for applications like deconvolution or star reduction. SX's star masks can be very messy with parts of the nebulosity or galaxy cores (M82 in this example) being included in the star mask. SN to the left produces much cleaner star masks.
Overall, I'm happy to have both tools in my arsenal. And I also want to very much thank the creator of StarNet for giving us this fantastic tool for free!
Cheers,
Reinhard
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Starnet V2 has upped it's game, that's for sure! It does at least as well as StarXTerminator and surpasses it on some images.
Regards,
Loran
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
|
SX did better in the core of the Rosette than SN and saved me from having to manually mask and patch. For me it's going to be case-by-case which gets used.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
D. Jung:
I started using StarXterminator, which works well in most cases, but in this particular case Starnet V2 kicks azz.
From top to bottom: original, starxterminator, starnet v2 (all in linear mode with standard STF)
Yeah, my push was mostly towards star dense fields. Glad this worked out well for you!
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
I have also done some comparisons between Starnet2 and StarXTerminator. The results are mixed but I would lean towards Starnet2 being slightly better for now.
1st example: Top of the Elephant's Trunk. Starnet2 (SN, left) removes more of the very dim red star residuals than StarXTerminator (SX)
2nd example: The Wizard's nose. SN (left) again removes stars slightly better but also creates some artefacts. See the little rings slightly above and to the right of the "nose". What SN does much better is filling the residual areas with artificial noise which leads to a cleaner look. Nice!
3rd example: Rim of the Crescent. Here again SN (left) does remove stars slightly better than SX.
But, 4th example: Core of the Crescent. SN creates some crosshatch-like artefacts in the large area where the bright star was removed. SX just shows a blob.
Where SN is clearly better than SX is in the area of creating star masks for applications like deconvolution or star reduction. SX's star masks can be very messy with parts of the nebulosity or galaxy cores (M82 in this example) being included in the star mask. SN to the left produces much cleaner star masks.
Overall, I'm happy to have both tools in my arsenal. And I also want to very much thank the creator of StarNet for giving us this fantastic tool for free!
Cheers,
Reinhard
Reinhard,
Thanks for all the kind words! I love that the discussion moved to a point where we have to pixel-peep to see flaws in star removal.
One thing I am trying to understand is how people perceive the results though. In your example #4 there is a single crosshatch-like 'artefact' as you call it, but for SX there is twice as large blob around the same star plus multitude of smaller blobs around smaller stars. You do not consider those to be artifacts?
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Reinhard,
Thanks for all the kind words! I love that the discussion moved to a point where we have to pixel-peep to see flaws in star removal.
One thing I am trying to understand is how people perceive the results though. In your example #4 there is a single crosshatch-like 'artefact' as you call it, but for SX there is twice as large blob around the same star plus multitude of smaller blobs around smaller stars. You do not consider those to be artifacts?
Hi Nikita, and thanks for replying here!
I totally agree that the "blobs" created by StarXTerminator are also artefacts. I just thought it interesting where this cross-hatch pattern came from, that's why I mentioned it in particular.
As to your question as to how the results are perceived: I really like what I called the "artificial noise" that Starnet uses to replace the stars. That makes the areas of star removal much more natural looking as compared to the very noticeable flat blobs that SX creates. Also Startet is better at removing even faint star residuals.
And for the serious pixel peepers there will always be the need for some final touch-ups. I have the feeling that the artefacts that Starnet introduces could be easier to remove that the ones by SX but I haven't tried that yet.
Thanks again for your great work!
Reinhard
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Reinhard,
Thanks for all the kind words! I love that the discussion moved to a point where we have to pixel-peep to see flaws in star removal.
One thing I am trying to understand is how people perceive the results though. In your example #4 there is a single crosshatch-like 'artefact' as you call it, but for SX there is twice as large blob around the same star plus multitude of smaller blobs around smaller stars. You do not consider those to be artifacts?
Hi Nikita, and thanks for replying here!
I totally agree that the "blobs" created by StarXTerminator are also artefacts. I just thought it interesting where this cross-hatch pattern came from, that's why I mentioned it in particular.
As to your question as to how the results are perceived: I really like what I called the "artificial noise" that Starnet uses to replace the stars. That makes the areas of star removal much more natural looking as compared to the very noticeable flat blobs that SX creates. Also Startet is better at removing even faint star residuals.
And for the serious pixel peepers there will always be the need for some final touch-ups. I have the feeling that the artefacts that Starnet introduces could be easier to remove that the ones by SX but I haven't tried that yet.
Thanks again for your great work!
Reinhard
Reinhard,
Thanks for clarifying! The pattern is from SN trying to generate noisy pattern, but there is no noise where huge stars are, so that's the result...
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Andy Wray:
Just starting to try out Starnet version 2. It seems to be quite a big improvement. I'd welcome other people's thoughts?
Here is a small version of an original Crescent capture:
Here's a version after star removal with version 1:
and here's the same after using version 2:
and this is what the version 2 star mask looks like:
I think it's particularly impressive when it comes to big stars and diffraction spikes which we as Newtonian users live with and for some of us actually like |
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Andy Wray:
Just starting to try out Starnet version 2. It seems to be quite a big improvement. I'd welcome other people's thoughts?
Here is a small version of an original Crescent capture:
Here's a version after star removal with version 1:
and here's the same after using version 2:
and this is what the version 2 star mask looks like:
I think it's particularly impressive when it comes to big stars and diffraction spikes which we as Newtonian users live with and for some of us actually like 
Can I ask how you get Starnet V2 to produce just the star field?
All I get is a starless image. Stars seem to have been simply deleted.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Christian Koll:
I have just one word for the results of StarnetV2: FLAWLESS!
This software is an amazing gift and has changed the way we process our images forever.
Chris
It’s a huge improvement but not “flawless” I’ve just been processing the California Nebula and it left MENKIB behind, whereas StarXTerminator removed it…
It’s really impressive but needs to be used in conjunction with all other tools.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
|
I find it a huge improvement over StarNet1, I did purchase StarXTerminator but find StarNet2 is pretty awesome too.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Hello to all,
Here is my contribution to the analysis of the results between StarNet (v1 and v2) and StarXTerminator (v6 and v7).
https://millenniumphoton.com/starnet-vs-starxterminator-le-match/
I hope it can be useful for you...
The article can be automatically translated into English by clicking on the flag directly on the page...
JB
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Die Launische Diva:
There are some disclaimers in the SN v2 announcement in Cloudiynights that make me think that SN is going to become a pay product.
I don't think this is the case, even if I believe the developer has every right to do so. I suspect that something more sad has happened: Someone else used Starnet's output data in order to train his/her own software and probably he/she is making a profit out of it. It is sad that others exploit his generosity. Whatever Nikita decides, he needs our support.
Despite the bug on linear images, Starnet v2 is a great improvement. Maybe the temporary stretch applied is not invertible (in the mathematical sense)?
In my old PI installation, Starnet v2 can happily coexist with the older version, and surprisingly, the older Starnet is compatible with the newer tensorflow.dll file.
Thank you Nikita for this truly innovative tool!
For the inquiring reader, this demonstrates how a malicious user may benefit from someone else's deep learning model. Figure 1 of the paper shows two potential infringers. The second (right) infringer has access only on the input/output of the model, an not on its internals.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Really hope there will be a command line version for mac. Don’t want to buy pixinsight just for that tool 😅
I found a link to sourceforge page for MAC/linux commandline tool
https://sourceforge.net/projects/starnet/files/v2.0/
but...
"This project's file releases are suspended because this project has not made source code available under an Open Source license, as required by SourceForge's Terms of Use.
"
 |
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Really hope there will be a command line version for mac. Don’t want to buy pixinsight just for that tool 😅
https://starnetastro.com/ |
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Die Launische Diva:
There are some disclaimers in the SN v2 announcement in Cloudiynights that make me think that SN is going to become a pay product.
I don't think this is the case, even if I believe the developer has every right to do so. I suspect that something more sad has happened: Someone else used Starnet's output data in order to train his/her own software and probably he/she is making a profit out of it. It is sad that others exploit his generosity. Whatever Nikita decides, he needs our support.
Despite the bug on linear images, Starnet v2 is a great improvement. Maybe the temporary stretch applied is not invertible (in the mathematical sense)?
In my old PI installation, Starnet v2 can happily coexist with the older version, and surprisingly, the older Starnet is compatible with the newer tensorflow.dll file.
Thank you Nikita for this truly innovative tool!
Starnet V2 is a fantastic tool - thank you so much Nikita for all your work on this and for making it freely available to us.
On the linear issue - people may wish to note that the GeneralisedHyperbolicStretch tool (see here) has a checkbox allowing a precise mathematical inversion of the GHS stretch making it easy to stretch-destar-unstretch. I have used this to great effect with Starnet.
|
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
Before StarNet v.2, I used the original StarNet algorithm to produce a starmask and then, via a set of processing steps, let Photoshop replace the stars by using the starmask and so-called content aware fill, which worked excellent as shown bottom left below.
With StarNet 2, the results are comparable to the content-aware filling. Sometimes a little better, and sometimes a little worse. Taking conveniance into account, I definetely go for StarNet 2.
For those interested in the Photoshop procedure, it is described here:
https://www.astrobin.com/forum/c/astrophotography/deep-sky-processing-techniques/a-helpful-tip-plastic-surgery-for-improved-star-free-images/
1. (top left): Original detail from the Wizard Nebula
2. (top right): Original StarNet++
3. (bottom left): Photoshop content aware method
4. (bottom right): StarNet++ version 2
Martin
 |
You cannot like this item. Reason: "ANONYMOUS".
You cannot remove your like from this item.
Editing a post is only allowed within 24 hours after creating it.
You cannot Like this post because the topic is closed.
0.90
0.00
3.82
7.84
2.41
0.00
3.53
4.44
12.54
11.14
23.42
3.21
