“Stanford student Billy Ramirez, a member of the class of 2027, told the publication that he remembered posing for the original photo, only to discover that he had been replaced by an AI-generated Black woman."
It is worth remembering that AI replaces the ENTIRE image, even if you ask it to change only one thing. That’s why every single person in the image looks different.
You werent talking about any specific system at all until that user mentioned Affinity. I think you’re moving the goal posts here as youre now trying to frame the discussion as if you were talking about some other yet unnamed system the whole time when your original comment mentioned nothing of the sort.
If that is in fact how the system in question works.
We know that’s how the system works by the fact…
We know how THIS system works - but…
They cast doubt, I point out proof, they deflect to other systems they might be right about. Sounds like they moved the goalposts to me. I don’t need to know which system it is to know it does what we can see it do.
It is worth remembering that AI replaces the ENTIRE image, even if you ask it to change only one thing.
This is what they replied to. You made a blanket statement about using AI to edit photos not anything about any particular program or system. Furthermore, you can’t know which system was used here as you dont know whether all these changes were intentionally done or simply a byproduct of using something that changes the entire image.
That wasn’t proof of anything. Not only does it not make any sense in the context of the discussion, it was incorrect (“every single person in the image looks different”), as there are more than three people in the image that you probably didn’t notice, and they appear to be largely untouched.
This article is about Stanford U using AI on a photo of their students. No one was talking about what system they used. Then lIlIlIlIlIlIl@lemmy.world started talking about Affinity, which is making it seem like people who are pro AI are moving the goalpost to avoid talking about the issue. The issue being that AI is being used inappropriately.
It seems there are different AI systems. One that can modify a specific region and one that needs to modify the whole image. I doubt that anyone who came to this post cares what system Stanford used. They just hate that this is happening.
You are moving the goalpost, changing the subject, and talking about a new subject. When Susaga@sh.itjust.works brought this up, you did turn on him and continued to talk about AI systems.
Why would they request to retexture the stone tiles at the top to have stronger speckles? That’s not just “we turned up the contrast”, cause the rest of the image has lower contrast.
It’s kinda hard to judge the contrast difference, given that the left image is an arctifact-riddled JPG, and the right image is an artifact-riddled photo of a print.
I’m just saying, a generative AI model does not inherently need to apply changes to the entire image. That is a very outdated understanding of the process, which tells me that it’s probably been a hot minute since you’ve been paying any attention to AI, and therefore you probably shouldn’t speak with such an authoritative tone on the subject.
Regional control (especially ControlNet) was a thing before the current models got popular.
Which is precisely why I hate the fact that the general lazy models are the only ones that are popular. Where the hell is the tooling to do fancy controlled geometric transforms plus style transfer in a controlled manner, etc? It’s all just “write a prompt and run it a hundred times”, nobody want to put human effort into the system anymore
As someone else pointed out, there are systems where you can isolate and replace only a portion of the image while leaving the rest unchanged, but you’re correct in this case, as you point out in your last sentence.
This sort of system takes the input image which it blurs and adds noise, and then regenerates from that rather than an initial pure-noise seed. You can mask or add more noise to the areas that you want and more dramatic change. Whole image-to-image diffusion; mostly uses the same mechanisms as a regular text-to-image diffusion.
This is inaccurate, current models can target specific zones in an image. Midjourney was capable of that ages ago, and Midjourney is already basically ancient history at this point.
“Stanford student Billy Ramirez, a member of the class of 2027, told the publication that he remembered posing for the original photo, only to discover that he had been replaced by an AI-generated Black woman."
It is worth remembering that AI replaces the ENTIRE image, even if you ask it to change only one thing. That’s why every single person in the image looks different.
The Asian girl and the white dude both got mildly more attractive. But Billy got that glow up!
The middle dude is a mash-up of the two guys in the original. One face and one hair.
Everyday I am more astounded at how well transracial people are passing!
If that is in fact how the system in question works. There are systems which allow you to edit regions, like Affinity
We know that’s how the system works by the fact that every single person in the image looks different.
We know how THIS system works - but Affinity is just one example of how you can modify an area with AI, not just the whole image.
Are you asserting how every piece of software works, based on a single image example? Because that’s wildly incorrect.
We aren’t talking about other systems. Don’t move the goalposts.
You don’t know what system you even are talking about.
You werent talking about any specific system at all until that user mentioned Affinity. I think you’re moving the goal posts here as youre now trying to frame the discussion as if you were talking about some other yet unnamed system the whole time when your original comment mentioned nothing of the sort.
They cast doubt, I point out proof, they deflect to other systems they might be right about. Sounds like they moved the goalposts to me. I don’t need to know which system it is to know it does what we can see it do.
This is what they replied to. You made a blanket statement about using AI to edit photos not anything about any particular program or system. Furthermore, you can’t know which system was used here as you dont know whether all these changes were intentionally done or simply a byproduct of using something that changes the entire image.
That wasn’t proof of anything. Not only does it not make any sense in the context of the discussion, it was incorrect (“every single person in the image looks different”), as there are more than three people in the image that you probably didn’t notice, and they appear to be largely untouched.
Brother… are you an AI?
As a neutral observer, you clearly and unambiguously moved the goalpost.
This article is about Stanford U using AI on a photo of their students. No one was talking about what system they used. Then lIlIlIlIlIlIl@lemmy.world started talking about Affinity, which is making it seem like people who are pro AI are moving the goalpost to avoid talking about the issue. The issue being that AI is being used inappropriately.
It seems there are different AI systems. One that can modify a specific region and one that needs to modify the whole image. I doubt that anyone who came to this post cares what system Stanford used. They just hate that this is happening.
You are moving the goalpost, changing the subject, and talking about a new subject. When Susaga@sh.itjust.works brought this up, you did turn on him and continued to talk about AI systems.
Or they ran a prompt with more than one unique request in it?
Why would they request to retexture the stone tiles at the top to have stronger speckles? That’s not just “we turned up the contrast”, cause the rest of the image has lower contrast.
It’s kinda hard to judge the contrast difference, given that the left image is an arctifact-riddled JPG, and the right image is an artifact-riddled photo of a print.
I’m just saying, a generative AI model does not inherently need to apply changes to the entire image. That is a very outdated understanding of the process, which tells me that it’s probably been a hot minute since you’ve been paying any attention to AI, and therefore you probably shouldn’t speak with such an authoritative tone on the subject.
Regional control (especially ControlNet) was a thing before the current models got popular.
Which is precisely why I hate the fact that the general lazy models are the only ones that are popular. Where the hell is the tooling to do fancy controlled geometric transforms plus style transfer in a controlled manner, etc? It’s all just “write a prompt and run it a hundred times”, nobody want to put human effort into the system anymore
As someone else pointed out, there are systems where you can isolate and replace only a portion of the image while leaving the rest unchanged, but you’re correct in this case, as you point out in your last sentence.
This sort of system takes the input image which it blurs and adds noise, and then regenerates from that rather than an initial pure-noise seed. You can mask or add more noise to the areas that you want and more dramatic change. Whole image-to-image diffusion; mostly uses the same mechanisms as a regular text-to-image diffusion.
This is inaccurate, current models can target specific zones in an image. Midjourney was capable of that ages ago, and Midjourney is already basically ancient history at this point.
They merged the two white guys…
Added a black woman…
And took like 20lbs off the Asian woman. Its shitty, but I don’t understand why it’s news outside of Stanford.
Reminds me of something Dean Pelton would do tho.
If the camera adds 10 pounds then it’s basically break even.
The camera doesn’t shorten that neck extension that the Asian girl got, though!
Maybe Pierce would’ve done it for marketing class and said something misogynistic about making the Asian woman thinner
You don’t understand why it’s news outside Stanford…