That said, now that I've tried this out - I'm not convinced by the nay-sayers that it's "generating details that arn't there" , but simply refining what is, and in a very advanced manner.
Hmmm?...?...yeah, I think I agree at least from the comparison you have posted ............but !..no worms!!! ok??
Interesting debate kicking off here Andy on the ethics of Image processing. Personally, I fall into the team that says that anything that is not entirely based on the captured data and thereby mathematically reversible is straying too far. My understanding from reading this thread is that the AI suite is using external data sources to build its algorithms with a composite image resulting.
This is not to say that these techniques should not be used and it becomes a personal choice for the image processor. My view is that it is acceptable as long as one has full disclosure which you have clearly done.
Rodney, I'd buy you an Internet-beer for that. Really well said.
Topaz AI (and Starnet++ for that matter, which creates similar artifacts) is to be avoided if you would like to do astrophotography. The same obviously goes for selective processing (e.g. creating a hand-drawn mask to manipulate part of the image). If you're doing photography, then adding stuff that wasn't there in the first place is not ok, particularly if it's not an "accident" (e.g. accidental ringing/Gibbs phenomenon).
If you do astroart, then anything goes obviously, but I think this particular sub-section of the forum is mostly populated by photographers, not artists. As such, we assume that what you are showing us is real and not the result of arbitrary manipulation of the image, or the result of a hallucination of a neural net.
It's not just Topaz AI either; many "standard" noise reduction routines are specific to terrestrial scenes, and built to reconstruct a scene. They are predisposed to reconstruct edges and geometrical shapes, whereas these are virtually non-existent in outer space. Such reconstructions fail miserably in - particularly - DSO datasets, ending up "reconstructing" detail where non exists. The same goes for debayering algorithms - the most "advanced" ones (for example those based on AHD) yield the worst SNR, as they try to "reconstruct" edges and detail based on random noise. Stacking that - now random detail - yields worse outcomes. This is one of the reasons why something "advanced" like PixInsight only offers something simple like VNG - precisely because it doesn't do anything fancy.
The best (and most widely accepted in astronomy) noise reduction algorithms for astrophotography only remove energy from your data and do not introduce it.
Good on you to drop in Ivo so we can hear it directly from the horse's mouth.
Quote:
Originally Posted by irwjager
As such, we assume that what you are showing us is real and not the result of arbitrary manipulation of the image, or the result of a hallucination of a neural net.
Well that illustrate my initial concerns in a big way. If a neural network can interpolate a jelly fish wideshot into a dog's painting I don't even want to think what it's doing to an astro image. Maybe the lower more subtle setting gives you cats?
Quote:
Originally Posted by irwjager
The same goes for debayering algorithms - the most "advanced" ones (for example those based on AHD) yield the worst SNR, as they try to "reconstruct" edges and detail based on random noise. Stacking that - now random detail - yields worse outcomes. This is one of the reasons why something "advanced" like PixInsight only offers something simple like VNG - precisely because it doesn't do anything fancy.
What's your take on bayer drizzle and drizzle integration. That's ok? Been doing it a lot of that with the FSQ and as far as I can tell it doesn't manufacture details but I thought I'd check with you anyway.
If you do astroart, then anything goes obviously, but I think this particular sub-section of the forum is mostly populated by photographers, not artists.
That argument seems a little simplistic Ivo - can we not be/do both ie: create art from scientifically valid astro images?
I'm hugely in favour of pushing established boundaries, all the way over the cliff if need be to create interesting and appealing images, and I'm not alone in that quest!
The debate about what's real or not is hugely subjective, as none of us have a warp capable starship to prove any of it, so why wouldn't you delight in something that helps ground based imaging punch through atmospheric distortion?
This is simply the next tool in refinement of post processes. Did everyone scream it's not real when digital imaging arrived? or CMOS, or wavelet sharpening, or starmasks, etc...
Semantics aside, some of these arguments sound like the church telling Copernicus that the Sun revolved around the earth and anything else is heresy!
So lighten up peeps, it's relatively new but AI is here to stay, and it's learning - just hope no-one founds Skynet anytime soon and we'll all be fine
That argument seems a little simplistic Ivo - can we not be/do both ie: create art from scientifically valid astro images?
I'm hugely in favour of pushing established boundaries, all the way over the cliff if need be to create interesting and appealing images, and I'm not alone in that quest!
The debate about what's real or not is hugely subjective, as none of us have a warp capable starship to prove any of it, so why wouldn't you delight in something that helps ground based imaging punch through atmospheric distortion?
This is simply the next tool in refinement of post processes. Did everyone scream it's not real when digital imaging arrived? or CMOS, or wavelet sharpening, or starmasks, etc...
Semantics aside, some of these arguments sound like the church telling Copernicus that the Sun revolved around the earth and anything else is heresy!
So lighten up peeps, it's relatively new but AI is here to stay, and it's learning - just hope no-one founds Skynet anytime soon and we'll all be fine
Andy, mate... what are you on? What you're describing is astro phantasy not astro photography. Maybe IIS needs an additional section called astro art then that would work.
What's your take on bayer drizzle and drizzle integration.
There's nothing about drizzling that makes up detail, so I'd say go your hardest
(And if it's good enough for the fine folks at NASA and the HST, then it's good enough for me!)
Quote:
Originally Posted by Andy01
That argument seems a little simplistic Ivo - can we not be/do both ie: create art from scientifically valid astro images?
I totally agree of course that every image is an interpretation of barely visible photon counts. That's already 50% art right there.
Quote:
The debate about what's real or not is hugely subjective, as none of us have a warp capable starship to prove any of it,
We don't need to! Us mere mortals with our puny scopes can readily compare our images to - for example - majestic Hubble closeups to see if what we captured is real if it is i doubt.
Quote:
so why wouldn't you delight in something that helps ground based imaging punch through atmospheric distortion?
I'd be all for it, but that's unfortunately not at all what we're looking at here.
(and we have deconvolution - an actual working tool - for that very purpose )
Quote:
This is simply the next tool in refinement of post processes. Did everyone scream it's not real when digital imaging arrived? or CMOS, or wavelet sharpening, or starmasks, etc...
Even if you are not across how neural nets (and training thereof) work fundamentally, we still don't have to guess whether something is real. We can hop on to the Hubble Heritage website, look up our object and download a dataset (great fun in itself!). Or we can check some of the work done by our peers on AstroBin etc.
Quote:
AI is here to stay, and it's learning - just hope no-one founds Skynet anytime soon and we'll all be fine
AI has been around for a long, long time (and what's considered AI keeps being pushed; remember when a chess computer was considered "AI"?). Like you, I'm also 100% convinced a practical application of deep learning for AP will come along some day, but this sort of simplistic data augmentation by detail hallucination is definitely not it.
(and let me be clear that I really, really want this to happen; I studied AI at the University of Amsterdam and I'm an AP algorithm nut )
No need to wait, Sky AI to the rescue, replace the whole sky in one go (includes atmospheric effects)
Marc, I usually respect your opinions, but Unless you have something meaningful to add to my OP, please stop wasting my time with this.
We all read your opinion & not everyone agreed with it. These repetitive niggles are getting boring now. Thanks
Wow. ANDY great stuff. As some of you know I've been out of the "game" for 8 years and only recently started imaging again. You all should be very proud of your images because my 8 year time warp perspective has blown me away by the quality of images. This image is such high quality
Well done Andy and good to read everyone's perspective on AI (which didn't exist 8 years ago)
I’m not across this AI stuff and so am unsure of what it consists of so can’t comment from a technical viewpoint however........
....my view is if the digital data that was captured by the imager is the only data used in the picture then that is still an original picture that the imager can call their own work. It’s no different from using a sharpen tool, decon or even assigning colours to a NB filter.
Having said that though, if data is being pulled from an external source and added to the imagers data then it’s a composite and should only be partially attributed to the imager.
I don’t think there a difference between ‘a small enhancement’ or ‘replaced 50% with other people’s work’. Skill, equipment, understanding and knowledge is what makes a great imager. Adding other people’s data to your own makes a great artist.
.....not pointing fingers andy, as I said, I don’t know how this AI stuff works.
Marc, I usually respect your opinions, but Unless you have something meaningful to add to my OP, please stop wasting my time with this.
We all read your opinion & not everyone agreed with it. These repetitive niggles are getting boring now. Thanks
Awwww... don't be a sour puss sensei. You asked for the debate to take place in your post.
I’m not across this AI stuff and so am unsure of what it consists of so can’t comment from a technical viewpoint however........
....my view is if the digital data that was captured by the imager is the only data used in the picture then that is still an original picture that the imager can call their own work. It’s no different from using a sharpen tool, decon or even assigning colours to a NB filter.
Having said that though, if data is being pulled from an external source and added to the imagers data then it’s a composite and should only be partially attributed to the imager.
I don’t think there a difference between ‘a small enhancement’ or ‘replaced 50% with other people’s work’. Skill, equipment, understanding and knowledge is what makes a great imager. Adding other people’s data to your own makes a great artist.
.....not pointing fingers andy, as I said, I don’t know how this AI stuff works.
That is not how the Topaz suite works though, it is not importing data from other sources and inserting it, it is extrapolating detail that is not visible from detail that is. The question becomes how much sharpening it can do before it is inventing detail, not revealing it.
People importing data from other sources is a whole other kettle of fish.
That is not how the Topaz suite works though, it is not importing data from other sources and inserting it, it is extrapolating detail that is not visible from detail that is. The question becomes how much sharpening it can do before it is inventing detail, not revealing it.
People importing data from other sources is a whole other kettle of fish.
As I said.... I’m not sure how it works so having had that explained (purely sharpening) then I don’t see an issue with it.
Ivo knows neural networks and deep learning. He does it for a living. He said AI for astro photography it's not quite ready yet but one day might be and he's looking forward to it.
Andy, I'm not voicing an opinion, just pulling your leg a little . You know what they say about "opinions". Everybody's got one.
I explained that I quantitatively showed that Denoise AI and Sharpen AI are manufacturing details even on the lowest settings.
You say that you don't think they do and you are happy to use them. I have absolutely zero problem with that. But you can't call it astro photography mate. It doesn't work this way.