GraXpert with AI

[Jockinireland]

I watched a video by Cuiv about the new AI Driven GraXpert. So thought I'd give it a go.

Its easy to download, install and use. I trialled it on several different images of single galaxies to fields almost completely full of dust. To some with really difficult gradients.

What I found was that on something plain and simple, there was not much to be gained. But on an image with lots of dust and little real background (like my recent cepheus dust image) there was a benefit. When I do DBE on that type of image, I will use a reference image on AB to help me place sample points. This can be very longwinded and difficult. Importing the image to GraXpert is a simple matter and then it's essentially a one click process

On this type of image I found GraXpert did an excellent job for much less work. There were slight differences between the manual DBE output and GraXpert but I could not say which was more "correct" and really i was not to worried about it. Both looked perfectly fine to me.

On really complex gradients I also found that GX produced a very similar, possibly slightly better, image than DBE and again with minimal work by me.

So , GX is easy to use, simplifies the process significantly and gives output at least as satisfactory as DBE for minimal user input. I will certainly keep it and use it for anything other than simple gradients where sample points are easy to place.

Cuivs video is here

[Petrol]

Thanks for posting this. I tried it when I saw Cuivs vid but to my newbie eye, couldn't see any improvement on my images. That said, I have a simple approach to DBE that VisibleDark posted

Correct me if I'm wrong but if you have a fairly flat field, then the gradients are not that difficult to deal with. I'm really hoping to learn something from your post Jockinireland.
Again, thanks.
Pete

[Juno16]

Excellent review David!
I also have had good success with DBE, but like you, I tried out GraXpert on both simple and complex gradients and found it to be very comparable to DBE, only a heck of a lot faster!
I also like how GraXpert can work with and save files in Pixinsight’s proprietary xisf file format (amongst others).
Great stuff to add to the toolbox at an unbeatable price (free)!
Is AI stretching next???

[Jockinireland]

Thanks for posting this. I tried it when I saw Cuivs vid but to my newbie eye, couldn't see any improvement on my images. That said, I have a simple approach to DBE that VisibleDark posted

Correct me if I'm wrong but if you have a fairly flat field, then the gradients are not that difficult to deal with. I'm really hoping to learn something from your post Jockinireland.
Again, thanks.
Pete

Hi Pete, yes I used that process by Visibledark for a while but in truth I found that it was a bit hit and miss. Particularly on images with limited true background or complex gradients.

My thinking right now, and I'll need to work with it for a while to really be sure, is that Graxpert AI is not going to produce a "much better" result than DBE. The results are pretty similar. The difference is that, for certain types of images or complex gradients, to get a good result with DBE takes me time and significant effort - working out where to place the samples (using a reference image if need be), trialing various settings, working out the gradient symmetries etc. Whereas Graxpert AI produces an equivalent result with essentially one click. I found that Visibledarks simple method did not really cut the mustard on such difficult images.

And yes, on simple images, with a simple gradient, DBE, and even Visibledarks method is fine and I'll continue to use DBE for those. Throwing down a few sample points and hitting go is not any more work than exporting the image then importing it back after Graxpert. (And dont forget that on really simple gradients ABE can sometimes give a pretty good result straight off - I use it sometimes just to see) But on anything thats going to take more work than that then I'll probably use Graxpert. In fact I'll probably do both for a while to really asses Graxpert.

[Jockinireland]

Excellent review David!
I also have had good success with DBE, but like you, I tried out GraXpert on both simple and complex gradients and found it to be very comparable to DBE, only a heck of a lot faster!
I also like how GraXpert can work with and save files in Pixinsight’s proprietary xisf file format (amongst others).
Great stuff to add to the toolbox at an unbeatable price (free)!
Is AI stretching next???

Thanks Jim

"Is AI stretching next"? I am a biologist so I really dont understand how AI/neural networks actually work so really I dont know if stretching would be something that AI could do well. My gut instinct would be that things like gradient removal and deconvolution are things that maths (or just "math" as you Americans wrongly call it ) can handle. But stretching is something that is much more driven by personal taste and judgement. So I'm not really sure how well it would work, for me at least.

But I think that if it can be done there will be some clever chap training his AI (StretXpert??) right now. I think that the amount of new tools, AI or otherwise, that our hobby is being given right now is just astonishing.

Take care,

David

[Juno16]

Hi David,

I was kind of joking about stretching, but also slightly serious.

Yes, stretching is absolutely a personal preference. As we all know, many different outcomes (infinite) are possible during the stretching process. Many very nice and very different variants are possible.

For me, I would say that stretching is my most challenging processing step and the one that I work at the longest (and probably still have the most to learn).

I have no idea about the maths around AI, but Photoshop AI (content aware fill I believe) and ChatGPT, but they both will give alternative outputs or responses. Conditions can be applied to every ChatGPT request and the AI responds accordingly.
You have probably also tried ChatGPT like I have, and I am amazed at the responses that I have gotten when presenting an astrophotography processing question.

I think that StretXpert coming and maybe not too far off. But who knows.

We have spent years developing skills to process our data. Possibly soon, all it will take is a few clicks!

[Petrol]

Thanks for posting this. I tried it when I saw Cuivs vid but to my newbie eye, couldn't see any improvement on my images. That said, I have a simple approach to DBE that VisibleDark posted

Correct me if I'm wrong but if you have a fairly flat field, then the gradients are not that difficult to deal with. I'm really hoping to learn something from your post Jockinireland.
Again, thanks.
Pete

Hi Pete, yes I used that process by Visibledark for a while but in truth I found that it was a bit hit and miss. Particularly on images with limited true background or complex gradients.

My thinking right now, and I'll need to work with it for a while to really be sure, is that Graxpert AI is not going to produce a "much better" result than DBE. The results are pretty similar. The difference is that, for certain types of images or complex gradients, to get a good result with DBE takes me time and significant effort - working out where to place the samples (using a reference image if need be), trialing various settings, working out the gradient symmetries etc. Whereas Graxpert AI produces an equivalent result with essentially one click. I found that Visibledarks simple method did not really cut the mustard on such difficult images.

And yes, on simple images, with a simple gradient, DBE, and even Visibledarks method is fine and I'll continue to use DBE for those. Throwing down a few sample points and hitting go is not any more work than exporting the image then importing it back after Graxpert. (And dont forget that on really simple gradients ABE can sometimes give a pretty good result straight off - I use it sometimes just to see) But on anything thats going to take more work than that then I'll probably use Graxpert. In fact I'll probably do both for a while to really asses Graxpert.

Thanks for your detailed explination David. Some of how this processing works is a little over my head. That said, I'll take all the help I can and will now use Graxpert AI instead of DBE. On a sidenote, I bought normalise scale gradient which is another process I have to go through though I'm not entirely sure what it actually does other than take a long time! Recently I've put a lot of effort trying to optimise the imaging trains on my rigs in the hope that better data needs less processing.
Once again, thank you for sharing your expertise and making things quicker and easier.

Pete

[Jockinireland]

Thanks for posting this. I tried it when I saw Cuivs vid but to my newbie eye, couldn't see any improvement on my images. That said, I have a simple approach to DBE that VisibleDark posted

Correct me if I'm wrong but if you have a fairly flat field, then the gradients are not that difficult to deal with. I'm really hoping to learn something from your post Jockinireland.
Again, thanks.
Pete

Hi Pete, yes I used that process by Visibledark for a while but in truth I found that it was a bit hit and miss. Particularly on images with limited true background or complex gradients.

My thinking right now, and I'll need to work with it for a while to really be sure, is that Graxpert AI is not going to produce a "much better" result than DBE. The results are pretty similar. The difference is that, for certain types of images or complex gradients, to get a good result with DBE takes me time and significant effort - working out where to place the samples (using a reference image if need be), trialing various settings, working out the gradient symmetries etc. Whereas Graxpert AI produces an equivalent result with essentially one click. I found that Visibledarks simple method did not really cut the mustard on such difficult images.

And yes, on simple images, with a simple gradient, DBE, and even Visibledarks method is fine and I'll continue to use DBE for those. Throwing down a few sample points and hitting go is not any more work than exporting the image then importing it back after Graxpert. (And dont forget that on really simple gradients ABE can sometimes give a pretty good result straight off - I use it sometimes just to see) But on anything thats going to take more work than that then I'll probably use Graxpert. In fact I'll probably do both for a while to really asses Graxpert.

Thanks for your detailed explination David. Some of how this processing works is a little over my head. That said, I'll take all the help I can and will now use Graxpert AI instead of DBE. On a sidenote, I bought normalise scale gradient which is another process I have to go through though I'm not entirely sure what it actually does other than take a long time! Recently I've put a lot of effort trying to optimise the imaging trains on my rigs in the hope that better data needs less processing.
Once again, thank you for sharing your expertise and making things quicker and easier.

Pete

Hi Pete,

So, I am far from expert but my understanding is that what NSG does is to minimise the effects of gradient (light pollution, passing clouds etc) that change during an imaging run. So it basically looks at all of the individual light frames, assesses the gradient in each and then normalises them against a reference frame. This is done before integration and, as I understand it, leads to a better, less gradient impacted final stacked frame.

I used NSG a few times when it first came out and found it did make a noticeable difference - a stack with NSG did look flatter and cleaner than a stack of the same lights without NSG.

However, fairly soon after that, Pixinsight made significant changes to their LocalNormalisation process, which does pretty much the same thing. They also integrated it into the WBPP process which means that you dont need to do it as a seperate process.

I did some quick comparisons and found that, using my uncalibrated V1 Eyeball, I could not see any difference between the two processes.I expect that some of the Pixinsight Maths geniuses could do some analysis to show a mathematical difference but, honestly, that's beyond my level of interest. If it looks good to my eye, thats good enough for me.

So, I no longer use NSG but include LocalNormalisation in my standard WBPP process, irrespective of whether I have variable gradients or not. Now, be aware that it does increase the processing time for WBPP quite significantly, but the way I work is that I load up WBPP hit start and then go off and do something else for as many hours as it takes. So its not an issue for me.

Your man Shawn did a video in which he assessed the two and basically reached the same conclusion as I did.

[Juno16]

@Petrol
@Jockinireland

I did a very limited comparison (one dataset) and like David, found no visible difference between NSG and LN with my data.

I did however, compare several datasets with and without LN and there was a significant difference. The LN masterlights were much "smoother."
LN is included in my standard WBPP workflow and the only cost is a few extra minutes of time.

Just to note: LocalNormalization is found in WBPP under the Lights tab. Make sure the box is checked.
Also to note: Several PI WBPP versions back, LN was set by default to choose a "Single Best Frame" for the reference frame. This default setting changed to "Integration of best frames" in later WBPP revisions. So, just take a quick peek to insure that "Integration of best frames" is selected at the Reference frame generation.

Take care.

[Petrol]

@Jockinireland
@Juno16
First of all apologies for going off topic somewhat. Thanks for your detailed explination David, I understand now.
You guys have saved me an enormous amount of processing time and have taken the hassle out of it all.
I'll ditch NSG
Stack everything in WBPP using LocalNormalisation
Then run GraXpert
Easy!

Again thanks
Pete