Yes yes something something this is inevitable therefore the galaxy brained thing is to be completely fine with how it's applied in an unfettered capitalistic system. Also something something if you're against this technology wrecking careers, reputations, and sometimes lives you might be too emotional and against Materialism(tm). :very-intelligent:
Can they stand john up in a police station with Chinese words on the wall?
Just asking for the few thousand photos with perfectly repeated artifacts in the Chinese Police "hacked documents" last year that were meant to implicate them in mass detention.
You know, the ones where a couple dozen people had a jacket with EXACTLY the same thread on the chest pulled and facing the same direction even when shown wearing it in different states of dress and different police stations, among many other generated artifacts.
spoiler
The necks and ears on a bunch of people literally not being attached to their head properly because the generated parts of the image not lining up that got missed in QC were a bigger tell, but yeah. Oh and there were no photos of the jacket WITHOUT the artifacts, to clear up the sample size worries. I know I sound like a conspiracy theorist but tech has BEEN capable of this shit for a WHILE. In the PUBLIC.
I remember that, it was so shitty and obvious too.
I remember that coming up several months ago. Started on GenZedong and I initially saw it on the CTH discord, and I remember pretty thoroughly debunking it.
It's correct that there was a propaganda element at play. The photos are most likely real photos used for identification and facial recognition purposes, which is nothing new in China, or in many other places for that matter. So, real, but not particularly noteworthy, and probably not even people who went to a detention camp. It could be ID photos for all we know.
I clearly remember 2 images where one is a guy, the other has the exact some composition, face, clothes and everything, except with longer hair and feminine make up. Like if you choose the gender swap option in faceapp.
As for resolution, there was a website called "this person does not exist.com" or something that generates HD faces with great detail long before the stable diffusion business. The tech is not new.
I don't think all 1000 images are fake. Whether they are images from Xinjiang need to be proven but I believe they started with around 100 "real"(as in actual people photographed with cameras, I don't think they generated from scratch) images and used some process to "breed" or "multiply" them in order to have the 1000. There's no way all 1000 is real and all the glaring artifacts can't simply be explained away.
That site does 1024x1024 images, I think I remember concluding that the source photographs would have had to have been at least 2048 initially. And it does bear the usual suspects as far as artifacts go -- I can see incoherent backgrounds and I notice that light reflections on the eyes are very inconsistent and make no sense, as the most obvious ones.
I can tell you that if they had 100 real images, they'd have been better off making 1000 real images than attempting to finetune a face generator. But like I said, the images don't really prove anything of significance if they are authentic. They could be prisoner intake, or they could be state issued ID photos. So I'm not worried about it.
There was literally peoples faces merged into other faces in some sets, including mouths inside of eye sockets. There was all sorts of discussion about face direction normalization for facial recognition and other nonsense like that, but for the most part none of that is worth spending too much effort on because it's not my main point. I'm sure some images are real to some extent, some are altered for processing purposes, or some are just caused by faulty equipment. The file corruption one is a mixed bag because I agree on the obvious corruption artifacts like an image having a hard line where it color shifts/stops lining up as was typical with shit digital cameras back in the day are likely just artifacts.
But none of the artifacts typically associated with image corruption or camera malfunction present themselves in ways that actually corrupt a single object/subject IN an image. Generally corruption of a single object or subject IN a composed image (with multiple distinct items) is because... the composition had artifacts, not the file format. Now unless China was experiencing localized reality artifacts when the photos were taken or the camera somehow created them before the image was rendered to a digital file (similar to the lines discussed above but something more unique I'm unaware of), then the composition artifacts occurred because they were generated. I'm not at all claiming to be any kind of expert here and I generally don't say shit like this lightly because I'm incredibly wary of holding beliefs that don't have fairly solid foundations. I know that probs doesn't mean much from a stranger online but generally I'm in the opposite camp and would prefer to wait longer and have more and overwhelming evidence before making claims, and the jacket thing was just what kicked off my curiosity to look through more images in the first place.
The clothing artifact IS because it was the same jacket. I agree. I just think it was the same Jacket in the reference image that the model was pulling pieces from to make the generated ones. From what I could determine the jacket in question would have had to have been shared around multiple prisoners (totally reasonable for processing reasons I can understand this or invent reasons it's plausible) multiple DEPARTMENTS (less plausible, but I can still stretch to get there probably, but it lists where they were processed in the documents with each image on their website) but then on top of all that most of the images people are wearing whatever they want, wearing different jackets(some other clothing "clones" exist too, but the jacket was the most standing out one because of the thread) and it just all seems rather unclear why it would be needed in the first place or why it's seemingly exempt in many other cases.
That's when out of curiosity I went looking further and found tons of images in the image directory that didn't appear to be linked on the website, some of which had things as I described above like mouths in places they shouldn't be, noses clearly not sitting in the right spot on the face or pointing a different direction than the head is facing, ear lobes rendering too high or not at all, or inside the cheek (kind of in front of the other face bits) etc.
I agree at least SOME (maybe even a majority) of the images are real. I just can't accept the idea they're police files and not like dmv or passport photos or some shit someone used to generate more of the same and then threw together spooky explanations for to get that sweet anti-china funding or because it's their job. At least not without harder evidence. I'm not as much insisting it's all a conspiracy(of the magical and not boring kind) as much as I am insisting it more than meets my threshold of "I'm not taking the US at it's word, that alone is shady enough plus look at this weird shit, either start explaining or fuck off".
More dismissing the claims rather than making ones of my own. The claims of my own are simply that more and more I have less trust in images and video without trusting the source, especially when some anomalies are present, and especially when they come from literal fash spies.
If you're willing to concede that maybe even a majority of the images are real, why stay committed to the idea that some are AI generated? It'd simply be more practical to show 500 real images rather than show 1000 images where half are fake and to risk detection, just because of how difficult it would be to make any amount of fake images of sufficient quality.
I don't trust it by default either, which is why I choose to completely ignore the given context for the images and simply acknowledge what I can see. I see pictures of a number of people, mostly seeming to be the same ethnicity, they're in a room with a fairly boring background wearing dark clothes that don't fit, their heads are all rotated the same way. It's clearly facial recognition photos based off of those facts. What are they for? We don't know. Probably just standard ID photos. People being of the same ethnicity doesn't prove targeting either since they could be cherrypicked photos. That's enough to discredit it without even questioning the authenticity of the photos themselves and as a whole it is far more in line with standard propaganda tactics.
I mostly agree with the premise of all of that and was/is the way I dismiss it's purpose in the first place. I just ALSO think at least a non zero (fairly large) portion were doctored/generated too. It's not so much a commitment as it is a thing I was curious about and then saw with my own eyes. You can go and browse the top level directory on the website hosting the "leaks" by deleting the filename from the url. Regardless as you said it really doesn't matter for the purposes of denying their purpose to at least reasonably aware of the shit our government does people. While I agree just using larger photo sets with false representation is something our government would do or would find easier, I don't agree it's something a grifter with some access to a stolen archive of photos wouldn't do to scam money OUT of our government, who would eat it up because aligned interests/relative ease. As to why they needed say 2800 instead of I dunno, 1400? No idea, couldn't really be bothered to speculate on it but perhaps there was some reason to enlarge the dataset, financial or propaganda or overpromising a "leak" or who knows. Never underestimate the ineptitude of people who willingly refuse to analyze the world around them lest they become a communist.
The ability to train existing open models to a fairly decent capability has come a long way in the last few years though. As early as least year I was able to get pretty good results capable of fooling a lot of people I know IRL using nothing but a google compute instance with gpu and 25-50 images of either random or specific subjects depending on the purpose before I got too bored because despite being easy it IS time consuming, and continued with the main purpose of training anti-ai detection models as part of thing I did that I can't link to without doxxing myself. Nothing to consider myself any kind of authority of any kind on the matter though, just saying some models are a lot better than others, and depending on relative strengths of the model and how well it aligns with a purpose it's definitely at least as reasonable as "These are all normal images of eyes with teeth".
Edit: Just to add, a TON of the people tangentially related to/in the periphery of the PMC/"Think Tank"/"That other term for 'non political' organizations that exist solely to fuck with other countries for profit but pretend to be charities" circles are just straight grifters. It's a huge majority of them. They will sell/buy/fabricate anything if they can reasonably spin it as fuel for their home countries political war. Hell Zenz himself sold the idea of a genocide on moving the decimal place in a fucking table and implying some massive number of women were being sterilized instead of the real story of single digits (a similar percentage to the rest of the populace) getting reversible contraceptive proceedures (IUDs) and the public just ate that shit up and even when you bring it up it gets waved away.