10.26
#...
·
9
likes
|
|---|
|
Hi folks, this is a heads up that I recently presented on the Astro Imaging Channel. Among other subjects I discussed the use (and ethics) of Topaz AI software for Astrophotography to create dynamic images. Tune in anytime as I unpacked this can of worms ... https://www.youtube.com/watch?v=GjTWAM7O4zo cheers Andy |
1.51
#...
·
5
likes
|
|---|
| @Andy 01 Why did you single out Topaz? We are all using Photoshop or similar which allows unnatural manipulation of images. It is just up to each user how far he goes. Tomas |
|
10.26
#...
·
1
like
|
|---|
| @Tomas Chylek Photoshop has been around for a long time, and while it has had many evolutions and updates, as you can see demonstrated in my presentation - Topaz Ai is new and a game changer - especially the latest Denoise AI. Cheers Andy |
|
1.51
#...
·
4
likes
|
|---|
| @Andy 01 I am lately less enthusiastic with Topaz Denoise as a game changer. It was good for some of my images but turned unusable for others. Even with 0 sharpening it still introduces artifacts. But again, other functions in Photoshop or GIMP also introduce artifacts if you are not careful. |
18.84
#...
·
12
likes
|
|---|
|
Hello All, I think this question is not valid in my point of view. All images presented here (or in general) are pure art, which are an individual's interpretation and not scientific data at all. They are processed in many different tools and every step adds, removes or tampers the RAW data. Sometimes in a lossless way, sometimes in a lossy way, which is the vast majority. Even the way they are gathered is already impacting the "truth". Therefore you can only claim "this image is closer" to "reality", but even that is hard to quantize, because you would need the 100% for reference. Therefore I agree with @Tomas Chylek and consider this question as not legitimate question at all. If you want to discuss "the truth" you would have to present un-manuipulated, single RAW fits from different detectors, and even that can not claim to be the truth depending on the chips capabilities. I think it is absolutely necessary to accept that any presentation is an expression of personal taste of the creator. The idea to present the "reality" is a pure illusion. CS Rüdiger |
3.82
#...
·
5
likes
|
|---|
Tomas Chylek: You should try Low Light option, it is not mentioned in this video, but that looks to be the best option for AP. |
3.34
#... |
|---|
Ruediger: IMHO there are some shades of grey to this question. RAW data is some sort of (scientific) data, which can be transformed by specific algorithms in order to produce more appealing images. There are different levels of transforming the data from applying an algorithm to the whole of a picture to selective data manipulation by applying masks for example. Another level is introducing signal that has never been in data in the first place which are then artefacts. You can get there easily with Topaz. Moreover, to my knowledge Denoise AI is kind of an heuristic application which additionally performs differently on different hardware setups as I've heard. So reproducibility is an issue. Don't get me wrong, I bought a license of Denoise AI, I used it pretty boldly, introduced artefacts (unintentionally), accepted them and shared the result nonetheless, since it just looked better than the version not processed with Topaz. But the version before Topaz processing was less artificial for sure. It's not about talking about "truth" (that's not what science is about anyway, that's rather "reality" ), but being more or less far away from the original data. IMHO there is a difference between Starnet/Topaz denoise and a color preserving stretch, for example. People just should be aware of that. CS Stefan |
3.01
#...
·
1
like
|
|---|
Andy 01: Topaz AI is for sure Not a game changer. cs johny |
|
10.26
#... |
|---|
| @sky-watcher (johny) Hi Johny, I respect you’re entitled to your opinion, however given the brevity of your remarks, would you please back up your statement with some documented observations or facts? cheers Andy |
|
18.84
#...
·
2
likes
|
|---|
|
Hi @Stefan Böckler I agree to some extend, but I actually do not see fifty shades of grey (  sorry could not resist). Right at the beginning of pre-processing we remove hot and cold pixel, clusters or even bigger structures like clusters or trails. Or clip the data, or have other cosmetic filters. This is an irreversible tempering of data and here it already loses the title "truth" or "real".I agree, that it is a very subjective and individual question, which will be answered by everyone differently. I also think there is no real answer to it. From the scientific perspective it can be answered rather digital. The data is tampered and so it is art. Everyone has to decide to which extend he or she wants to alter the data. And everyone who is watching an image has to be aware that he or she is watching an interpretation. Nothing more, nothing less. Everything else is cheating yourself.  |
|
3.01
#...
·
1
like
|
|---|
Andy 01: I don't think that I need to present any documented observations or facts. I tried out Topaz DeNoise AI, Topaz Sharpen AI, Topaz Gigapixel AI, Topaz Adjust AI and I experienced what I expected. They are good to use in daylight photography, and in some cases also in astrophotography. But that's all, Topaz AI is for sure Not a game changer cs johny |
0.00
#...
·
1
like
|
|---|
|
Hello Folks I think Denoise AI is good for beginners because it's easy to sharpen and remove noise especially with the low light mode. Unfortunately, stars are nibbled at very quickly there or very fine details are quickly destroyed. On 3 different PCs I got 3 different results with the same settings for the same picture. It is very strange. Sorry for the translation I use google translate  greetings Danny |
|
3.82
#...
·
4
likes
|
|---|
Topaz AI is for sure Not a game changer The problem with Topaz, that it can totally destroy the original details in the way that the result is still an eye candy picture (at least at the first sight). (There were some examples also here on AB) Too heavy noise reduction or too strong sharpening in PS or in PI is visible immediately, but since in Topaz the 2 effects are happening same time driven by own new logic, there will be nice looking artefacts with lot of "details" and very low noise. I do not know what exactly we call game changer, but this is definitely a new challenge in AP. However it is also possible to use in a moderate way, low noise option, recover original details can work quite correct, but for a beginner it is not very clear immediately, therefore I agree, that it has sense to discuss the usage here. |
3.82
#...
·
7
likes
|
|---|
|
When I use Denoise or Sharpen, I almost never apply them to the entire image. I have Topaz set up as a plugin in Photoshop and only apply it to a masked area. I also keep the sliders on the low end of the scale. A little goes a long ways with these tools. They can do some wonderful things - or not - depending on the image and how you use it. I really don't see that using Topaz is any different than, say, using any of the processes in PI. Topaz is just another wrench in the toolkit. CS, Loran |
|
1.51
#... |
|---|
|
here is a sample of my clearly over-done Topaz Denoise: https://www.astrobin.com/bgcpcq/0/?nc=user and here more resonably Topaz denoised: https://www.astrobin.com/bgcpcq/K/?nc=user which one is more pleasing to look at? Tomas |
0.00
#...
·
2
likes
|
|---|
Danny: *** Not surprising you got 3 different results. Since you were running 3 different instances of the AI. AI does not follow the same pre determined “if this then that” kinda logic that most software follows. I think this is what is a game changer about Topaz Denoise, the use of AI in order to remove noise from an image is an amazing step towards the future.*** |
9.90
#...
·
6
likes
|
|---|
Andy 01: Hi Andy, Really nice talk, I finally had the time to watch the video. You raised all the good questions, however I felt that you were convinced in the end by the AI techniques proposed by Topaz. I think there are some good reasons for being skeptical about the AI application to astrophotographic details and data. I am not saying that we cannot use Topaz AI for denoising or sharpening, but that it is much more dangerous to lose control over what we are doing and depart from reality. The dangerous aspect of this tool is that it creates images that look like reality since it has learned from the database how real, sharp, noise free images look like. So yes, the finite product looks real. The problem is that it is not necessarily real (it begs the question to say that that's what the actual detailed picture should look like). Where there is no information (by definition, blurred images or noisy images have less information) I don't understand how AI (or HI) can uncover the information. It can make a best guess based on previous data and it can produce something that could be real. If there are 3 strands of hair in reality, but the image does not contain that information, it just has an average thick strand of uniform pixels, no one can infer from that data that there are really 3 strands of hair. Topaz can produce a realistic image with 3 strands of hair, but that does not mean it is real. The case becomes even more complicated when we're talking about astrophotography. As you said, this software or AI was trained for faces or dogs or I don't know what particular daylight photo theme. It has a real reference that can for example teach it what particular human anatomical traits are more probable given some details, so it can infer it. You cannot apply what that AI has learned from a portrait database to galaxies and nebular filaments or cloud densities. The software will create something, but if it isn't actually there, I don't see the point of doing it. I have seen Topaz images where clearly the resolution and details are artificial. You cannot produce real details using AI when it was impossible for the telescope to have produced them in the first place. If a 100mm refractor + AI gives you the same information as a 20" CDK, what's the point of searching the actual photons with diameter, seeing, good guiding and perfect focus? That's just an illusion. Another aspect of AI and deep learning is that it is a black box. We don't know exactly how Alpha Go is thinking and beating the World Champion. Same stands for these tools. Starnet is very different. You have the original, the real initial image. You can check it for artifacts and errors. Not the case for this type of AI. We have Hubble as you have shown, but sharp, Hubble like details are not necessarily real (remember the M101 debate). Voilà, some worms came out : ). Best regards, Bogdan |
19.68
#...
·
18
likes
|
|---|
|
I have been following this discussion for a few days, and I would like to add my perspective as a former research scientist. By way of background, my doctoral work was in quantitative ecology. My research was published in refereed scientific journals, and included observational science (re-describing marine species), experimental science (both field and laboratory), and theoretical science (synthesizing quantitative models based on observational and experimental data and analysis). From my perspective, very little of what happens here at Astrobin is science. There are notable exception, such as the discovery and cataloging of new objects. Beyond that, most of us are making pictures of objects in space. To qualify as science work must be repeatable by someone else. In other words, by precisely repeating all the materials, protocols, procedures, and timings of your work, the same result can be accomplished by someone else (in experimental science, statistical analysis is used to determine whether outcomes are likely the same). We all collect photographic data (which is impacted by the native variability of different equipment, procedures, sky conditions, etc.), and then use these data to create an image. Once we bring these data into Photoshop, Pixinsight, Topaz, or nearly any other imaging software, subjectivity takes over and repeatability is not possible. In other words, we are making an image that is personal expression of our taste, sensibilities, or intention. It is for that reason that I do not perceive any ethical conflict in image created with any imaging software, unless of course, the imager is maintaining that the resulting image is representative of some natural history phenomenon or other scientifically testable characteristic. For Astrobin I make “pretty space picture,” and I believe that is the case for most people posting their work at Astrobin. Just to underscore that point, we have a Fine Art Astrophotography group. Manipulating your data in anyway that pleases you is supported by the group because in the end, the goal is to make images that are beautiful, dramatic, and inspirational. CS, Gary |
9.15
#...
·
3
likes
|
|---|
|
Hi Andy, I watched your presentation over the weekend and I really enjoyed it. I have expressed my thoughts on AI based tools in another thread and will not repeat them here. In summary, these are image processing tools with their benefits and shortcomings like every other tool in the box. You can wreck an image by overusing PI deconvolution and this is no different. I use machine learning as a research scientist every day and you learn to understand when you can trust the output and when it looks suspicious. Also, people should not confuse training of a neural network with inference with a trained neural network, which is what we do when we use Starnet or Topaz DenoiseAI. Inference, is perfectly deterministic and will always give the same answer given the same input data (right or wrong). I liked the examples you provided but even more I enjoyed your discussion on dynamic images and aesthetics. Many of us have the technical aspects of image processing pretty much down pat. What I learn from people like you, Alan Pham, Gary Lopez to name a few is how to modulate the aesthetic impact of an image, why and image evokes an emotional response. All the round tight stars in the world will not do that. So thank you for really emphasizing this topic in your talk and presenting your point of view. Cheers, Luca |
|
10.26
#...
·
4
likes
|
|---|
Bogdan Borz: Luca Marinelli: Hi Luca & Bogdan, thanks for your feedback, glad you got something out of the presentation. Thanks too everyone else for your input, it's nice to have a solid (robust?) discussion about something we clearly all feel strongly about! Cheers all. Andy |
11.07
#...
·
3
likes
|
|---|
|
Thanks Andy good presentation. If there is a take-away, at least for me, it's be careful and conservative with Topaz. That has certainly been my experience to this point with Topaz AI. I still like Denoise 5.1 too. I think Topaz is a valid tool for what we do. I would suggest these guidelines based on my experience: 1) As you pointed out, you normally want to protect stars in one way or the other. 2) I have found that you need to be really careful with the dark details in bright areas like galaxy and nebula cores. It really wants to make up details in those areas. If it seems to be making up details there, consider masking out the core in whole or part, try the low light version, or use the retain original details slider. 3) I would also say that I've never seen using the sharpen slider add anything of value - meaning I've either left it at the default or turned it off. 4) Always check the low light setting to see whether it does a better job of reducing noise without creating false detail. Cheers, Kevin |
6.08
#...
·
2
likes
|
|---|
Gary Lopez: Well said, and too the point. I do this to make pretty images of space. It's a hobby not science for the most part. Photoshop, PI, Topaz are tools we use, sometime overkill but in most part it's well done. This is suppoused to be a fun, and as far as introducing artifacts due to over-processing, I like looking at images as a hole, not inlarged 500%. |
|
6.08
#...
·
3
likes
|
|---|
|
Andy, watched your "Can of worms" it was great. From the above post. I'd also like to add that I use most all the AI programs from Topaz and have used Topaz labs for years and think they all work quite well. When Topaz first came out with the AI versions, they started with an "AI clear" that I think works great but is no longer available. I think the information AI is working with is there, you just don't see it untell it runs the process. I think of it as doing what you would do doing a deconvolution and then bringing it out with a layer mask, but Topaz dose it for you. |
2.15
#...
·
4
likes
|
|---|
Gary Lopez: Gary Lopez: Apologies in advance for sounding like a lawyer, but I don't know any other languages, other than Texan, so I hope that all y'all will forgive both deficiencies. Mr. Lopez' view is consistent with how the law defines and protects intellectual property. For example, a scientific method or apparatus can be protected under the law in most jurisdictions through a patent or as a trade secret. On the other hand, a scientific method or apparatus cannot be protected as a copyright because a copyright exclusively protects artistic works. Do you see where I am going here? Photographs are considered artistic works and may be protected as copyrights (and possibly trademark law depending on how the artistic work is used), but photographs are not protected as patents or trade secrets. The artistic works presented on Astrobin and other social media platforms are manipulated using many different techniques with different parameters, some of which include camera filters, stacking parameters, and post processing software. It is well settled law that a photograph is an artistic work, in the same category as a poem, song lyric, instrumental composition, and literary work. Whether a software-driven processed image violates some ethereal ethical representation of the truth is like debating Van Gogh's interpretation of the Starry Night he saw while looking out his window. This is not an issue of ethics; you either like Van Gogh's Starry Night artistic presentation of the night sky or not. |
|
2.15
#...
·
1
like
|
|---|
| To me, the more interesting question is whether AI-generated images truly have an author as we understand that term. A topic more worthy of debate is whether an AI-generated image has an "author" as that term is understood under copyright law and, if so, who is it? The software programmer? Arguably, it is the person who licensed the code from the developer and loaded the image, but that person didn't actually create the resulting image, did she? This would be a great law review article, and I may have to tackle this. Does an AI-generated image have an author as that term is understood under copyright law, because if the answer is no, well then, none of our AI-generated images are protected as copyrights. |
