That being said, there were a few comments on here about gaussian blur and deconvolution, which I would like to tackle. First, I need to mention that I do not have an maths/engineering background. I am familiar with some concepts, as I've used deconvolution via FFT several years ago during my PhD, but while I am aware of the process, I don't know all the details. I certainly didn't know that the image that was gaussian blurred could be sharpened perfectly - I will have to look into that. In fact, I used gaussian blur to redact some private information (like in screenshots), and it's very helpful to know if I haven't redacted anything and the data is recoverable. Wow.
I would love to learn more about the types of blur that cannot be deconvoluted.
However, please have in mind that in my experiment:
1) I also downsampled the image to 170x170, which, as far as I know, is an information-destructive process
2) The camera doesn't have the access to my original gaussian blurred image, but that image + whatever blur and distortion was introduced when I was taking the photo from far away, (whatever algo they are using doesn't have access to the original blurred image to run a perfect deconvolution on)
3) Lastly, I also clipped the highlights in the last example, which is also destructive (non-reversible), and the AI hallucinated details there as well
So I am comfortable saying that it's not deconvolution which "unblurs" the image and sharpens the details, but what I said - an AI model trained on moon images that uses image matching and a neural network to fill in the data.
Thank you again for your engagement and your thoughtful comments, I really appreciate them, and have learned a lot just by reading them!