How to spot Generative AI? (even if it has all 10 fingers and toes)
Вставка
- Опубліковано 6 лис 2024
- I've seen quite a few "how to spot ai" videos that were simply talking about errors in the construction of the image.
Too many fingers, weird eyes, weird faces, etc...
But AIs are getting better;
And I'm not hearing anyone talk about the errors baked into the training data itself.
What makes an AI image stand out?
-The Blue channel on real Digital Photos will ALWAYS be more noisy than Red and Green channels.
-Consider the hardware used to take the "photo" or make the "digital painting"
-JPEG compression is image wide, chunks of slightly off tint/hue colors, with purple and green tendencies.
-JPEG compression should be more visible around curves in an image, they are formed from 8x8 pixel blocks, but may appear as lattices or grids within individual blocks (DCT Patterns). Sometimes standing out as 2x2 or 4x4 blocks because of the patterning.
-If it doesn't have BOTH JPEG issues, then its likely AI.
-So, non-square JPEG artifacts in PNGs with just random "JPEG Green & Purple" tinted pixels;
-Or odd noise patterns mixed throughout the image (per object noise patterns)
-Red, Green, & Blue channels all have the same noise means its a digital image, not specifically AI, but if its a "photo" with uniform channels, it's AI generated.
-Even "digitally painted" AI art styles will add inconsistent noise to every object in the scene, which an artist's general Style AND Approach/How to use the program wouldn't do; when using photoshop, procreate, gimp, etc.
-Lines crossing 64x64 pixel boundaries tend to not stay congruent. (12:04, as I mention it, look at the tree-ground line between the horse's legs on the right of the image)
-If you spot stochastic sampling (like pixel gaps between similar colors randomly in the render), I'd put my money on cgi render, not ai
Update : Given the comments, I've started to make an ai detector python program that can run windowed or from command-line, or basically, automated image checking of the issues I talk about in the video.
I'll update with a github link soon.
Note / Mess-ups -
I said Encryption when I'd mean Encoding, default JPEGs are Encoded, some reason my brain was just stuck on it.
5:00- Took the photo's last year, I just suck at keepin' my damn hands still.
6:00- I've been on sets and productions, but I'm a technical artist. I'm what comes after film or rendering or need to create pipeline tools to work with film or camera signals for work. Mostly writing code to work with images/video.
7:00-
The sharpening "appears" like JPEG compression, but the lack of the chunking I was talking about is the indication of "style" over "file compression" over "ai blatantly using random noise all over the place".
There is visible JPEG artifacts in the tan/beige area, but again, looks like "attempted style," not "JPEG compression"
7:15-
There is potential for straight line without jpeg compression, but you'd notice a blocky edge where the curved part meets the straight line. So, the area above what I have selected.
7:55-
Random JPEG noise but no distinct squares.
9:18- *Horizontal lines
I made the music in Aiva, and hold the license.
It's an ai music generator you can customize pretty deeply.
Please stop with the "encoding" comments, I get it... really....
My brain just never kerrr-chunk'ed to saying `enc-oding` from `enc-ryption` for jpegs
I appreciate it, but you don't need to explain it to tell me about it, please.
Reading the comments used to be fun, now its a haunting red dot.
Don't feed the haters bro. Just ignore them and they'll go find something better to do.
@@DragnEYE Oh, its not hate really. It is a valid error I made in speaking; but its not helpful at this point, and becoming quite bothersome.
@@TrancorWD understandable homie, just fuck em' tho. turn up litty. the video was great.
I’m sure it will sync in now. If I had this happen to me for a technical term I would learn it PERMANENTLY. Trauma is the best teacher 😅
haunting red dot sounds fun
cod any one?
As the old meme goes "I can tell by the pixels that it's fake"
We went from old sayings to old memes
@@1x5x7 lol
As foretold in the ancient texts.
Age old adage, "check the pix[els]"
I can tell by the pixels, and from seeing quite a few -shoops- ai's in my time
Harkens back to the old 4chan days of “This is shopped, I can tell by the pixels, having seen many pixels in my time”.
never gets old, images will be always made of pixels.
@@SilverSpoon_unless it's a vecor
Then il Screenshot the vectorgraphic and get pixels
@@Voyajer. To view a vector, it still has to be rasterized.
@@Xetttt we will also not see a vector of a human face. Your computer would blow or something
Nice to have a forensic method rather than just noticing everything is vibes not reason.
We do have an ‘inner method’, call it likeness, I’ll probably make a video on that later. But all you need to know is that when you feel something is off, it’s not just about the uncanny valley vibe, but rather because it’s different from a lot of things you’ve seen before. If you look at people’ faces a lot, you’ll not notice the asymmetry of their faces because you’ve seen so many faces like that, but the second you flip that face in a picture, you’ll notice how asymmetrical it is. I call it ‘the contrast theory’. Any imagery or audio that happen to contrast every ounce of information you’ve already learned will be noted by your mind automatically.
Even if a 5 year old drew their parents, you can still tell it’s something from a human being just simply because the drawing ‘makes sense’, it doesn’t violate any of your memory.
To be fair, AI image gens are pretty good at getting lighting right enough and sometimes even structure, but it doesn’t understand weight or physics the same way we do, whenever you see an AI generated image, you can tell that the picture has no element of physics in its depiction, you don’t need to actively seek it out. It just goes against everything you’ve learned passively and that’s a huge problem
LLM can't understand concept is a silver lining we will always have against it no matter how much training it has. So feeling an image is "off" is always good sign to look closer to see if it's AI generate.
999
You call it vibes, i call it heuristics
Never before have I thought "thank God for jpeg artifacts"
I love jpeg artifacts
same reason people like 8 bit even though its objectively awful sound variety wise
@@Forcoy Or how vinyl records have better mixing than CDs due to the limitations of the medium.
@@jamesphillips2285 Any limitation as soon as it can be avoided will have people tripping over themselves to emulate it 5 years later
ive intentionally added JPEG artifacts on multiple occasions there quite cool, might even be my favorite photo file type
So this is why they're trying to kill jpeg
there is a "wet plastic" texture that all ai images have to some extent, it's a very vague texture but it's noticeable and i can usually tell ai images from a glance due to it
YEAH I've heard it described as "creamy" which is exactly it. I've still been fooled before despite that
I find AI 'photography' outputs to be generally hard to look at because they have so much focus and smooth detail on the bits the AI can grasp (faces, eyes sometimes, skin without pores, fur on an animal's face) but blur and smudge everything that AI cannot grasp (edges between subject and background, edges of portions of the body and face, overlapping corners of subject and background or subject over subject). It makes my eyes strain for some reason. I hate it.
The ai sheen
I thought this was deliberate so that people can't make convincing deep fakes. I remember when DALLE3 was first introduced, it could generate very realistic images to the point that you couldn't tell it was AI other than by the 1024x1024 image size. Hell, I think they banned "DSLR camera" in the prompt because that was one of the ways to generate those photorealistic pictures.
That's largely dependent on the model used. I've seen some photorealism ones on the Stable Diffusion subreddit recently that can't really be identified as AI by that quick "vibe check". But most of the images you see on social media and whatnot are very plastic-y looking
i love how his taskbar is full of work programs (Photoshop, VSCode, GitHub, OBS, etc), but also has minecraft at hand for whenever he needs to play some survival
:D Hey, I got a 8 month old hardcore world goin strong!
... Is now a bad time to promote an optifine shader pack I wrote called `procPromo` on my ProcStack github? haha
Bro has good taste!
@@TrancorWDdefinitely not a bad time haha. You gotta promote yourself somehow!
@@TrancorWD W hardcore player
average autistic taskbar
AI model: "Do I look like I know what a jpeg is? I was just told to make a picture of a god-dang hot dog."
* epic club house remix *
This is an awesome reference
IM NOT THE ONLY ONE THAT THOUGHT OF THAT
"I'll tell you hwhat!"
@@idontwantobefishingforfish as the song goes:
J- peg... jpeg J- peg... jpeg
ai spotting went from "it barely looks like an object" to hyperspecific analysis
the cartoony dall e squares that looked like abstract shapes
And it is all going down from here in the next 1+ years
@@alface935 If by down you mean up then I agree :) time for the middle class to know what it means to have their jobs replaced by automation!
@@tsjbb It is all going down from here = It is going to get worst from here and foward
@tsjbb ...no...ai piqued in college xD this is all it will ever be unless someone starts all over again and,makes something completely different but starting from the lessons ai thought us. It's like saying ok, in this bowl im throwing ingredients for cookie dough, it tasty all the way up to the baked cookie, but it's imperfect cuz you're not a masterchef, you just followed instructions. And now you wanna try and improve your cookies...by grinding them down again and adding more of the same ingredients and baking again- you'll just get the same cookie as before + burnt chunks and if you keep doing that you're just gonna end up with a burnt mess
huh it makes sense that it would learn the artifacts but i never thought of that before! that's interesting. I also noticed that usually, AI generated pictures have really high sharpness, especially on faces and folds
Fair call! A lot of people use GFPGAN to specifically fix ai faces, it unwraps your face flat, then sharpens it up (And usually makes the face look younger) and wraps the 'fixed' face back.
You can tell if they are using it by seeing sharper hair around the face, but not on the rest of the head.
Edit: I said ESRGAN initially, that is an image upresser AI, not face fixer; and also a second ai to use.
@@TrancorWD oh wow ok, thanks for the information! What is this erdgan thing? Is it also machine learning?
@@pepn it's a secondary ai you run after you generate your image. I think it was initially designed by Tencent in China to help "enhance...enhance" security footage.
But it is free to use, and many do use it
Many ai generators have it already included or make it real easy to install it.
@@TrancorWD I tried looking it up and thought "hey this seems kinda cool to fix up blurry photos, maybe gen AI finally have some good use cases other than killing out art". Then I zoom in. It was pretty bad. I tried the one on Replicate and it made the cat's whiskers look like tendrils, it was so uncanny.
@@TrancorWD Okay. Thanks for the valuable information, in the video and here in the comments ! Cheers
Just like phishing attacks, the techniques will continue to improve and become more common, making them more difficult and time-consuming to detect. The best method to "stay safe" is to consider whether the source is trustworthy and what they are trying to tell you.
Having said that, for the forensics nerds who can work to evaluate whether an image is AI or not even with minimal context... this is pretty cool.
Phishing attack techniques haven't really advanced or improved at all. If anything they intentionally remain flawed so they can specifically target people who don't pay attention. People who are too wise to fool aren't worth the effort when there's someone else that is easy to fool.
@@PURENTmost attacks work this way, but Spearphishing can be very complex.
In this case it's different because of how generative models work fundamentally, I don't really see a feasible way to exclude the anomalous artifacts form the model, the problem with machine learning is that you can't really control how the model arrives to a conclusion, you can only control the training set to hope for a desired outcome, but since all models have been trained with JPEGs and PNGs they have included the artifacts as part of the pattern, can't really reverse that back, the only way is to start training from scratch using only a training set of PNGs and making sure there aren't any screen captures of JPEGs. It's not impossible but you would limit your dataset significantly and it would be way too much work to vet every single image in your training set, definitively harder than blindly just scraping the web, it would not be worth it
@@QualityDoggo Even in spear phishing, they always look for the weakest point. Which usually means someone gullible.
Artifacts are something tooo niche. They're not detectable by a wide public, so no need to get rid of them, nobody with money will see the difference. As were mentioned by Lambda_Ovine, removing artifacts demands building a training database from zero with a careful selection of pictures. PNGs aren't a popular internet format, so pngs are lying on disks of artists, so if somebody want them, they should negotiate with said artists. And I'm not sure if you can find many pngs in photography. Most cameras save everything in jpeg. I don't know if raw is even useful for AI, but not many photographers know how to dig out this raw from their devices. And I mean photographers in general, everyone who's making any photos. Oh, and funny thing. Many modern smartphones have automatic post processing of photos, which you can't turn off . So we feed a lot of extra artifacts already
Awesome video! I'd like to make some comments regarding JPEG compression that may make it easier for people to spot:
1. JPEG uses a grid of 8x8 blocks of pixels almost all the time, though you can occasionally find pictures encoded with 16x16 pixel blocks. Like you said, the borders between two blocks will be distinguished by occasional sharp, high contrast differences between pixel values in one or more color channels. You drew a lot of selection squares that weren't 8x8 pixels while explaining, so I thought it was worth clarifying.
2. The top left corner of an image will always lie on the first 8x8 block boundary, so you can use that to find all of the block borders on the image. Of course, not all images have a resolution that's divisible by 8. In those cases, the bottom row and rightmost column of blocks are actually stored in the image data, but they are partially cropped out when the image is displayed. If you overlay an 8x8 grid on top of a picture with JPEG compression and the edges of the blocks don't line up with the grid, then it's almost certainly AI!
Well, if the photo has been manipulated or the artwork is a photobash, it may have block boundaries moved if the source photos themselves ended up being scaled and moved. But still, you'll see fairly consistent blocks in one area, and the particularly noticeable artifacts should be clearly inside one square or show a boundary between squares.
I'm not sure if I fully understand video or this comment, I'm not familiar with graphics technologies. Like what if there 2 vastly different colours in that 8x8 square?
or cropped, but yeah I get your point
@@realdragon In the case of an image having two contrasting colors that divide a column of blocks in half, it still wouldn't be perfectly vertically sharp. Only the boundaries between the blocks are able to be perfectly sharp.
The reason for this is that each block doesn't actually store its pixel values directly. The colors in a JPEG are actually made by taking the discrete cosine transform of the blocks rows/columns of pixels in the original image and, once they are in the frequency domain, discarding the highest frequency signals. Because it gets rid of high frequency signals, each block can only show a 'blurred' version of the original values that its region contained before compression.
If you've ever saved a screenshot with text on it as a JPEG and zoomed in to notice weird banding, noise, and artifacts around the text, that is why. JPEGs just don't do sharp, geometrical shapes like text well due to how the format was designed.
I know that might sound confusing, but there are some really good videos on UA-cam explaining JPEG compression with visuals if you're curious.
@@monkeysfromvenus So each 8x8 block is just Fourier transformation? Yea that make sense, thanks
My assumption for why the noise patterns are consistent is that AI generated images are created purely /from/ noise.
AI models are given a pattern of pure noise and told "There is a picture of a horse here, please denoise it until the horse is visible." and it shuffles the pixel values around until it resembles something marked "horse" in the training data.
What you end up looking at isn't a noisy picture of a horse, it's a horsey picture of noise.
„A horsey picture of noise“ is possible the best explanation of anything I’ve ever heard
This is the best explanation I've seen, thank you 🤣
That’s what stable diffusion is.
@@ThatEverydayEnthusiast Yup, I know - I was just explaining the actual mechanism behind the denoising and how that would result in the abnormal-for-real-images consistent noise patterns.
And majority of people (even artists) think it's all stitched collaged shit which isn't even true
Misinformation should be stopped
Great video! I had never realized that compression artifacts show up in AI-generated images, that's interesting.
Thats what i hate about ai art the most if you look close you just see camera like compresion artifacts on a waifu painting.
everything shows... thats why its important to use consistent datasets when training your model
I kind of noticed them already, but the smear/shine was a greater tellsign if something was AI or not.
It's like a cargo cult. Imitation without understanding.
compression artifacts are not due to the output by the ai generation but mostly by the renderer of the software the ai generator uses (aka browser).
People have been extruding raw image data from the ai generation directly from the GPU with 0 artifacts but this can damage the gpu itself.
"The model was trained on errors"
Holy shit. That's... kind of a significant oversight when you think about it.
AI art is all oversight. Like how millions and millions of dollars are being funneled into this technology, that copyright and law just has a caught up with, and then when people ask copyright and law questions the big people making all the money at the top of it just wave their hands and say, "Don't worry about that."
Yeah. Definitely forward thinking technology, right?
It wouldn't be if the samples were properly tagged. That is, you want the model to be informed "this sample has jpeg errors on it", but also have samples without it. This way it can utilize jpeg samples, but also learn how to not commit jpeg errors.
JPEG compression has been so widespread for such a long time that people are trained to expect JPEG artifacts.
The same as how 24 frames per second is said to give something a 'film quality'. Not because it's better than 48 or 60 fps, but because people have been trained on seeing films at 24 fps in movie theatres.
it depends which images or videos the model is trained on, among other things.
the people who originally designed these models aren't villains rubbing their hands together trying to create the Perfect Copycat, they were computer scientists interested in how well an AI could replicate realism. bar monetary incentive when it took off, they had no reason to want images that are absolutely indistinguishable from genuine counterparts.
This makes so much sense, and kind of explains that feeling you get when nothing is technically wrong but you can still tell it’s Ai generated, the noise pattern specifically really stood out for me.
No one is noticing a noise pattern without looking for it, consciously or otherwise.
@@Razumenyou kinda do, because you do notice if the patterns are inconsistent. It creates a weird fizzyness
@@deusex9731 You literally can't see it without an image editor.
@@Razumen Brain is strong with this one, but the mind is weak.
@@Razumen I gotta disagree, since these kinds of subtle effects are part of my post processing, adding and modifying noise, doing an emboss/ high pass filter that’s has very slight hue differences and like a 1-2 pixel offset adds noticeable sharpness, which when done intentionally can be controlled to draw attention to certain points. The Ai version is very thoughtless and random, but you do notice these things, color channels and stuff only serve to make to make things more obvious.
As an artist I identify AI art by the sheer vibes, but having a more scientific approach for it is certainly going to come in handy
especially when other people are like “erm how do you know it’s ai? it looks perfect to me” and i’m like “idk! i cAN JUST TELL!!1!1”
For me the tried and true method is just looking at an artist's portfolio. If something seems off it takes about a minute going through the rest of their work to see if it's because of their style and skill level or if they're generating stuff.
@@user-xsn5ozskwgAgreed!! if their entire profile is just the same thing that already felt like ai posted 30 times with only slight variations thats a good tell,, i see it on pinterest often
@@pancakestrouble2429yeah and those I can just tell people started accusing actual artists as being Ai. The i can just tell crowd are the worse.
I’m an artist and the way I tell is:
Overall style (a lot of ai art has an obvious style) -> proportions -> hands -> coherency (do patterns match up? Does hair disappear into nothingness?) -> small details (ai has a hard time making minute patterns consistent, like detailed eyes) -> image size ( ai has a very hard time above 1k pixels)
The hair's often too manicured - no flyaways or baby hairs, all waves make a perfectly satisfying S-curve. Looks like an old illustrated magazine cover come to life.
Alternatively, it does have flyaways, but they're weird, puffy masses instead of anything that behaves like actual hair.
the thing is you can 100% have perfect hair irl
@@neatoburrito3170 not the best description since that just sounds like normal hair to me but you can most definitely replicate "weird puffy masses" in real life
@@TheParklifeChoseMeHow tho, please teach me
@@mo-s- backcombing can give you a weird puffy mass of hair, Google Simon Gallup (back when he was younger), I'm obsessed with his poofy hair
Great video - thanks for sharing with the community! A few minor corrections/notes:
a) You're using the term "encryption" in place of "encoding".
b) The term you're looking for to refer to the JPEG square is a macroblock (usually 16x16).
c) The green/purple colouring at edges is called "chromatic abberation" (CA), and it is NOT a JPEG compression artifact - it's an optical lens artifact. Seeing it in itself doesn't suggest generative AI. However, seeing it in very discreet portions of the image, but not in others (and not having a gradual transition in the extent of CA between areas in the image) suggests that the image is a composite.
d) Just because the image is composited, it doesn't necessarily prove that it was created via generative AI. That is, it can be generated by a human, not necessarily by AI. Additional nuance is necessary to make the distinction, but my only point here is that signs of compositing can only very roughly be inferred to suggest generative AI, and it should not be used as a hard rule of thumb. (I'm not suggesting that you were necessarily blurring the distinction, just think that it's a point worth noting for everyone.)
Thanks for the input!
I did want to mention you can get the green&purple artifacts from jpeg compression in addition to chromatic aberration from lensing effects on a digital camera.
I'm talking about the jpeg artifacts that look like "deep fried memes".
It's from the random noise it puts into the compression to reduce the amount of data between near by blocks, to reduce file size.
It appears like DCT Patterns, being grids, lattices, or other ways of mapping the colors to be "similar" to others to reduce file size.
It's not solely an optical effect of Chromatic Aberration, likely it will be side by side with it in a digital photo.
I mentioned (D) at the end of my video after talking about pre-multiplication, matting, comping, and background removal having tons of ways to do it by hand. That it isn't an indicator of AI, but that it stands out if someone says "look at this photo"
Don't forget that chromatic abberation is often an intentional effect some artists add on top of their art, it's been popular for a few years now. Though in those cases it's often visible from afar and not only zoomed in.
the "encryption" thing in particular is jarring, it's a simple mix-up of similar words but it sounds extremely weird to people familiar with the difference and potentially has a disproportional impact on credibility. not worth throwing the video out over, just wanted to emphasize it as important for future videos
tbh there's a number off ways you can get the green/purple colouring, it's not actually cut and dry. it can be an artifact of certain types of image or video encoding with different colour channels or the incorrect way colour mixing is usually handled for eg
One thing that color issues at edges can come from in compressed images is Chroma Subsampling - our eyes suck at color so 1/2 or even 1/4 the chroma resolution with full luma resolution can, overall, look better than you’d think.
JPEG images frequently (but not always) use it. (And basically every video outside of production had quarter res chroma)
i always tell its AI by the art style and lighting. AI almost exclusively has an over correct composition, sharp contrast of colors , intense dreamy lighting. its like that episode of rick and morty where the Aliens could only manufacture positive memories. it rarely deviates from that look
AND something to consider from a 3D artist. many artists that like realism will intentionally add fake noise or film-grain to their renders. so dont go around harassing 3D Artists for having noise in the art.
Unlike the rick and morty episode, it *can* deviate from that look with a bit more effort, especially if the work is AI-assisted rather than purely AI (think someone using img2img to finish rendering a draft). So while these things are often good indicators a work might be AI-generated, their absence doesn't necessarily mean they aren't.
Yes, they add noise, but will it follow the same patterns as AI noise? That seems improbable.
@@petersmythe6462 it can
since the noise is artificial the pattern will stay the same across RGB channels
@@feffy380 AI users by the philosophy of this tech. dont usually put in that kind of effort. very few of them really try. the rest of them like to spam the same thing everywhere or scam people with minimal effort.
The lighting in an Ai picture is probably the easiest indicator for me. Either that, or how overly glossy and ‘goopy’ it looks. I hope this doesn’t cause any actual artists to become harassed though. This Ai art thing being pushed around really shows how little they care about us artists.
AI images make me feel so queasy its like someone took a screenshot of an uncanny nightmare
‘Can kind of tell this is fake’ those horses proportions are HORRIFIC. I would never think that was even slightly real LMAO.
Edit: it’s utterly tedious replying to a bunch of comments who havn’t even read the comments before them, I won’t be replying to anymore replies to this comment
I noticed the weird ass background in the left first
@sociallyineptsnapper my eyes were drawn straight to the middle horse with the giraffe neck.
These techniques are going to he necessary however as the technology improves...
I just noticed the colors were totally wrong. Super unnatural color, like Legos or something.
@@Gandhi_Physiqueyou can put photos through a filter for that effect.
like why does the middle one have boobs
Ai learning defects will never not be funny to me, because when people who can't tell the difference between ai generated and real images start to teach their model on ai Images they've made, that have saturated their sources of authentic media, the ai is probably going to start basing their future generations on stuff it may have already made, likely exacerbating it's issues if there were no preventative measures already in place.
Almost like downloading an image over and over again.
I think it'd be interesting if there was a model made to do just that, generating images it learned from other ai images, just to see the results of it, y'know?
But what do I know really?
isnt that already happening with text bots? they made so much crap, that new bots that are scraping for data get a bunch of already bot made data
See, but that would yield actual interesting results and they don't want that, they want *content*.
good , i hope the ai decay continues so people can stop polluting the internet with their overhyped aibro bs and they get nothing of quality
One of the key steps in the quality of modern text-to-image systems is literally to train it on its own data, just filtered somewhat for quality.
I'm pretty sure this concept actually has a name, model collapse. I don't know where I saw it but I remember seeing someone inputting a generated image back into the AI, and within literally one iteration it already obviously looks worse, and by 4-5 times it's virtually unrecognizable. With the amount of AI work being posted online, then AI scraping those sites for training data, then generating more stuff that's posted, I think it's actually going to start being a huge problem.
JPEG and PNG aren't "encryption" formats.
JPEG macroblocks are always 8x8 pixels however the chrominance chanels are usually subsampled so they would be 16x16.
Very informative video overall :)
He probably meant to say encoding.
Yeah, encryption and encoding, I mix em up a bit since some formats have encryption, but you're right, it is encoding
Macroblocks! Thanks kindly, wasn't familiar with the term, just it's behavior.
That was kind of what I meant by only being familiar with them being set blocks.
But the artifacts generated seem grid like at 2x2 or 4x4 at times visually, I just assumed its the encoding algo
looking into it some more says newer macroblocks are 2x2 and 4x4 for h.263/h264 encoding, which I guess makes sense for how data moshing video looks.
There are other effects that recur in 2x2, like Bayer filters and 4:2:0 chroma subsampling (also present in jpeg). Interlacing affects every other row (but doesn't randomly go vertical). Other formats may have other subdivisions, including JPEG 2000.
Looking at the technology of the training images is a good tactic. That's something you can't just get out of the ai. The better the model gets, the larger the training data gets, and thus they have less chance to be choosy with the images.
Don't forget that AI generated images can actually poison the data set
1:44 that's impressive; never knew you could 'encrypt' in jpeg
In regards to that other comment, I understood everything you said perfectly fine! I'm not even sure how you could explain it more effectively... Very informative video, I had no clue there were such decisive methods of detecting AI that are _mostly_ invisible to the average person. Thanks a ton for making this.
And for the record, I may be biased but I like more unscripted content. It's offers more emotion and feels a lot more welcoming than some of the stiffness you see with more scripted creators. Its definitely a skill to master, but it already seems that you have experience with how articulate you are.
Thanks kindly! I do like seeing that in other video makers too, that off the cuff kinda emotional side. But I definitely miss spoke slightly at times.
Nothing wrong wrong (7:07 is bordering on wrong tho), but I should have re-shot some sections in here.
Is that Reg in your pfp?
@@oxfordcommaisthegreatest perhaps...
@@dovahfruit9503 Nice! I love that anime
A lot of not entirely accurate information in this video:
1. CCD sensors haven't been used in phones since 2006 and in DSLRs since 2010. CCDs are still in limited use today but it's mostly for scientific applications - like astrophotography cameras, microscopes, x-rays etc. The amount of training data that comes from CCDs is likely so minuscule it has little to no effect on the output.
2. Different noise patterns between CCD and CMOS sensors is not a unique property - noise patterns also differ between different CMOS sensors. Sensors can be different physical size, have different amounts of pixels on them, they can have different bayer patterns, or same patterns but with different, more or less efficient absorption characteristics or lack a bayer filter in the first place as well as numerous other differences that will have an effect on noise. Furthermore noise isn't only generated in the sensor itself - elements like the ADC will generate their own noise and add it to the image.
3. A lot of training data doesn't have inherent noise at all since models are also trained on digital artworks, not just photos. Furthermore there are numerous different models and I know of several that are taking measures to eliminate jpeg artifacts from their training data (either by simply rejecting images or by using some upscaling/denoising software to deal with it). So while the method you're employing might work on the model you're using, it might not work on others.
4. Black borders like those you've shown around the snow are possible in real images and are most commonly a result of overzealous sharpening algorithms in some consumer cameras (can produce white borders as well).
5. Scrutinizing pixels and photo artifacts like that requires a lot of knowledge and will lead to a lot of false positives when done by people unfamiliar with how cameras and compression algorithms work. Similarly how a lot of the UFO footage features ordinary things that people mistook for UFOs due to their lack of knowledge how cameras work - great example is that one video taken with some night vision camera that happened to have a triangular aperture which made the bokeh appear also triangular - then an out of focus plane in the sky looked like a triangle in the footage which people mistook for a UFO. Lens flares and other lens reflections have also been commonly mistaken for UFOs in ordinary footage.
6. I really like the method of comparing noise patterns across different color channels, that being said again that will only work against images pretending to be photos and not digital artworks and also I would be cautious because it's possible some models might be able to reproduce this effect as well.
7. So how do we spot fakes? Well for starters there are the obvious inconsistencies like fingers and toes, but there are other, less pronounced inconsistencies that you can notice when looking a bit more closely - like weird asymmetry, incorrect or distorted anatomical features, small background objects appearing unclear and/or distorted. Nonsensical or jumbled text. And finally the overall feel - a lot of the time, at least for now, the AI images still have that distinct AI feel which you can learn to recognize - although be mindful it's not always 100% accurate.
But unfortunately sometimes you just won't be able to tell and that's a scary thing. I suspect a AI image detectors will become more common as times goes on and we're gonna see an arms race between AI image generators and AI image detectors in the future.
Yeah, this video is literally just the technobabble version of "it's AI because it looks weird to me". Tons of assumptions with little actual facts to back it up. Nothing he touched on in this video applies to any of my AI-generated images.
Another inconsistency to look for is continuity errors between intersecting objects, such as a tail passing behind legs, or the background being cut in half by the subject - AI struggles with knowing where the other end of the object should come out at and what it should look like. So for example, a table behind the subject will come out the other end too high or too low, or a different color, pattern, shape, etc. It's a dead giveaway because it's not something a human artist would ever struggle with, since artists have layers to work with, but the AI doesn't.
This a great comment. I zeroed in on the first item you stated. The fact that we never had a single iPhone model with a CCD sensor should be a pretty good indication that the majority of the training data is on photos from CMOS sensor cameras. Even in the SLR world, most of the prosumer cameras moved away from CCD in 2006 (and even earlier for Canon it seems).
Also, noise pattern peeping across channels can result fairly easily in false positives, as it is rather common to apply noise reduction much more aggressively to chroma (colour) than luma (luminosity), which yields very similar looking noise patterns across the channels, as colour noise is usually visually much more offensive than luminosity noise.
Very cool, unfortunately I didn’t understand a single thing you said
Hmm, maybe I should have written up a script or something rather than rambling off the cuff like this.
I guess I'm mostly just saying "Random green & purple pixels that don't look boxy/gridded, it's likely AI generated"
Thanks for the feedback!
@TrancorWD mostly it would have helped to be a bit more explicit in what you were doing. I was mostly following but I see where OP was coming from.
There were some steps before you pressed record, and some terminology used, that might seem obvious to you, but are not so obvious to the people this video will reach.
I think this was a pretty good explanation. JPEGs has been with us for all of digital photography & web browsers, PNGs for half that time.
If we as “creators” or just “users” can’t know how our everyday tools work, and most spend 13+ years at school - is that _really_ education? Or just “work training”…? Shouldn’t education be about all of life - including the tech we use 24/7…?
Yep, questions. There are answers too - not pixelated at all, go find’em... 😅
1) if noise on every channel is the same, it's ai. 2) if image have jpeg artifacts (squared boxes of pixels and colors where it shouldn't be) then it's ai.
@@FabbrizioPlays fair call on explaining specific terms and better. I get what Musiqtee is sayin, but I could help provide some of that education in those cases. Thanks kindly!
One thing I have a concern about, we don’t know how many times that image has been converted by default from one format to another (jpeg-> png, vise versa) by social medias unless it’s verified that the post you sourced from is the original image in any lossless format.
However there is one thing that can help with this: The generated image *still* made fake artifacts no matter the original format. if you carefully extract the compression artifacts from the generated artifacts, this technique still can apply. (Though with a lesser level of confidence)
Great video, thank you for your insight. Amazing as always to see technical artists going into detail about interesting topics
I think, it's a very useful video, if you're checking not the authenticity of a random picture in internet, but a picture you ordered from somebody or their portfolio
in 5 years we're gonna have to calculate quadratics to make sure that it isnt AI 😭😭😭
uhuhu
Highly doubt that. Generative ai has barely improved in the past year.
@@IHaveALotOfBraincells Thats definitely untrue. The main improvement thats happened is speed and price, but its still improved
i mean those can easily be done computationally and in 5 years time im sure therell be a program that can detect this kinda thing effectively
@@thatoneguy9582 But how many will use it? It needs to be integrated into image and social media sites to detect automatically, with a disclaimer that says "This image is likely to be generated by AI" or something. And will that be cheap enough for sites to integrate?
Something that typically sets me off instantly is faces having a lot of sheen
This made so much sense and I LOVE how you describe things
AI always has kind of a wrong look even if nothing in the image looks theoretically "wrong" aka no messed up eyes or hands
and I never realized its because the error patterns its generating are just incorrect in general!! Thanks for the video
Always remember: AI emulates pictures. It has no clue, how reality looks like. The lighting will be off. Grain and noise will look weird. There will be stuff, that just doesn't work. And: AI often repeats. You see the same mistakes again and again. Weird artifacts, lines that don't work, strange proportions. A fast look can deceive the eye. But the more you look at an AI-pic, the more you see the flaws and mistakes and they will stay. No matter, what the Techbros want to sell you, AI will not develop fast enough, to understand reality.
I see this as a new field in photography. Photoforesics. The job, that reveals AI-Fakes for fact checking and stuff. Wait for it. This will be a thing in the future, because of all the Fakenews.
If you're watching this the first time, basically, zoom into the pixels and read for the unrealistic/extreme/unnatural noise patterns. They tend to happen even in the best of models. You usually don't even need to zoom into pixels to spot weird patterns.
Awesome video. Learned something new not only about AI but also about JPEGs. Never knew that they almost always consist of 8x8 blocks. For some reason I thought that the compression algorithm used in them was a bit smarter than that.
People trying to show off in the comments by pointing out the difference between encryption and encoding are kind of annoying though
This really depends on the type of artifacting. some image filters and blurring/smearing tools can cause similar effects to what's seen in your second example, though in a more consistent way. Asides from that, most of these are a pretty accurate way of telling
Edit: this also varies depending on what dataset the ai is using. Channel noise is generally a more consistent tell
I can’t wait until the internet gets so saturated by AI images that models get trained on AI images. It’ll either cause a feedback loop where all slight mistakes are repeated until the AI deteriorates, or the developers will be forced to find a way to differentiate AI and real images so their AI is trained only on real images. Either way it’s gonna help tell real from AI apart.
Ofc people that train the AIs are to stoopid to notice that
When it comes to drawings, the main things that I notice are when lines are depicted that an artist wouldn't draw, when the art style isn't consistent, artifacts, etc.
stable diffusion when unstable infusion walks in
It's cool to know some more technical details about why I can look at an AI image and think it's an AI image despite it looking pretty damn well constructed.
I often can tell at a glance even on profile pictures, usually something about lighting/shading or just the overall image that I can’t point my finger on
Clearly didn't watch this video...
7:04 I feel it should be noted that vertical and horizontally aligned edges generate different errors - like you see in the example image - because the waveforms used to make these edges will be perfectly aligned
This is really fascinating. I've never seen anyone point this stuff out before. Other people do this terrible thing where they describe mistakes a beginner who actually was trying could have made, which makes more people demonize beginner artists. I appreciate what I learned here today.
It would be great if a checking-website is made by implementing this method for general users to quickly find out if it's AI or not.
Also, great video.
Btw virtually no modern device uses a CCD, it is almost entirely only CMOS sensors, unless you are taking some pretty retro photos
This is such a fantastic video. I now regularly sit and wonder about certain images being AI generated because, even as someone with a passive interest in photography/graphic imagery or whatever, I have to think to myself "is this even a real photo?". I haven't messed with photoshop in a while but now I'll be doing a deep dive on every questionable image i come across. Thanks for taking the time to do this.
Whoa, 3 horses standing infront of powdered sugar???
it's common in equestria. but that's not sugar, they need help.
@@SilverSpoon_ I shoulda renamed the file at least.
@@SilverSpoon_Pinkie pie's Sugar stash
Please stop saying encryption I am sobbing 😭
Besides that great video, nice to see someone delving into this a bit
I was always able to subconsciously feel like something was wrong in an AI-generated image but I would never be able to tell why, now I know! This guide is extremely valuable, thank you for making this!
I doubt this is the main reason - such artifacts are often really subtle. I think that AI just has look to it - some subtle stuff that of course can exist on manmade pictures as well, but are much more common in AI. Perhaps because AI is often used for stuff that is hard to make with other means, like some photorealistic and impossible landscapes - I doubt that improvements in AI will make this AI look and feel entirely disappear
The striped color pattern was the result of a buggy VAE that shipped with SDXL, that VAE was since replaced for a different one that does not generate the striped color pattern.
This was incredibly fascinating, somehow I never consciously thought about AI replicating flaws like jpeg compression... but obviously it would, because it was fed low quality training data.
God please dont let the ai see this
The reason for the weird horizontal banding might have something to do with how the model works. All a GAN is doing is composing functions, either linear, trigonometric or gaussian in nature. Because those functions are continuous, at small scales the individual functions show through, causing the strange noise patterns and banding artifacts.
From my own analysis, if you use the sharpen filter a couple of times, you can start to see strange artifacts.
EDIT: duplicate the image three times. Use the curves modifier to remove all but one color channel from each image. Set the three altered images to the subtract blend mode, then merge them with the original. You should have a (seemingly) black image. Go back into the level curves and ramp up the very left side of the gamma chart until you start to see bands of color emerge. These are the functions. Look for the bands in areas with little detail, or areas all of one color. That’s where composing smaller waves would add unwanted detail. The more detailed areas show a noise texture, but that’s because the waves there are smaller than the width of a pixel.
Someone is already training a model to address these shortcomings.
Good detective work. It's remarkably easy to obfuscate generative artefacts like you're demonstrating, especially because they are high frequency in nature. For those not applying that extra layer of complexity, I agree mostly with what you're presenting as a detection methodology for this class of artefact. I work in the field of image forensics.
ai after this video: "ok dude imma fix those then."
jpeg does a couple of things that causes blockyness:
1. chroma subsampling: the RGB values are converted to YUV color space (one brightness channel and two chroma channels U and V). These U and V channels are stored at half resolution because our eyes are worse at seeing differences in color than those in brightness.
2. discrete cosine transform (DCT): each channel will be spatially compressed by taking each 8x8 region, applying a DCT (basically a fourier transform but different), and removing high frequency components. since each 8x8 region is compressed in isolation, seams can start to appear
And yes, for the nerds out there: i am intentionally leaving 4:2:2 and 4:4:4 subsampling out of the picture for simplicity, and they aren't very common to begin with
I can only spot AI generated images because I spent ungodly amount of time on artstation using those Chinese artist's posts as reference around 2015-2019. My problem now is pinning down the originator of the ai diffusion default woman face. It has to come from artstation or several thousand gravure faces. that woman face is so noticeable.
FYI, JPEG & PNG are encodings; encryption is a whole different thing.
Very nice video, that said.
The noise pattern was something I didn't think about before. Great tip!! Would this tip also apply to computer generated images such as digital art/renders?
Generally speaking, yes, but it's tougher to spot.
In cgi renders the RGB color channels will look like the 3 horses in my video. The same shared noise pattern on each color channel.
So if something looks like "pixar style" it will be a lot harder to tell.
But you'll be able to spot the weird JPEG compression colors not in a grid.
It would also apply to genuine compression artifacts, like AV1 video noise resynthesis. The telltale is not the existence of artifacts but the incongruence.
The middle horse is closer to a giraffe, than a horse😂
You also get identical noise patterns on 3D renders, so that's not necessarily an indicator of generative AI.
They are similar, but not quite still.
Any ray tracers or photon based renderers, the noise will at least use the same colors from elsewhere in the scene, ai noise is usually a Gaussian 3d noise or Perlin random color noise.
You wont see random yellows from a neon sign or random cream pixels from a candle light on the other side of an AI image.
Then many physical based renderer's will have color banding, but with "moved" pixels as the noise.
Also the OpenDenoiser that is commonly used these days in renderers use AI as well. It will connect pixel colors together like jellyfish tentacles, removing the noise by making masses of same colored pixels. Which would be a strong indicator of a fixed up 3d rendered image.
@@TrancorWD The noise seed is static in Blender's Cycles renderer unless you check the box to randomize it. It's deterministic as far as I'm aware
@@TrancorWD Good point about AI denoise
@@afjerIt's always randomized, the seed may change, but path tracers have a "random value" aspect of how they shoot light paths, you can actually adjust the randomness in blender, it results in vastly less noise or even zero noise renders but with the possibility of introducing huge and strange artifacts.
I'm definitely sending this around so people can find out for themselves what likely is or isn't AI generated. I've noticed so many people post things they claim to be their art but a reverse image search reveals that they traced and didn't credit or it is actually AI generated. I can only hope that people are going to learn what to look out for when it comes to AI so that people who tend to fake "skill" they don't have can be found out, and maybe then these people will finally learn to get the skills they were faking through the time, effort, and creativity necessary to earn said skills.
Tank you good sir. I will use this to fight the good fight.
we stand in solidarity o7
I've been able to tell pretty easily from a glance whether something is AI or not before this, but this does help a lot, and probably could help people who can't tell that easily, too.
FYI the term I think you mean is encoding, not encryption, encryption usually refers to obfuscating a piece of information to be unintelligible without some kind of specific knowledge like a key or algorithm.
Thank you now I will add "jpeg artifacts" to my negative conditioning
It usually won’t understand that. Just tried it on a couple image generators.
This reminds me of an artist who was accused of using AI because one of their pieces featured a character with six fingers per hand. The character in question was Stanford Pines from Gravity Falls, who does indeed have six fingers per hand.
polydactyly time!
im glad my art is mostly physical, it would be hard for someone to claim my crochet designs are fake
Great video! A couple minor corrections:
PNGs and JPEGs don't involve any encryption, only encoding/compression.
Also, PNGs can theoretically contain JPEG artifacts if they were converted from a JPEG at some point. In the repost hellscape that is the internet this is not too uncommon; somebody might post a non-animated GIF on one site, then somebody will copy that over to a social media site which autoconverts it to JPEG when posting, then somebody else will screenshot it as a PNG and post it somewhere else, etc.
The ai music is creepy asf but I enjoy the forensics
Software developer here. The JPEG compression algorithm uses regions of 8x8 pixels, and those are called a "block".
Nitpick: the JPEG algorithm would be called an "encoding" algorithm, not an "encryption".
i'm pretty sure the alternating green and magenta lines are an artifact of camera sensors known as "moire," which happens because of the way the color filter array works in a camera sensor.
very useful video, the jpeg compression definitely explains the feeling of "this image makes sense visually, but something is wrong" in the subconscious of the brain even if the jpeg compression is minimal in the generated image
Am I tweaking or is 2:25 among us??😭
No way😭😭
Where
@@vtvl27 look closely
LOLLL I SEE IT HOW DID U SPOT THAT 😭😭😭😭
@@hampsterdancelover idk I jus saw it and was like wth😭
Hey there! You make some great points here, but it’s good to note that many are fallible. Please take care when speaking about these methods.
Image formats like JPEG and PNG are not “encryption” - they only provide encoding schemes. JPEG and other lossy formats provide “compression”. The difference here is minor - encryption assumes that the receiving end has a special authorization, like a password or key file. However, it might be important to maintain your authority when speaking to others about these topics.
Also, while PNG is a lossless format, images can be encoded in a lossy format (like JPEG) before being re-encoded in PNG.
In other words, it’s possible (and pretty likely online) that PNG files you download will be compressed in a lossy form beforehand.
Yes yes, many have acknowledged the follies in my word choice. I appreciate it though,
I mention that point as well in the video kinda quickly, if the image did go back and forth, there will be noticeable issues in isolated areas still. The jpeg encoding artifacts wouldn't line up correctly, or be the wrong size, or it becomes compounded artifacts. It does make it more difficult, that's for sure.
ok, So I'm 5:48 in and I get that's how you can tell for photos taken with a physical camera, but how about digital, a lot of artists will use programs like krita, paint, maybe even an adobe software, how would you tell the difference then? or as another fact, I use 3D software like blender however there's other options like maya or C4D how could you tell the difference there either?
I guess it depends on the post render processes. Blender has Intel's ai assisted OpenImageDenoise to help with ray tracing scatter for lower ray count renders. So if you see OpImgDenoise's kinda jellyfish connected pixels, that is a good indicator of a raytracer and a post process to "fix" lower sample rates from a CGI program. (Image wide)
In cases like maya & c4d, I'd say its more of what renderer are they using, Octane, Arnold, Redshift, Karma, Cycles, RenderMan, etc.. All ray tracers/photon based will have random pixels from other objects in a scene;
Coherent color scatter from other objects on the other side of the image.
Current Diffusion Models can't support that level of "Attention" and wont have random color data from the other side of the image.
For photoshop, krita, procreate, etc. how diffusion models work, they can't generate flat colors, without post processing. There will be random noise, if you sharpen it and sharpen it, you'll see random colors, not ray traced color noise; usually a Gaussian or tight Perlin random noise.
Amazing video!! Thank you so much for making this. I recently saw a pretty damn convincing generated image that looked very convincingly human drawn, had it not been for the poster admitting to using AI I don't think we'd have known. It scared me that one day we wouldn't be able to tell. This gives me hope that we'll still be able to know :)
The issue with this classic arms race, is now the models will be trained on how to combat this detection. 1) training on one color channel then outputting r g b, then combining to combat the noise. 2) identifying what an image looks like before and after jpeg compression & how to reverse jpeg compression. 3) only training on pngs without jpeg compression & jpegs to identify the correct output
that is assuming anyone actually cares lol. most ai generated images are so low quality it’s immediately obvious without even thinking. because who wants to spend resources on some garbage that’s gonna look fake anyways?
For now. Those peddling the stuff do spend. Those abusing it for fraud are happy to hit mostly the more gullible targets - saves them the trouble of trying to convince people with a bit more awareness.
@@0LoneTech not to mention the conspiracy theorists that are just unaware images theyre using as evidence are completely AI generated... I hope there's extensions sometime in the future for browsers to help people be aware of when an AI generated image crosses their path...
@@theopendoorevI love your pfp
Gonna be the pedantic one to say that image compression is not encryption. Compression is (usually) lossless meaning you cannot restore the original from the compressed file by reversing the compression algorithm. Encryption (by design and by principle) is reversible.
That being said this is a great explanation and yeah that sensor noise is going to be hard to replicate especially if training data is missing exif and sensor data. (if they even use metadata at all for training)
I forgot to say that the word you're looking for to describe artifacts from JPEG compression is encoding artifacts. where the wavelet tables just don't have the precision or depth to recreate the noise in each channel as accurately as the original.
The best way to recognize AI generated content is to expose yourself to it. Look at AI stuff on different sites and galleries, you will begin to notice certain details that will stand out like a sore thumb once seen in the wild. Also one thing I have noticed to work is to ask yourself "Why would the artist put that thing there".
indeed. you have to decipher human error from ai error. while yes, an ai could make the the same mistakes a human would, a human would not make the same mistakes an ai would. inconsistencies are one of these, take eyelashes. sometimes the other eyelash is in a different style for whatever reason. why is there a 6th finger coming out of their wrist? those are not very human mistakes. especially the finger.
Yeah because surrealist art doesnt exist.
Dali, sir. Why is that bloody clock melting.
@@TallicaMan1986 Surrealism isnt usually only allocated to a one part of a drawing of an entirely different genre. Trust me, you dont need to play Surrealist or Ai any time soon because the difference is rather noticeable.
@@Anonymouthful yeah for now. Can yall stop talking about the present when we discuss Ai. It's saves us a lot of time rather then going over the same old. Yeah for now. Have you seen Will Smith eating Spaghetti recently? Compare that to last year.
The problem you people have isn't even a problem outside of some legal grey areas. The problem you have with ai is the Growing Pains once this stuff starts evening then we cam start complaining because as of now. Picking out images from a tech that's exponentially evolving is a fools errand and borderline paranoia, emotions and irrationality all balled into one. There's No Gotch ya with Ai that will stick for any meaning of time. The meta is changing monthly. By Decemeber, guaranteed another crazy break through will have happened by then.
Learning how the technology is being developed will be better for everyone in the long run.
@@TallicaMan1986 If the development comes with the utter rape of an industry and culture of creative people (which is has been shamelessly shown to be) I doubt "yall" will be too happy with the out come as every artist in the world will push against un etchial AI generation.
One of the easiest ways to tell based on visuals alone is the lighting. It will look unnaturally perfect and glow-y, often the light source won't make sense (too bright, too dark, wrong angle, etc). There may be extra light WITHOUT a source, and colors won't reflect on surrounding materials like real life. If an image looks too good and it's not by a photographer, it's probably AI.
AI causes me so much of dichotomia. On one hand, I am huge fan of this technology and I can see so many use for it. On the other hand, zero-effort AI "artists" are taking my followers by doing nothing.
Yes, and automation took humans off the assembly line. That never got reverted and neither will this.
Nice techniques. For art, I've had pretty good luck looking at fine details and how details join, overlap and interfere with each other. AI always generates these things ways that no human artist would draw it.
The maroon and green colors in that error seems a bit more like color fringing to me. If Im not misunderstanding it, its a lens design issue that happens as light of different wavelength bend differently through the glass. It shows up most obviously in high contrast, high light situations, like taking a picture of a tree against the sky on a clear day, though I recall most of the green color fringing Ive had been a bit more of a yellow green, but lens coatings make alot of difference there. Doesnt matter much to your point, but it's interesting to think about the AI accidentally emulating a wierd normalization of a physical phonominon, a compression algorithm, and electrical noise all combined.
That is another source (chromatic aberration). It shows up in real life, e.g. look at the bottom of a pool with water. The cause is different wavelengths refract differently. It will have a direction. A similar digital effect is subpixel font rendering.
Yet another reminder to learn how channels in PS work. Great vid, thanks!!
The tell that will give away 97% of AI slop: the image aspect ratio is square. That's it
Damn, vine is doomed.
I'm SO glad you STARTED with the noise pattern! :D That's the biggest one.
1:28 a CCD? is there another acronym that doesnt stand for a "charged coupled device" sensor? every phone uses a cmos sensor
I have some CCD cameras and they look nothing like my phone. Also they don't have rolling shutter unlike my phone. If you see this comment then please reply because I think I'm shadow banned off UA-cam :(
@@HexOverride yeah I think the guy just mixed them up. ccds became mostly obscure in recent times, I've got a 1997 camcorder with a ccd sensor
@@Killbayne i kinda prefer ccds because they dont have rolling shutter
Even if they did shadow ban you, they do not do it consistently. Your posts may be hidden or deleted by a plethora of untold factors and there is no review process. It's really another ML system training on its own garbage.
Most of the artefacts you pinpoint as telltale signs of AI generation could be the result of post-processing the image, whether the source is AI or non-AI: resizing, sharpening, addition of noise (to simulate grain), magic wand selection of certain areas for secondary colour adjustment, etc.
As a professional colourist, and an old dude who's been using Photoshop and After Effects since the mid-90s, I've encountered those kinds of incongruities for the last 30 years, way before generative AI was a thing. (By the way, I'm also an AI user myself, for creating and treating both static and moving images).
Interesting! You keep saying "encrypting", though, when I think you mean "encoding".
You gotta boost the audio mate, this video was very quiet, but thank you for explaining this to people. You even taught me some things! Ive been doing photoshop semi-professionally for 18 years!
I am now going around saving various images as jpegs and then screenshotting them and saving the screenshot as .png so I can convince people in a couple of years that the image was AI generated.
For me, it’s just the overall look, like I always notice it when it’s “artwork” it’s just so glossy and obvious
Another problem of ai is that is beginning to train itself with ai generated content which over time will only make it look worse
They have been doing that for years now…
Great video! It's great that we have some methods for this. Quick note: I'm pretty sure it's JPEG compression or encoding, not encryption.
Regarding the thing in the thumbnail specifically, PNG can have JPEG artifacts if a conversion from JPG to PNG was performed, which can be done either automatically by the hosting site the image was uploaded to or manually by the user. By itself it is not an indication of generative AI.
The thing is that AI doesn't understand how the error works and so applies it 'incorrectly'. If you convert a png to jpeg, the jpeg artifacts will consistently show up across the whole image in the same way. An AI will only apply it on some edges, not others (for instance)
It's only a matter of time before A.I. learns to mimic jpeg compression. This video will age like fine milk.
new job,Photo authenticator
honestly learned a lot about how images in general work with this video. learning is awesome and ai bros are terrified of it