PI GradientCorrection process in 1.8.9.2 Build 1601

THAT'S SO COOL!!!!!!

It works much better and It's so much easier than any software I've ever seen before.

Hit and miss as well. It's an obvious win for some of my images, but not as efficient as DBE with others. I've noticed that GradientCorrection (GC) can't handle edges that are not trimmed. I'm also running into an issue where nebulosity is over corrected, but when I reduce the settings to correct for the nebulosity, dark structures begin disappearing.

I‘ve had some problems with uncorrectable gradients in a few of my images (light leak, now fixed) - and GC cleaned them up extremely well with standard settings. ABE, DBE and Graxpert all failed, even with much fiddling. Great tool.

I tried Gradient Correction (GC) yesterday with terrific results on some older data. I like having one more option in addition to AI Graxpert, DBE and ABE. 

It’s important to remember that GC is meant for use on images with additive gradients (where using subtraction with DBE is appropriate). It appears to be less effective on multiplicative gradients (such as vignetting where division would be used in DBE).

Tried it as well and like it a lot. For me, Scale and Smoothness were the most important parameters, I left the others at default. Automatic convergence I ticked as well. And with bright and prominent objects in the image (I tested M 81 as well), Structure Protection is necessary. Yet I noticed sometimes a slight glow at the image edges (not all edges, and not in all cases), although the images were trimmed before GC. Need more experience here…

I think for "easy" images with an isolated object and lots of true sky background surrounding it, DBE is done quickly and reliably. But with more complicated cases like nebulosity all over the image or lots of dark clouds, GradientCorrection seems to me to be a better alternative.

Although tweaking and trial-and-error is necessary. I think we're all waiting for MARS! 

And yes, like Richard wrote above, showing the resulting background model would be great, as well as the live view.

Colleagues,


I will also test this feature … looks like
a fantastic feature… I will come back with some examples…

CS
Johannes

I too have been using and loving GraXpert AI for quite some time.  Yesterday I gave GC a whirl and was completely thrilled with the results. My environment has typically required separate treatment of gradients on each filter. I picked a terrible example - RGB low directly over the top of Dallas TX and GC in two tries just nailed it for the RGB made for non-corrected RGB masters. GC is really easy to use too. Great job PI team!!

Just install the new PI 1-8-9-2 build on top of your existing PI installation meaning do not uninstall first..  GradientCorrection will be there in processes!

Dale Penkala:
I got the notice yesterday about the GC tool and I’m excited to try it. I have been using DBE and GraXpert and kinda was liking GX but this looks and sounds very promising!
I finished up 2 images last night, 1 a planetary where I used all NB data and another M97 $ M108 with just UV/IR Cut filter and plan to give it a test run!

Quick question, has anyone had any problems installing it like some of the other updates in PI?

Dale

Hi Dale,
Just as Ron said. I installed 1601 right over the existing copy of PI (build 1597?). 
Checked for updates after installation and all processes were there. 

If you use cuda, make sure that you save the tensorflow.dll from your present (old) installation (in the bin folder) and replace the new tensorflow.dll with the old version. 

Jim

Jim Raskett:

Dale Penkala:
I got the notice yesterday about the GC tool and I’m excited to try it. I have been using DBE and GraXpert and kinda was liking GX but this looks and sounds very promising!
I finished up 2 images last night, 1 a planetary where I used all NB data and another M97 $ M108 with just UV/IR Cut filter and plan to give it a test run!

Quick question, has anyone had any problems installing it like some of the other updates in PI?

Dale

Hi Dale,
Just as Ron said. I installed 1601 right over the existing copy of PI (build 1597?). 
Checked for updates after installation and all processes were there. 

If you use cuda, make sure that you save the tensorflow.dll from your present (old) installation (in the bin folder) and replace the new tensorflow.dll with the old version. 

Jim

Thanks @Ron Crouch & @Jim Raskett for that info!

Dale

Just had a play with it on a challenging image, and to be really practical I agree it does ideally need a preview option. There are too many sliders to change /play with for a process without a live/semi-live preview. I tried it (default settings apart from automatic convergence ticked) on a really difficult/weird circular radial/ring-like gradient I have on one integration done with a sloppy altaz mount without a camera rotator using several hours of very short 10 sec exposures (so that each individual sub had no field rotation problems), and it didn't do anywhere near as well as the latest Graxpert AI version on that tricky gradient. It did do slightly better than ABE, but only slightly. I am using it on a very difficult gradient that might even not be an additive gradient, but I have  no idea about that. (How do I tell if a gradient is or isn't additive? Is there some easy way to tell/know, or an analysis process/script that can tell me that?)

Next I'll try playing with the sliders to see that helps it deal with that gradient, but without a live preview that might be a painful iterative process, when i know Graxpert handles it well enough with its default AI settings.

An edit/update on the above. Just by lowering the scale slider so the process does more work on smaller scale gradients, which this ring gradient with its narrowish rings is, the result is much improved on the defaults without losing much of the real image structures ( my first try at a much lower scale setting  [1.9]  was too low and removed some real structures as well as the gradient, but scale 2.7 worked nicely). I'd say with that one change from the defaults (plus automatic convergence on) it now gives me a better result than Graxpert's AI effort. I'm impressed after that simple single slider change.

Cheers,
Chris