Fake Ai Portraits Are Fooling Everyone 

Gotta love that the quiz mocks you after youre forced to enter your email (to view results) even if I was good over 50%.

Thanks Pratik. What's wild is these images were all created in July of 2023 and it has just taken me 5 months to run the test and produce the video. The images in this test are already very outdated! I think a lot of photographers are oblivious to how the general public sees and views our work. One interesting thing that came out of this video is how many normal non photographers thought the real images were AI because of the lighting and image quality alone. I don't think we EVER think about this but we might be at a point where our overly stylized REAL photos will be viewed exactly the same as AI work. It's already happening in many of our Fstoppers Critiques. It makes me wonder, if the average person thinks real photography is AI, why should we view AI creations so negatively and act like our work is so much better? -P

at least someone who gets it

They were all modeled after Peter and Martin Schoeller. There are some AI images in Peter's style and Martin's as well as real ones as well. It's a mixed bag. The thing is though, most normal people don't know anything about catch lights and many people actually thought they looked weird and fake which brings up a whole other conversation. What does it mean when normal people think our highly stylized photos are now AI? -P

@@FStoppers 31/44. I saw the same thing in the images. A lot of them has the same catch light in the eyes. Some was easier to see than others.

@@FStoppers I totally agree! who doesn't work in photography for sure has a hard time guessing who has done these photos. For me, since AI is a tool, we photographers need to communicate to our clients that we create images from real life, not synthetically like an AI does, and is not easy!

@@ZanonGabriele The thing is would the client even care?

@@CallMeRabbitzUSVI depends on the application. If the goal is finding a photo for a flyer no. If the client wants a reportage of an event, or something very specific.. I hope so!

how many people make an actual living from classical painting today VS when photography was invented?

Some of the AI have the same lighting too. Only photographers would notice this though. The general public actually thinks it looks strange 🤷🏻‍♂️ -P

@@FStoppers I got 34, but I timed out 4 of them, and I never heard about Peter Hurley, but I've used SD tons, and I know how hard it is the get even eye reflections and how much generation you need to do to actually get ok results, so I attacked it from that perspective. Another thing that AI usually messes up is the edge, but as these photos all are shot against a background not real things it's harder to catch the AI. I usually look for things like plants, architecture that doesn't make sense, sizes that look off etc.

I m happy I still managed to score 70% on this quiz, but I think after MJ6 it would probably be lower

I haven't used MJ6 yet but I feel like what I was generating with my system in Stable Diffusion might be better than MJ6. I'll have to compare. I posted a generation of Natalie Portman on the Fstoppers Article using MJ5.1 and my algorithm and there is a night and day difference. -P

That seems like a weird stance. If I need code to make a website or automation work correctly and grab it from somewhere else, does that make me more or less intelligent? -P

@@FStoppers yes but wiser?

@@FStoppers If you steal most of your code then yes, you are not intelligent

Sadly it won't work, the Open Source community is already skirting around those protections. It's just security theatre I'm afraid

I dunno, I could increase the time for sure but I placed a time limit on the real people as well. I asked them them to talk through their thought process sometime just so I had content to share in video form. I actually think the digital version is easier unless maybe you take the test on a small cell phone. On my desktop computer, the digital copies show artifacts better than the smaller 4x6 prints do in person. Pixels kind of blur together in printed form where as on a 4k monitor it's easier to see. But the point isn't to cause you to score poorly. These images were all created in August 2023 and I'm sure today they would be even better and in a year they will be even better than now. The theme is most people aren't able to identify AI images EVEN when told that's what they are here to do. The ramifications of this technology is going to be monumental. -P

@@FStoppers Only allowing people to see the images for only a few seconds is manipulation of the results (Now you never claim it to be scientific) but for a better test increase he amount of time one takes a look at each image on an exam. You are treating it like people are scrolling through Instagram which might seem clever at first. Forget that when we are given a quiz we tend to look over the question ( or in this case pictures ) over and over to make sure we understand it and can answer accordingly. I mean hell people take longer than that to respond to Buzzfeed quizzes on which Hufflepuff are you?

It already is, alot of surgeons talk about how their clients would come in with a clearly filtered picture will be asked to make them look just like that

The in person tests didn’t have a limit since I needed them to explain their thought process but overall they went through them in 10 seconds or so. The in person test is way harder than the online test IMO. -P

I show it in the stable Diffusion section of this video. -P

Most people say "I loved that ending" but then failed to realize it happened a few times earlier. -P

I FELT like something was off but I couldn't figure out what or why until he said it at the end!!!

No. AI created images have no human input. It seems like it does because you are putting in prompts and manipulating data sets for a desired outcome but it isn't a human creation. The courts have ruled as such too. I've heard it compared to a horse owner commissioning a painter to create a painting of their horse looking majestic in front of their house in late day sun. The final resulting painting followed those suggestions but the person who commissioned the painting is no more the author than the horse. The painter alone is the creator and it is 100% his work. Capturing a photograph and taking it into lightroom and moving the sliders to oblivion is still a human action. Now elements created from generative fill are up for debate but as a whole AI created art and art manipulated by a technician in photoshop or lightroom are completely different things. -P

@@FStoppers - What I was saying was, Would an image straight out of camera without PS or LR be confused with an AI image ? Were these (camera) images manipulated with PS/LR ?

Yes, these headshots are professional headshots so the models often had make up, styling, and slight retouching. They were not straight out of camera. But neither were the AI images in the sense that they were told to replicate the style of these real images. Now if I took real images straight out of the camera with no retouching, I could also have the AI try to replicate that too. The goal of this test was to have two sets of images, real and AI, that looked similar in style. -P

@@FStoppers Wait you cited the legal precedent in this comment but not the other one where it was relevant? Are you picking and choosing your responses based on the Original comment? Not sarcasm I am genuinely confused unto how your replying to these comments. Seems inconsistent

Not sure what you are referencing. Legally, AI created images cannot be copyrighted. But also if you copy an image and transform it enough, you may not be infringing on it. It’s complicated and all the AI stuff hasn’t been tested in court yet. -P

Both. The images used as real photographs are definitely unique in their lighting but usually they are minimally retouched. That said, modern photography whether it be headshots or advertising images are ALWAYS retouched to some degree so it's a fair comparison. What is an interesting question that has come out of this is, are these real photos that are super stylized with lighting going so far that people now view them as AI generated even though they are real? -P

Yes and no. They are definitely smarter in some ways and probably out of touch in others. It's kind of like the street smarts vs book smarts. If you've traveled the world, experienced different cultures, worked with AI yourself, seen the latest gadgets and tech at shows like CES (or in your friends' homes), and are generally on top of what the cutting edge of technology can offer, then you are probably going to recognize the subtleties of AI created art better than say someone who isn't on the internet, doesn't know what technology is capable of, lives pay check to paycheck and doesn't have the means to invest in this sort of debate or curiosity. I know because I have friends from my home state that are very rural and while they are very smart in many ways, they are pretty gullible in other ways too. Being rich and dumb aren't mutually exclusive but it can happen. Anecdotal I know but sometimes it's pretty wild hearing and seeing what my neighbors are involved in. For example, one guy is about to buy a $100m device that launches packages into space and into orbit. It's like being surrounded by Elon Musk type people sometimes. Sorry for the long answer. -P

​@@FStoppers thanks for the response! Long answer appreciated 🙏. Looking at my comment in retrospect, it now feels a little confrontational, which wasn't my intention at all; and I apologize. Yes, I agree with you about being more traveled and having exposure to different industries & cultures having a huge impact on one's intuitions about technology, but I think my main point was just that we shouldn't expect them to have any better results simply due to them being entrepreneurs, CEOs, investors, etc. Perhaps it was just the way it was presented, maybe it made me feel a little defensive, because their intellect was being validated by their financial success. And as has become clear, Elon Musk is a great exploiter, but not generally a great person (thinking specifically about the anti-union & anti-equity stances, lawsuits over racism at Tesla which he lost, his joking about Bernie Sanders being dead, his suggesting taking voting rights away from certain US citizens, and much more). I feel for you if you are surrounded by these types on a regular basis. I am occasionally, and it usually leaves me feeling hopeless for humanity lol. I once had a very tense conversation with a couple of millionaire+ investors about the detriments of the stock market and how it guarantees ever-widening income disparities, all varieties of discrimination, and infinite war. My argument was so clear and rational that they actually agreed with me! Then, after about 5 seconds, the one friend turned to the other and immediately resumed his questions about bitcoin investments. Smh. Anyway, I bet you'd get similar test results if you went to a ghetto in Harlem, or Skidrow in LA, or a PTA meeting. It would be interesting to see how AI software engineers themselves would do on this test. I wonder if there'd be a significant difference. I imagine so!

Lots to digest here and I appreciate your response. 1) I had one engineer take the test who is very up to date on AI as he does a lot of programming in his engineering. His score was 30/44 which was basically right in the average of all scores (68% correct). I have another friend who is making insane AI stuff that helped me get stable diffusion running. He has made a full Fortnite scene of our entire neighborhood and has a $300,000 computer just for AI stuff (neural nodes that connect to your forehead and the Augmented reality glasses to go with it). I should have gotten him on camera to see how he did but the AI stuff he creates is less hyper real and more anime metaverse type stuff. I regret not keeping him arms distance away and letting him run the test because he would have been the best person aside from Pye Jirsa who is a professional photographer (only photog who did the test in person). 2) Yeah no one here is like Elon Musk in personality but I just mean they dream huge dreams and then actually make them happen. When someone says they are going to build and design a $15m yacht, I never believe them but we just filmed the rental video for that yacht 3 years after they told me they wanted to do it. Someone else says they want to write their first book and get it to the top of the NYT best sellers list and hot damn, they did it! We've been trying to solve this weird sargasm issue here (think large plumes of seaweed that come up on the beaches during the summer) and someone moves here and suggests we raise money and build a $70M artificial reef to prevent the sargasm from ever even approaching our beach...it sounds insane but they are working on it now. So yeah, if they were narcissistic like Elon and kind of meddling in chaos and conspiracy theories then I couldn't deal with it either. But no one here, or anyone I've met at least, is like that at all. 3) I'm not so sure about the results I'd get from Skidrow or Harlem. My experiences with Skidrow would tell me I couldn't even get anyone to stop and take the test for more than 10 seconds and I'd probably get the wildest answers that aren't even options on my quiz. Good discussion though, there are so many crazy interesting conversations to be had about this test. Maybe Lee and I will do an Fstoppers Live about some of the results of this test and the other tests I haven't released yet which are also wild too (I tested other things than just headshots). -P

How will anyone know what images were used in the model to produce the work itself? You'd have to file a case and go through discovery and subpoena the actual companies just in the hopes that you find your client's work was part of the data set. That seems impossible. Unless the AI spits out an image that is 95% the same, I don't think you can prove it. Then there is the "transformative" part of copyright law that allows people right now to change a copyright image "significantly" and not be infringing. That significantly isn't as large of a barrier as you might think. -P

@@FStoppers You are forgetting two important things 1. The datasets for each model is known (for now) and Getty images was clearly apart of that set which means legislation has to come early on before a fresh new community driven dataset emerges which is untraceable. 2. That transformative nature only applies if the works are done by humans. ( There is also legal precedent for this, which is currently trying to be appealed by some Ai company and it doesnt look to be going well) So if the computer is what is creating the image then that image does not have legal protections which makes it toxic for companies that would like to own their IP (Intellectual property). I find it convenient (And a little suspicious) that you leave these two things out of your arguments?

@@FStoppers Some of the AI photos I seen, hardly any, really look like that have been taken from photo, even with prompt to ask for ethnicity etc. I think there is definitely a test case to establish if AI 'Transformative' can be applied to a none human entity. Its going to be interesting. I think there has to be a solid way of telling AI NOT to scan your images for learning and for it to comply! JMI

Yep, it's an Fstoppers quiz. The results are sent to you -P

True, but that is the very nature of modern photography today. We have to compare apples to apples. The cover image of Martha Steward's Sports Illustrated Swimsuit issue is no where near being un retouched. Probably no celebrity portrait in any magazine looks anything like what it did straight out of the camera. Straight out of camera is not the standard by which modern photography should be judged. -P

How did you do on the test? -P

Curious if you are talking about the quiz? Those images are all Peter Hurley’s and he is known for not retouching his portraits much at all. The waxiness you are seeing is a combination of lighting, wide open aperture, and makeup (on women). They are pretty far from the heavily retouched images you might see on the cover of a magazine (very little skin retouch, no liquify tool, etc). Doesn’t mean you can’t see they look unique but I can spot the real skin away fairly easily but prob because I’m a photographer. -P

@@FStoppers would like to see them at 50mp+

We actually have another video we are working on where we ask people to look at posters printed at 8,16, and 200 megapixels. No one can tell the difference 😂 -P

It's the way he is dressed, big hair but black top will obscure everything except his head. Also lol this comment

I also did this test. Landscapes are actually more difficult for AI to create in a realistic way. Architecture was the genre that fooled the most people. -

That makes sense, actually. I just naturally assume it's better at angular stuff😄@@FStoppers

Yeah it's kind of surprising. You'd think reproducing random forests, trees, oceans, and cliffs would be super easy but for some reason they look so magical and unrealistic. I probably just need a better base model or something. With Architecture it definitely has a difficult time getting windows and motifs that repeat to look good but it's very good at creating modern, wonderlust type interiors that look like they could be real. You'd think architecture and landscape could be very similar but from my creations at least the landscapes always look more fake than the interiors. -P

Yeah, it definitely has a problem with repeating patterns and very high res detail. Kinda ends up looking like a cartoon a lot of the time. I guess it prefers reading graphic novels to looking at fine art photography😂 @@FStoppers

I had grade school kids in this video! -P

At the moment, that might be the only reliable way to succeed at this. Once it fixes the eyes, it will be near impossible. -P

Pye scored the best of anyone. He missed only 5. Their was a junior high school kid who got 7 wrong and the girl who said she thought the guy looked hot (and was sweating) she did really well too with 9 missed. I think Pye just knew what to look for and he was the only photographer who I let take the test but that was just because he came to visit and I was curious. -P

Interesting, I partly agree since machine learning is nothing new. Machine learning used to identify and recreat. För 20 years ago I used it for identifying letters and words. Real AI is more like a virtual humanoid with a real "brain" . Think like getting a call from your mom and you can have a real conversation with creative input, not just historical flashbacks. So real and even up to date. So real that it's impossible to distinguish it in any part.

You might not have a choice in it...they may start using AI themselves and the panic starts slowly creeping in year after year. -P

@@FStoppers ur right, fascinating and scary times to live in. Let's see what it'll bring for us all

kind of like Santa...let's see if we were naughty or nice! -P

Your client's perspective: Looks like we can get rid of that guy with his camera and studio to save money now.

That’s a 68% pass rate. Maybe in college that would be a D+ but for a test like this I wouldn’t consider that acceptable. -P

@@FStoppers I stand offended; will certainly unsubscribe, and change thumbs up to thumbs down 😛

The end of this video says it all.

Apparently Tesla just erased all their self driving code they built for a decade for Ai created code and the self driving aspect works better than ever because of it. Pretty telling. -P

I did, it’s in the video. The AI was able to get about 90% correct. What that tells me is at some point we will use AI to validate truth and bias but then that AI will start to become pretty powerful. -P

@@FStoppers thank you! I must have missed that part!

It's not Natalie Portman at all! -P

How long do you stare at a billboard or tv advertisement? -P

@@FStoppers False equivalency, we are not evaluating if a billboard is real or not while driving

People like to say this but at some point when it has combined 20 things into a new transformative thing, I kind of think it has created it. I think about this a lot with music. Let's use guitar playing for example. If Eddie Van Halen was simply cutting up licks from Allan Holdsworth, Jimmy Page, Eric Clapton, Pete Townsend, and Jimi Hendrix then by your logic he didn't create magic, he just chopped up pre existing material and really didn't create anything new. We all know that to be false but AI is kind of doing that same thing. Humans have influences and borrow and steal the same way AI is but we are quick to say AI isn't creating anything, it's just referencing stuff. But so are humans right? -P

@@FStoppers I think you are missing what I'm saying. Let's use your music example we are not evaluating what is art, what your video asked if we can tell real from fake so, it will be like if we took samples from all of those musicians and repurpose then into a track where we ripped the direct recordings of each instrument then ask people if they think that that guitar solo is real? Do you know see how your video is structured? If Ai was told (or prompted) to draw a portrait of a human by description using only colors and it came back with those images THEN I would be impressed. But I'm not impressed when 100s of companies use literal billions of dollars 💲💲 over the better part of a decade where they stole already existing art and photos without the makers and creators permission, (Not to mention no compensation) cut it up and resold it as "Artifical intelligence". To me that isn't impressive 🤷‍♂️

I kind of get what you are saying but my worry is many people will take your view and say “oh well this isn’t real AI, or this isn’t general AI” and will scoff at this lesser version we are seeing now. Then suddenly, and I mean suddenly, the next version of AI will be so powerful that there won’t be any time to prepare for the human reaction to it. We should be preparing and perhaps even limiting it now before we see the really impressive stuff. -P

If you are scoring less than 50/50, you'd be better off flipping a coin to determine if the image is AI or real. It's like when they say a monkey could pick better stocks. -P

​@@FStoppers You could easily flip a coin and end up doing worse than those candidates who got 30-40% if the sample size isn't high, e.g. only 10-20 photos and only one round of guessing. If everyone really went through all of the 100 photos, then yeah, the coin flip would likely be be better.

The test was 44 images. It’s still 50/50 regardless of its 2 images or 1000. Yes the larger sample size allows it to even out but anything above 10 images is prob going to be close to 50/50. -P

@@Elias-nj6gi You are coming at it from a purely mathematical stance instead of Applied Maths. They're different

I think that day is approaching closely. -P

Also paying you 100s if not 1000s of dollars vs a free app on their phone is a pretty easy trade for most people to make. Even if the end result isn't the best. I've already seen Ai Christmas cards 🎄🤶🎅

I hope you are right but I’m still skeptical. If what you say are true, no one should pick an Ai image over a real image be it a headshot, landscape, or portrait session….yet in my tests they have done exactly that time and time again. -P

Those same imperfections was what led people to discover the AI images too. Once a person picked out something, halfway through that same criteria screwed them over. It was pretty funny to watch happen. It was eyes, then hair, then neck skin, then freckles, then ear piercings. Very few signs stayed consistent and AI can surely replicate any of those things well. -P

Thank you! Finally someone that understands that Ai is not doing anything special except theft on a massive scale. Imagine if those Ai companies had to pay licenses for each image that used to produce an image? They would've gone bankrupt before they even touched the ground. But they hid behind "Research Data" so they could say it was all for research and not for profit. Well OpenAi is now a for profit company and Microsoft, Google and the like are all gobbling it up and forcing it down our throats. if we do nothing they would've gotten away with it. Taking and synthesizing what makes us human. Disgusting!

I'm not so sure about that. 1) If AI uses the little (or large) thumbnails on say Google Images for it's input, how would you ever know your image was used in an AI derivative image? I have another video coming out showing these derivatives and not only can no one tell which is the non ai version, I doubt any photographer would see the AI version and think their image was the original. It's impossible to tell. 2) Copyright laws currently allow works of art to be altered greatly to become new derivatives. Richard Prince is a great example. Another great example is the Obama Hope image but Shepard Fairey settled that out of court so we aren't sure how it would have been ruled on. The point is that if an image is altered enough it can avoid copyright infringement. "Enough" in today's standards is pretty lenient and I'd argue AI derivatives are even more "transformative". I fear the argument you are trying to make is a lot like seeing Napster and thinking we can put the genie back in the bottle and the music industry will be saved. It won't be. Streaming is coming and artists will never be paid the same. With AI, it's even worse. -P

@@FStoppers In the short term you're correct but every action eventually spawns an equal and opposite reaction in society - it happens at many levels and often it arrives before we recognise it. Germaine Greer and other feminist writers take the world by storm and now we see the rise of the Manospere , Trump , Anti woke comedians and so on - action ----> reaction . AI is no different Great video BTW - loved it

Ai left by itself with no new information starts to go crazy and hallucinate. Its a problem that is incredibly hard to solve since Ai is treating each prompt as a problem to solve so it never knows what is good enough. It has no end goal so presented with the same human faces over and over again it will start to over correct its corrections and everything starts to fall apart. Thats why these companies need us to keep inputting more and more information en masse (Or better yet steal more and more) because without that new info their Ai's will start to lose it. So in essence the more we use Ai the more we make ourselves more irrelevant until most people have lost their jobs and society becomes a shell of what it formerly was

"I won't give them away" why? Also, why would anyone do a tutorial like that - you'll just take a halfway decent photo on any device and run it through an AI filter which will give you an endless amount of perfect headshots... unlikely at that stage this will involve any traditional photographer in the classic sense anymore for 90% of cases unfortunately.

Aren’t lens prices only going up each decade? -P

It will definitely replace some jobs faster than others. I can easily see how it will replace say the video editor but it's going to be a while before it can show up to your house and repair your cracked toilet. But yes, with enough time, barring any sort of global doomsday event spurred on by the uncertainty AI brings, AI will probably cause most or all jobs to cease to exist...eventually. -P

@@FStoppers Have you been following the progress of TESLA bot? I see little reason why in 10 years it won't be able to perform the vast majority of human labor.

@@johnzach2057 TESLA has a record of lets say... Fudging the data. For a true gauge look at Boston Dynamics which aren't very close to that

I think you are missing the point. This test was to show that normal people cannot tell the difference between real and Ai images. The first shoe to drop will most likely be in the advertising and marketing world. Why would any brand hire a photographer, model, makeup artist, creative director when AI can produce compelling photos like these? These headshots are very similar to beauty shots. Then there is editorial portraiture. Why hire a model for teen magazine or vogue when AI can make it for a tiny fraction of the cost? Yes real people will still want real photos of themselves but as this tech becomes increasing easy to use and realistic, I don’t doubt that many will choose to have an AI headshot made or make family photos themselves. Imagine throwing 20 snapshots off your phone into a program and it creates a near perfect group shot of 10 people with everyone looking accurate and happy. Most people HATE dressing up and dealing with a family portrait session during a vacation. Ai will no doubt be able to replicate that in a non cartoony way. I already have a few AI created holiday cards on my refrigerator. Makes me wonder if those families hired a photographer this year or they simply preferred the Ai generated image and mailed that out to everyone. It’s already happening! Normal portrait and headshot photographers prob have a decade or so until this is a reality but in the commercial, stock, editorial and advertising world, this is going to totally gut that sector. I predict in 5 years most magazines, blogs and advertisements will use a significant amount of Ai generated people. Even Hollywood is concerned about this. -P

Long live photography 👑

Everything is dead

Photography is very much alive

do you AI generate your wedding images or photograph them?