My point is that AI images don’t differ significantly enough from non-AI images. “AI images” is an extremely broad category.
If you are narrowing that category to, say, “all Dall-E images” or “all Midjourney images” or something, MAYBE. They tend to have a certain “look.” But even that strikes me as unlikely, and those are just a slice of the “AI images” pie.
As someone who has played around with Stable Diffusion and Flux, the “average color” of an image can vary dramatically based on what settings and models you’re running. AI can create remarkably real-looking images with proper variance in color and contrast, because it’s trained on real photos. Pixels, as I said, are pixels.
That’s not to mention anime or sketch or stained glass or any other medium imitation. And of course, image-to-image with in-painting, where only parts of an image are handled by the AI.
My point is that if there were overtly simple answers like, “all AI images average their color to a beige,” then there wouldn’t be all this worry about AI images. It would be easy to detect them. But things aren’t that simple, and if you spend a small amount of time looking into the depth that generating AI images has gained even in the last year, you’d realize how absurd a simple answer like that is.
So because you “make” AI generated images you are saying that they are magical and don’t follow the rules of their generation?
They are based on noise maps and inferred forwards from there. They leave a history in the pixels it’s how lots of people are detecting them.
Just because it’s trained on real photos does not mean it’s still a real photo. Just cause it looks fine doesn’t mean there isn’t stuff true beneath it.
In the video I linked they even talk about how the red blue green maps have the same values cause it started with a colorless pixel anyways. A real sensor doesn’t do that.
I worked with photographers and in Photoshop and did what you think you are doing. Working with images and pixels are not just pixels. That means nothing. Dogs are just dogs. There are still different breeds and types of dogs.
So because you “make” AI generated images you are saying that they are magical and don’t follow the rules of their generation?
That’s what you got from what I wrote?
There’s nothing “magical,” but the variety of AI images that can be produced belies the simplicity of their detection. Which has been my point this whole time.
They are based on noise maps and inferred forwards from there.
There are an infinite number of methods to diffuse noise into an image, and changes to any one of a wild number of variables produces a different image. Even with the same seed and model, different noise samplers can produce entirely different types of images. And there are a LOT of different samplers. And thousands of models.
Then there are millions of LORAs that can add or remove concepts or styles. There are ControlNets that let a generator adjust other features of the image generation, from things like poses to depth mapping to edge smoothing to color noise offsets and many many many more.
The number of tweaks that can be made by someone trying to generate a specific concept is insanely high. And the outputs are wildly different.
I don’t pretend to be an expert in this subject, I’ve barely scratched the surface.
In the video I linked they even talk about how the red blue green maps have the same values cause it started with a colorless pixel anyways. A real sensor doesn’t do that.
No, they give an extremely simple explanation of how noise maps work, and then speak as if it were law, “You’ll never see an AI image that’s mostly dark with a tiny little bit of light or mostly light with a tiny little bit of dark.” Or “You won’t have an AI photo of a flat sunny field with no dark spots.”
But that’s simply not true. It’s nonsense that sounds simple enough to be believable, but the reality isn’t that simple. Each step diffuses further from the initial noise maps. And you can adjust how that happens, whether it focuses more in lighter or darker areas, or in busier or smoother areas.
Just because someone on YouTube says something with confidence doesn’t mean they’re right. YouTubers often scratch the surface of whatever they’re researching to give an overview of the subject because that’s their job. I don’t fault them for it. But they aren’t experts.
(Neither am I, but I know enough to know how insanely much there is that I—and they—don’t know.)
None of the things they say in that video as though they are law or fact are things that haven’t already been thought of by people who know far more about the subject than these YouTubers (or me).
I did mention earlier that this sort of thing might be true for Dall-E or Midjourney or other cheap/free online services with no settings the user can tweak. AI images generated with as few steps as possible, with as little machine use as possible. They will be easier to spot, more uniform. But those aren’t all there is of AI images.
Another thing to consider: this technology is, at any given moment, at the worst it’s going to be going forward. The leaps and bounds that have been made in image diffusion even in the last year is remarkable. It is currently, sometimes, difficult to detect AI images. As time goes on, it will become harder.
Obviously it doesn’t apply to everything but nothing does. But I asked a question about using a different method of detecting AI images using the fact that color brightness does still average out and base values are usually identical and was met with condescension and incorrect information from you as well as to how color in pixel math work.
You started with dismissal and haven’t gotten better. It’s been an argument and an uphill battle to point out that this is true and yet you push it off because it’s easier to hold your position.
I wanted a conversation and you wanted to punch down. You still want to be from the pulpit of right because you like your toy. I’m done talking to you.
You certainly like to see a lot of your own words. You can dismiss me and keep talking. I don’t care anymore.
(I know you deleted this but I think it’s worth referring to.)
You are accusing me here, again, of dismissing you while simultaneously saying I type too much. These aren’t compatible.
Again, engaging with you and disagreeing isn’t dismissal. It’s conversation. It’s discussion.
Here, how about this:
I thought the video you linked was entertaining. It’s not my thing, but I can understand enjoying their style. And the claims they make are interesting for the value they hold in detecting simple, low-hanging AI fruit. I’ll grant you that.
But what I’m trying to tell you is that such a simple solution isn’t a robust one. It may work for, as I said above, low-hanging fruit. Fine. But again, if AI detection were that simple then people wouldn’t be trying to figure out how to consistently detect AI as the target continues to shift.
What I now find interesting is that you have shifted—when I addressed your video and your arguments—to attacking me and my writing rather than what I wrote. You downvote every reply I make and then try to act high and mighty about how I’m dismissing you or how I’m punching down. You dismiss me and then accuse me of it.
Why on earth are you taking this so personally? We’re talking about AI image generation, why is your pride involved?
I asked a question about using a different method of detecting AI images using the fact that color brightness does still average out and base values are usually identical and was met with condescension and incorrect information from you as well as to how color in pixel math work.
You asked a question about why tools don’t use an extremely simple method of detecting AI images. I said that wouldn’t work. Initially I misunderstood your question and my response was overly simple, but it wasn’t wrong. Simple methods of detecting AI images don’t work for all AI images.
You started with dismissal and haven’t gotten better.
I didn’t dismiss you. If I had I wouldn’t have bothered to respond. You hadn’t presented much besides a vague question initially, and I disagreed with it.
When you came back with more I presented my position, that AI image generation is much more varied and complicated than your question and YouTube video assume. Just because I’m disagreeing with you and providing context doesn’t mean I’m dismissing you. Dismissing you would be to say, “No, you’re wrong, go away.” Not to explain why the simple method you’re talking about isn’t feasible, broadly, for the entirety of AI images.
If I wanted to dismiss you, I wouldn’t bother wasting my time on a response.
It’s been an argument and an uphill battle to point out that this is true
And you’re accusing me of clinging to my position. 🙄
I wanted a conversation and you wanted to punch down. You still want to be from the pulpit of right because you like your toy.
Where on earth did you get the impression that I want to be right because I like AI image generation? Or that I wanted to punch down?
Someone disagreeing with you and responding to your argument without accepting it isn’t dismissal, it isn’t punching down, it isn’t condescension. It’s engagement with what you’re saying. Just because I don’t agree with you doesn’t mean I think I’m better than you or smarter than you or anything like that, it just means I think I’m right.
I’m saying it because it’s not only obvious with even a moments thought (you can literally just ask it for an entirely red image or whatever), but also because it’s easily provable.
Prompt: “Under the sea”
Image:
Average pixel colour:
Prompt: “a man with red hair wearing a red coat standing in front of a red background”
Image:
Average pixel colour:
So I ask you the same question. Did you just say that because you felt like it was true?
The average brightness values of those are both middle of the road grey. Sorry I should have rephrased as I misspoke calling it beige but the point still stands that has the most average toned color.
If you look they are middling around 50-60% where as a similar red photo intake would likely have a higher contrast and an average color with a higher brightness.
I’d expect that many images are going to be somewhere near 50% grey if you average their luminance out overall. That’s just the average of every colour though. The fact that averaging a range of things tends toward a standard distribution isn’t particularly surprising. Again though, it’s not hard to get a diffusion model to generate something outside of that expectation.
Prompt: “night sky”
Image:
Average colour:
Average brightness: 21%
Prompt: “lineless image of an old man drawn in yellow ink on white background”
Ehhh you’d be surprised how much small highlights and true dark values will skew an average.
You don’t really get blown out or completely under exposed parts of AI generated images. It keeps trying to add some saturation.
I will say I’m getting different brightness levels than you are but it’s not a big deal. Point made.
However I will say saturation is also pretty much 50% across the board for all the colors so maybe there is something there to use as an indicator. The average color always comes out grey toned somehow.
Testing it myself has been pretty spot on.
Just to point it out, here is a photo I took to mimic the red man. And the color average from it.
It doesn’t even have the highlights of the face and yet it’s already much brighter and has a much higher saturation.
Also, this is tangential to the rest of our conversation, but I appreciate the dedication to the comment chain required to actually set up something with similar composition to the red man image and take a picture of it. Even has some black in the image in roughly the same size and area as his sweater. :D
I’m not sure what you mean by the saturation being around 50% across the board. If I peek the HSB of all of the averages only that first teal-ish one appears to be around the mid point for saturation.
Isn’t there a whole thing about if you average out colors on AI generated photos you get a uniform beige color?
I don’t get why these tools don’t just do that but I guess you got to keep the marketing up of using AI to find a solution.
Either that’s not true of AI images or it’s true of all images. There aren’t answers that simple to this. Pixels are pixels.
What? That’s some extreme logic.
First of all why would it be true of all images? Real photos would have variance of contrast and color in different ways.
These guys literally point out average colors and contrast in AI images
Instead of engaging the conversation you just say pixels are pixels? Luke that means something smart?
My point is that AI images don’t differ significantly enough from non-AI images. “AI images” is an extremely broad category.
If you are narrowing that category to, say, “all Dall-E images” or “all Midjourney images” or something, MAYBE. They tend to have a certain “look.” But even that strikes me as unlikely, and those are just a slice of the “AI images” pie.
As someone who has played around with Stable Diffusion and Flux, the “average color” of an image can vary dramatically based on what settings and models you’re running. AI can create remarkably real-looking images with proper variance in color and contrast, because it’s trained on real photos. Pixels, as I said, are pixels.
That’s not to mention anime or sketch or stained glass or any other medium imitation. And of course, image-to-image with in-painting, where only parts of an image are handled by the AI.
My point is that if there were overtly simple answers like, “all AI images average their color to a beige,” then there wouldn’t be all this worry about AI images. It would be easy to detect them. But things aren’t that simple, and if you spend a small amount of time looking into the depth that generating AI images has gained even in the last year, you’d realize how absurd a simple answer like that is.
So because you “make” AI generated images you are saying that they are magical and don’t follow the rules of their generation?
They are based on noise maps and inferred forwards from there. They leave a history in the pixels it’s how lots of people are detecting them.
Just because it’s trained on real photos does not mean it’s still a real photo. Just cause it looks fine doesn’t mean there isn’t stuff true beneath it.
In the video I linked they even talk about how the red blue green maps have the same values cause it started with a colorless pixel anyways. A real sensor doesn’t do that.
I worked with photographers and in Photoshop and did what you think you are doing. Working with images and pixels are not just pixels. That means nothing. Dogs are just dogs. There are still different breeds and types of dogs.
That’s what you got from what I wrote?
There’s nothing “magical,” but the variety of AI images that can be produced belies the simplicity of their detection. Which has been my point this whole time.
There are an infinite number of methods to diffuse noise into an image, and changes to any one of a wild number of variables produces a different image. Even with the same seed and model, different noise samplers can produce entirely different types of images. And there are a LOT of different samplers. And thousands of models.
Then there are millions of LORAs that can add or remove concepts or styles. There are ControlNets that let a generator adjust other features of the image generation, from things like poses to depth mapping to edge smoothing to color noise offsets and many many many more.
The number of tweaks that can be made by someone trying to generate a specific concept is insanely high. And the outputs are wildly different.
I don’t pretend to be an expert in this subject, I’ve barely scratched the surface.
No, they give an extremely simple explanation of how noise maps work, and then speak as if it were law, “You’ll never see an AI image that’s mostly dark with a tiny little bit of light or mostly light with a tiny little bit of dark.” Or “You won’t have an AI photo of a flat sunny field with no dark spots.”
But that’s simply not true. It’s nonsense that sounds simple enough to be believable, but the reality isn’t that simple. Each step diffuses further from the initial noise maps. And you can adjust how that happens, whether it focuses more in lighter or darker areas, or in busier or smoother areas.
Just because someone on YouTube says something with confidence doesn’t mean they’re right. YouTubers often scratch the surface of whatever they’re researching to give an overview of the subject because that’s their job. I don’t fault them for it. But they aren’t experts.
(Neither am I, but I know enough to know how insanely much there is that I—and they—don’t know.)
None of the things they say in that video as though they are law or fact are things that haven’t already been thought of by people who know far more about the subject than these YouTubers (or me).
I did mention earlier that this sort of thing might be true for Dall-E or Midjourney or other cheap/free online services with no settings the user can tweak. AI images generated with as few steps as possible, with as little machine use as possible. They will be easier to spot, more uniform. But those aren’t all there is of AI images.
Another thing to consider: this technology is, at any given moment, at the worst it’s going to be going forward. The leaps and bounds that have been made in image diffusion even in the last year is remarkable. It is currently, sometimes, difficult to detect AI images. As time goes on, it will become harder.
(Which your video example even says.)
Obviously it doesn’t apply to everything but nothing does. But I asked a question about using a different method of detecting AI images using the fact that color brightness does still average out and base values are usually identical and was met with condescension and incorrect information from you as well as to how color in pixel math work.
You started with dismissal and haven’t gotten better. It’s been an argument and an uphill battle to point out that this is true and yet you push it off because it’s easier to hold your position.
I wanted a conversation and you wanted to punch down. You still want to be from the pulpit of right because you like your toy. I’m done talking to you.
(I know you deleted this but I think it’s worth referring to.)
You are accusing me here, again, of dismissing you while simultaneously saying I type too much. These aren’t compatible.
Again, engaging with you and disagreeing isn’t dismissal. It’s conversation. It’s discussion.
Here, how about this:
I thought the video you linked was entertaining. It’s not my thing, but I can understand enjoying their style. And the claims they make are interesting for the value they hold in detecting simple, low-hanging AI fruit. I’ll grant you that.
But what I’m trying to tell you is that such a simple solution isn’t a robust one. It may work for, as I said above, low-hanging fruit. Fine. But again, if AI detection were that simple then people wouldn’t be trying to figure out how to consistently detect AI as the target continues to shift.
What I now find interesting is that you have shifted—when I addressed your video and your arguments—to attacking me and my writing rather than what I wrote. You downvote every reply I make and then try to act high and mighty about how I’m dismissing you or how I’m punching down. You dismiss me and then accuse me of it.
Anyway, I hope you have a good day.
Why on earth are you taking this so personally? We’re talking about AI image generation, why is your pride involved?
You asked a question about why tools don’t use an extremely simple method of detecting AI images. I said that wouldn’t work. Initially I misunderstood your question and my response was overly simple, but it wasn’t wrong. Simple methods of detecting AI images don’t work for all AI images.
I didn’t dismiss you. If I had I wouldn’t have bothered to respond. You hadn’t presented much besides a vague question initially, and I disagreed with it.
When you came back with more I presented my position, that AI image generation is much more varied and complicated than your question and YouTube video assume. Just because I’m disagreeing with you and providing context doesn’t mean I’m dismissing you. Dismissing you would be to say, “No, you’re wrong, go away.” Not to explain why the simple method you’re talking about isn’t feasible, broadly, for the entirety of AI images.
If I wanted to dismiss you, I wouldn’t bother wasting my time on a response.
And you’re accusing me of clinging to my position. 🙄
Where on earth did you get the impression that I want to be right because I like AI image generation? Or that I wanted to punch down?
Someone disagreeing with you and responding to your argument without accepting it isn’t dismissal, it isn’t punching down, it isn’t condescension. It’s engagement with what you’re saying. Just because I don’t agree with you doesn’t mean I think I’m better than you or smarter than you or anything like that, it just means I think I’m right.
It is absolutely not true of all AI images. I’d be surprised if it’s even true about most AI images.
Just saying that because you feel like it’s true or because you’ve participated in that line of thought for even 5 seconds?
AI images come from a noise map, it’s true cause they generate from it in a consistent manner.
I’m saying it because it’s not only obvious with even a moments thought (you can literally just ask it for an entirely red image or whatever), but also because it’s easily provable.
Prompt: “Under the sea”
Image:
Average pixel colour:
Prompt: “a man with red hair wearing a red coat standing in front of a red background”
Image:
Average pixel colour:
So I ask you the same question. Did you just say that because you felt like it was true?
The average brightness values of those are both middle of the road grey. Sorry I should have rephrased as I misspoke calling it beige but the point still stands that has the most average toned color.
If you look they are middling around 50-60% where as a similar red photo intake would likely have a higher contrast and an average color with a higher brightness.
I’d expect that many images are going to be somewhere near 50% grey if you average their luminance out overall. That’s just the average of every colour though. The fact that averaging a range of things tends toward a standard distribution isn’t particularly surprising. Again though, it’s not hard to get a diffusion model to generate something outside of that expectation.
Prompt: “night sky”
Image:
Average colour:
Average brightness: 21%
Prompt: “lineless image of an old man drawn in yellow ink on white background”
Image:
Average colour:
Average brightness: 90%
Ehhh you’d be surprised how much small highlights and true dark values will skew an average. You don’t really get blown out or completely under exposed parts of AI generated images. It keeps trying to add some saturation.
I will say I’m getting different brightness levels than you are but it’s not a big deal. Point made.
However I will say saturation is also pretty much 50% across the board for all the colors so maybe there is something there to use as an indicator. The average color always comes out grey toned somehow.
Testing it myself has been pretty spot on.
Just to point it out, here is a photo I took to mimic the red man. And the color average from it.
It doesn’t even have the highlights of the face and yet it’s already much brighter and has a much higher saturation.
Also, this is tangential to the rest of our conversation, but I appreciate the dedication to the comment chain required to actually set up something with similar composition to the red man image and take a picture of it. Even has some black in the image in roughly the same size and area as his sweater. :D
I’m not sure what you mean by the saturation being around 50% across the board. If I peek the HSB of all of the averages only that first teal-ish one appears to be around the mid point for saturation.