The Truth about AI - Are ChatGPT and BlurXTerminator Images Even REAL?
In this video, we explore the world of AI processing with and to determine if the images are real or not. Get ready to dive into the world of astrophotogr...
The Space Koala Astrophotography by Luca Bartek
You explained this so perfectly and brought such important points up. The use of AI is exponentially growing with a acceleration I think I have never experienced for any technology before. Glad to see folks like you being critical but not dismissive. Thanks for your great contribution.
Glad you enjoyed the deep dive, it’s definitely a wild time to be exploring this tech!
(That was the YouTube recommended AI response, it seemed fitting to use it. Usually I don’t as they sound really fake haha)
huh- heavy stuff. i like the thought that You want us to understand what we are doing ;)
Great explanation of how ai works. Thank you
Knowing how bad AIs are at keeping a consistent visual record, like with the infamous “AI yellowing plague”, I cannot fathom how astrophotography post-processing would not drift from a journalistic, faithful enhancement approach to a full-on “Photoshop” prettification mindset. I get why you’re putting on the white gloves to talk about this, but honestly, you can drop the gloves, the white ones and the hockey ones. This debate has been heated since the dawn of astrophotography, so don’t be shy about pointing out that this AI-driven “prettification” approach is a slippery slope.
while I agree with you that this needs to be nipped at the bud (well it's a little too late for that lol) , I find that a many people using genAI to make their (e.g. smart scope) images prettier genuinely do not know any better. They are happy that they're able to create something out of "their" photo that is nice and detailed and often take it as an attack when you bluntly point out what actually happened. My aim here is to educate those willing to listen - and likely whoever is sitting through this video wants to learn
Just the explanation I was looking for. @11:05 If you only care about BlurX.
Well done. I have used ChatGpt to experiment on the result using one of my S50 photos.. In general it removed stars from the original image, added new stars which were not there and added detail e.g. dust lanes which were not there. Great picture it produced but was not mine. I did this because I was seeing photos which were AI generated on the S50 forums and was challenged when I said I believed it was AI. I have no problem people posting these images but believe they must say it was AI modified and also show the original photo. Anyway great video, I liked it so much I subscribed.
I have had the same experience as you.. some people won’t accept it’s not the original photo even after you point out the obvious differences. Thanks for the sub :)
The only questionable AI image enhancement I've used thus far is Upscayl. I'm limiting it to making wallpaper versions of my photos.
Very good video, clear and interesting! I still have a little reservation about BlurX when you say that he doesn't invent details. From what I understand, BlurX was trained on images with typical long exposure sampling, so if you use finer sampling (<1"/pixel), it risks adding "plausible" details rather than real ones. This makes me wonder if this tool could mask a very faint supernova by confusing it with noise, or transform a small, diffuse galaxy into a star because it's a more common pattern in its training data. It could even create fake structures, for example, non-existent bridges between stars, simply to make the image more visually consistent. This raises another question: if everyone uses the same AI-based tools, don't we risk homogenizing astrophotography? The diversity and originality of images could decrease, and ultimately, this could even weaken the bond between amateurs and professionals, compromising the scientific reality of the resulting photos. I'd be curious to hear your thoughts on this!
It cannot add real details that are not there because it doesn’t recognize the part of the sky it’s looking at (as it was trained on fake stars). However, artefacts *are* possible especially with lower SNR data
@the_space_koala Yes, I agree, and I understand that this model doesn't work (almost not, in reality) like a generative model: it doesn't invent if it doesn't know. But the line is thin, and I've seen in quite a few photos processed with Blurx that sometimes there are pixel bridges between very close stars, fine lines that replace contrasts that are not or barely defined. Having images with a low SNR increases the probability of errors. While a simple knowledge of the objects the user is processing can rule out certain errors, it's up to us to be careful if we let certain shots slip through the cracks... In any case, I like your approach and the subjects you cover!
L'IA générative n'est là que pour nous voler nos photos, d'ailleurs j'en mets très peu sur internet, et en faible résolution. Je vois parfois sur Facebook des images "parfaites" de Voie lactée + montagnes + cascade trop bien placées ; comme je connais ces endroits (Dolomites etc.) parfois je vois une Voie lactée pas du tout au sud (dans l’hémisphère nord). BlurXTerminator est un outil extraordinaire, en un clic il arrondit et réduit les étoiles distordues aux bords de mes images !
Never use Ai , what's round the corner . ..
Really appreciate this explanation and description. It clears up a lot of things for me. But one "grey" area still is the use of AI-driven tools to "remove" aberrations or distractions from an image. I routinely use such tools to "clean up" my terrestrial photographs (removing electrical lines, garbage bins, even people) and these areas are filled in with pixels "generated" from the surrounding pixels. You could argue that using AI-driven tools in astrophotography to clean up halos or gradients or to deal with that last bit of dust mote falls into the same grey area. But you are correct - we are not creating new astronomical structures - we are simply highlighting and enhancing the structural data that is already there. I have never used and never will use a large language model like ChatGPT in my astrophotography.
I agree this is a bit of a gray area - I think everyone kind of has to use their own judgement here. For me personally, I think it's fine to use these tools to clean up a hot pixel or other tiny technical glitches, but not to replace whole areas of the image. For example, I have seen people remove a giant diffraction spike from a photo taken with a Newtonian telescope, that was going through the whole image. Apart from the fact that it remained visible, it was also disturbing because you know it should be there!
Thanks for the video - very helpful. As a test, I asked AI to enhance an image of M27 that I was working on. I wasn't completely sold on the AI enhanced image, but did see some things I could improve on. I continued working on the image. Later I submitted the image and asked AI to evaluate it and make recommendations to improve it. Some of the suggestions were, indeed, helpful and I ended up with a better image as a result. If anyone is interested, here is the result I got from the inquiry . . . Image Analysis & Enhancement Suggestions Strengths of Your Image Excellent Focus: The nebula is sharp, with intricate detail visible in the filaments and shell. Color Balance: The cyan and red hues are well-separated, revealing the classic structure of M27. Star Field: The stars are round and well-distributed, with minimal trailing or distortion. Enhancement Opportunities If you’d like to further refine your image, here are some targeted suggestions: Increase Nebula Contrast: Use local contrast enhancement (such as LocalHistogramEqualization in PixInsight) to make the nebula’s inner details pop even more. Reduce Tiny Stars: The background is filled with small stars. Using star reduction techniques (like MorphologicalTransformation or StarNet/StarXTerminator) can help the nebula stand out more. Slight Noise Reduction: There is a faint graininess in the darkest background areas. Applying gentle noise reduction (MultiscaleLinearTransform or ACDNR) can smooth the sky without losing detail. Saturation Boost: If you want a more dramatic look, a subtle increase in color saturation, especially in the red emission regions, can create more visual impact. Vignette Correction: If there’s any residual darkening towards the corners, a flat-field correction or careful background extraction can help.
To me, AI has helped streamline processes that I just simply would not be able to comprehend or would have to take faaaaar too much time learning, something I can’t really afford just for a hobby. I take my own pictures, capture it using my own time and energy, but I like how I can now simplify the editing process to make sure what I captured is enhanced so that it can reflect the nature of what I photographed. I like tools like BlurXterminator because I know it isn’t adding to my image with its own data but is rather helping me arrange my data better. There’s a difference between enhancing and creating, and I could never use AI for the latter. AI as an assistant is brilliant, but it should never become the artist. There’s no fulfilment when that occurs; at that point I would give up this hobby!
This is a great explanation. It's pretty easy to tell (for most people) when AI starts to hallucinate vs when it stays true to the data that was already there. BlurX does an excellent job at that and it's a little bit like magic when you compare before and after.
thanks Naz! I agree. The concerning thing is that the general population (non astro people) are already not able to tell an AI astro image from a photo, and the rate at which it's improving it's only a matter of when, not if, they will be indistinguishable from real data.
That was a very important video and some great explanation! Thanks a lot!
Am into astrophotography but do not care about 'beautiful' images. Will not win any contest but still have lots of rewarding and very interesting fun.
great video on the ai subject
Another superb video my you! This should be "must viewing" for astro-imagers. Thanks! Your talks at NEAIC were also superb.
thank you for your kind words!
She’s really lovely. There, I said it. Sorry!
I think not only is using AI for generating "astronomical" images not astro-photography at all, it also takes away the real reason why I do this kind of photography - or photography in general (because I don't explicitly consider myself an astro-photograph, more like someone who is very interested in astronomy and photography, which is maybe not the same, you tell me). Spending hours and hours on capturing a nebula or an interesting structure in the night sky is appealing and rewarding (sometimes also frustrating when it all goes wrong, I admit), but telling a machine to do something for me, is taking away all the education around this. Each mistake I make, I also learn something new and that is what keeps me hooked to photography. Especially astro-photography is a challenge because of the effort it takes to get a decent result and much more effort and time it takes to get a better result. Where is the challenge in life for letting someone (or something) else do this for us? If it were a human, the challenge would be to teach what I already had learned, but a machine like AI does not give the feedback as a human does and then it is boring to me. Yes, I must admit that it gets my curiosity and I try things, but to use it as a substitute? No thanks. And if any program changes my images by guessing what was there, I will pass. This is about astro-photography, but I really think it counts for all kinds of photography, at least for those people who enjoy the process of photographing from start to finish.