TheDarkgg wrote:
doudou971 wrote:
zorglub123 wrote:
"… this AI is also capable of erasing changes made to photos to recover the original image."
Well let's see …. 1) I don't see how it would know what we wanted to edit … 2) editing is destructive … so the original info is no longer there .. how she finds her
When you touch up a photo, you apply filters (scrub, image, Gaussian blur, sometimes extraction, insertion, etc.). These filters that your eye does not perceive if the retouching is done well form "modification patterns" quite characteristic on a bitmap and that it is therefore not difficult to detect an AI.
You get like 1 million photos. You apply filters randomly on a proportion of them which allows you to have learning data for your A.I.
This is a particularly simple case study where the AI ​​is very strong (because the learning situation you can provide it to the machine all by yourself by making unambiguous positives and negatives in no time). No need to go and buy data from the big collectors we all know
Yes, by having the original image or thousands of the same retouched images.
In the case of a single retouching image, a face for example to which we would have erased beauty points, I do not see how an AI would be able to remove the retouching and re-display them.
Yes, AI is easy to detect retouching, a human also expresses it but from there reappear an info that has been destroyed …
The case that you explain does not remove the retouching and will never display the original image, the AI ​​will just add information that it assumes to be the right one and that was destroyed by retouching by having compared learning by other photos mainly the same (a sunset, a landscape, …).
The information that is destroyed will never be available again if the original is not there.
You haven't figured out how AI works.
And I never said that we could find the original image. That said, it is indeed possible to have the image that was probably present before retouching. But this is only a proba. The news has indeed made a big shortcut. For the example of moles, the AI ​​can detect the use of a tool on circular areas which, with training, corresponds to teenage faces more likely to pimples than to beauty that we can recreate.
To simplify normally, you constitute a matrix (norm) which is supposed, when you give it a column (a state of) to send you back the true or the false.
You provide your matrix with millions of entries and you correct the coefficients of your matrix as you go along so that the entries supposed to give true give true and those supposed to give false give very false (this mathematical process is generally called " descent of the gradient ". This algo dates from the 60s).
Here your positive entries are retouched images and the negative are unretouched images. The presence of modification patterns that is impossible to describe in general ends up being detected correctly by your matrix once you have provided it with enough learning data (here they are very simple to make so this is a simple case, a case of school that you learn at the beginning that you do AI).
So basically, if a guy has used tools that have been given in sufficient numbers to enter AI, it is highly likely that there will be detection.
If, on the other hand, the modification processes were not given at the learning input, the AI ​​will see nothing. However, everyone almost always uses the same retouching procedures because retouching consists precisely in using these precise tools!
I hope I have been clear.
