@jeffcliff@ceo_of_monoeye_dating@ai Jeff, last we talked you said something about hanging out with a flat earth chick and I gave you some globe earth ammunition. Did you two hang out?
honestly i don't really see it but i pretty much have to accept this reasoning because I don't think i'm going to get further down into this paper at this rate.
ie i'll come back to this thread the next time we disagree on an possibly deepfaked image to compare notes
@jeffcliff Moreover, you should be careful to avoid clipping - either up or down - else your image will look smooth when it was not. Ideally you want to be doing these operations in Python with OpenCV and not GIMP, but for your sake let's use GIMP.
Now, let's ignore color maps and do that with this image. Convert to greyscale (I'm going to be sloppy and assume that taking the Luminance is how GIMP converts to greyscale, but I could be wrong), apply the Sobel filter (only horizontal, or only vertical), and equalize. Here we are.
You should see what looks like a lot of mountain ridges in the "body," - which I have described as looking "sinewy." The edges are chaotic, even though this was not quite perceptible to a person before.
Compare that to the smooth surfaces, which are uniform in color. Compare that to the other rough surfaces (like the glove), which - does show such ridges but in a fashion typical of such a rubber glove.
@jeffcliff Now, if you simply apply the sobel filter, you will see the edges - but you will end up with spikes in the histogram. When this happens, the filtered image looks smooth when it is not, as the small differences in pixel values are meaningful. You need to do something to amplify those small differences so that you can actually see the noise. Equalizing the histograms or normalizing the image are both good ways to do this. This is why I :keku: at you for clobbering the noise earlier, when you should have amplified it to see what you were looking for.
Moreover, you should also take the sobel filters in both the X and Y directions, and view them separately.
Note that there is a good argument here for taking a Laplacian filter instead. I have not investigated this myself, but I suspect it to be a good idea - although what you will see will look different, and will require a new way of explaining it and a new heuristic.
Moreover, we only do this for the Luminance, because it is where the bulk of the interpretable image data is.
@jeffcliff Let's go over the basics. It is a fact that when AI is used to generate an image, it produces what is often called a "fingerprint." This fingerprint is usually in the form of noise of a particular frequency (that is, it is periodic), which is often high-frequency. It has also been observed in practice that this noise is also somewhat sinewy looking. If you are skeptical of this, you may start here: https://arxiv.org/abs/2101.09781
It is a second fact that taking the derivative of a function is equivalent to multiplying its fourier transform by the frequency (up to some constants, which may be complex). You know this from your Fourier Analysis course.
The Sobel Filter is equivalent to taking a single directional derivative. So, we should expect it to be a lazy way of spotting the noise generated by the AI.
Note that if we were talking among researchers, we would make a much better argument for using a Notch Filter instead - but as I am trying to give people a simple heuristic they can use themselves, I give to them the Sobel Filter, which is implemented everywhere, and not the Notch Filter.
This is the broad overview, and I think this is what you understand.
I couldn't possibly know, for example, in 2007, that i would get involved with bitcoin in march 2009. Betting the future of some child on a cryptocurrency that didn't exist yet, when i could barely get like my best friend and girlfriend on ripple (and even then both of them treated it as worthless 'jeff bucks') would have been supremely reckless, and given i was often homeless and broke, i was in no position to raise a child and it was looking like that wasn't going to change. Winning the lottery not once but twice was not something I could have predicted in retrospect. All I looked forward to was starving/freezing on the street when things inevitably fell apart. Every single man in my life who was having kids, their relationships fell apart & they were stuck paying child support* and it was leading them into ruin, one by one. Or they just were deadbeat dads, and I never wanted to be *that* guy. I knew that sooner or later I'd knock some chick up, and that the barely-treading-water level of survival i had would fall apart. And sure enough: i lost my job right after that, in 2007, and then the next one, in early 2009, having to live out of a tent for abit, until my 2 concurrent fulltime jobs gave me enough of a cushion to find a place.
but i figured, even if i did eventually get out of that kind of situation, by some miracle, there were so many kids that were *not* being taken care of, that needed a father figure in their life, that i'd much rather help them then add another mouth to feed. There were child prostitutes at the schools i helped instruct piano at, for example -- they were just some of the example of the 1 in 5 children in saskatchewan who live under the poverty line and who face food insecurity and whatnot. I figured that there'd be a single mom that would appreciate bailing her out of her mess, to the extent I could, and that would be better than just creating more children when there's already too many.
* and fun fact: in canada child support doesn't have to be proportional to your income. The judge can just up and say 'you're paying N$ per month' even if you only make (N-k)$/month , in effect meaning you have to borrow money to even pay the support, nevermind to survive.