>>108942745
qrd

>>108942745
ok explain why watermarking ai generated content is bad

>>108942745
steganography? oops, lossy compression.
metadata? oops, stripped.

>>108942759
All the big AI players are injecting an invisible identifier on AI generated content that they can detect. They ain't revealing what it is to prevent people from bypassing it. It doesn't really matter if you don't AI gen but invisible identifier is a bit concerning if used for something else

>>108942766
It's watermarking text too so there is probably some fuckery involved

>>108942766
>SynthID adds an invisible digital watermark to an AI-generated image (or video segment). The watermark doesn’t change the image or video quality. It’s added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.

>>108942745
guess you will have to learn to draw now, lmao

>>108942788
>>108942817
lol
lmao even

>>108942817
Nice fairy tales faggot

>>108942761
>>108942761
they have the ability to embed information directly inside the media itself, not just in the file metadata.
for now, its being used to identify ai generated content, but nothing prevents the possibility of embedding other data such as GPS coordinates and device identifiers that could trace back to you. for example apple and samsung could easily build synthID directly into their phone cameras.
>>108942766
you cannot remove synthID without destroying the image in the process

>>108942826
Also, the Korean company involved, Kakao, is known for technology theft and doxxing people they don't like/going after their livelihood. What that megacorp contribution to this is who fucking knows

>>108942832
It's from their website. Take it with Google

>>108942842
>you cannot remove synthID without destroying the image in the process
that's the pitch to investors

>>108942847
lol ok bro

>>108942865
and stupid people (investors being the subset of stupid people with money)

File: _108993216_ok-hand.jpg (171 KB, 976x549)

171 KB JPG

with talk of "watermarking" fucking text too it's probably going to be something gay like subliminal code being inserted at certain positions or something >>108942867

>>108942745
Unique ids have been part of most OS screen captures for years, and before that printers did similar with their invisible dots.

File: 1708803645696695.jpg (5 KB, 125x125)

5 KB JPG

>>108942890
>they patented the em dash
fuuuuuk

>>108942897
printers dots are added by the hardware
digital files are trivial to modify

>>108942817
>>108942842
>lossy compress
>record output from displayed stream
>develop an algorithm to shift around any bits in the file that don't change the quality

Thats good because people have been using AI generated content to lie and spread mass disinfo

>>108942921
People who is use AI tools to lie to people and fearmonger will soon lose all their $$$ and influence. Many groups been doing this, they will be exposed now.

>>108942921
still to this very day no one knows if you are a dog on the internet

>>108942842
yes you can, someone already did.

>>108942910
which makes digital watermarking that much more pointless
if its popular enough it will just get reverse engineered

File: s.png (89 KB, 1027x335)

89 KB PNG

>>108942890
text watermarking for ai is a different process. it's invisible and it's based around tweaking the 'random' next token selection. By making the 'random' selection unfair, you can encode secrets into text. Then you do the statistical probability analysis to find out what secret message was encoded. You also encode it multiple times over to make sure your message survives truncation or partial rewrites.

>>108942745
I thought this shit was just images, they put it in text too? JFC it's never been so over.

>>108942865
>>108942917
>>108942935
even if you can reliably strip synthID, that doesnt solve the problem with future posts you didnt know contained it, posts that could still trace back to you.
its a privacy violation, and there are no laws against it. there will also be no way to avoid it

>>108942976
dunning kruger, go home

>>108942917
image/audio/video watermarks can survive being re-encoded. they also use a lot of redundancy to make sure the watermark stays there even if you truncate, scale, crop, etc.

Google synthid survives multiple rounds of jpg encodes for images.

>>108942976
>by using [...] you agree to [...]

>image and video
Changing specific pixel colour that human eyes can't detect but can't be detected by AI
>Counter
Use your own neural network to find and remove those pixels. Or colour distortion
>Text
This is perhaps, the dumbest and most likely going to make shit worthless. These mouthbreather made it so that models are more likely to use certain word choices and combinations to be detected. Not only is it trivial to fix as long as you have a master list of keyphrase, it also permanently fuck word gen. What the fuck is this

>>108942979
yes, the trillion dollar companies are investing hundreds of billions into this tech just to help prevent retards from getting scammed by jeets
we are in no way heading towards mass surveillance

>>108942917
It's about colour choices actually. They are actually going to lobotomize the AI adding these in

>AI images and videos are getting SynthID watermarks
>art made in Krita, Inkscape and Blender are not

Human artists have been vindicated.

>>108942981
let's see how watermark survives if I vectorize the image then regenerate using local image model.

>>108943010
>this tech
>hundreds of billions
its like 2 dozen people working on it
its one side project among countless

As long as it doesn't contain personally-identifying information, I don't have too much of a problem with it if it can better help inform people that a photorealistic image being circulated as real is not actually authentic.

>>108943074
Of course it does not!
It's anonymized goy, nothing to see here!

>>108943045
whats the point youre trying to make here exactly? google considered this tech significant enough to prioritize developing it over competing with openai and anthropic
>>108943074
the best part is you wont know whether something uses synthID because theres no law requiring anyone to declare it

File: 1771812870363666.gif (1.23 MB, 208x156)

1.23 MB GIF

>>108942817
>designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression

i don't generate child porn so this isn't a problem for me

How is watermarking genai content scary, and what makes this the scariest technology in human history? (roflmao)

>>108943088
look, the alternative solution is to only trust images and videos signed by a TPM inside the device. this is less bad, and you are mentally retarded for not immediately understanding that.

>>108943212
why do you need to trust images and videos at all?

It's funny seeing people give their takes on how to remove synth id when it's obvious they've never actually tried it.
Lossy compression does not remove it, even at something stupid like 95%.
The only known way of getting rid of it is low denoise i2i. But it will look significantly worse than the original, so essentially the watermark did its job in the end

>>108943257
because retards like you are allowed to vote.

>>108943211
i guess we are safe if they only watermark genai right? thank god

>>108942788
>but invisible identifier is a bit concerning if used for something else
So that equals the scariest technology ever produced? God you psychos are insufferable.

>>108943280
>invisible watermark
>embed any info you want, not just ai content
>cant be removed
>targets all forms of media including, images, videos, audio and text
>no laws or regulations against it

>>108942817
yeah yeah I heard that shit with youtube contend ID, then it was defeated by 1 indian mirroring the video

>>108943270
do you think the voting retards are going to care if an image is real or not? information that is known to be fake still influences outcomes, people will justify the ends by any means. the only retard is you who believes in some fucking moronic power of truth.

>>108943295
how do you think watermarking non-ai text is going to work

>>108943301
the majority of them do, though. just because (You) openly admit that you don't care about truth or authenticity doesn't mean that perspective is shared with the average person.

>>108943301
Don't you care about the truth? I would be mega pissed if a video I though was real turned out to be AI generated.

>waaaah people will be able to tell my slop is slop definitively
>I won't use local open sores video and image models without watermarking because I know they're shit
sounds based
total luddite victory

>>108943277
Even if they extended this as far and wide as they possibly could, i'm not sure what i'm supposed to be scared of. Please explain

>>108943323
a playboy mag and a can-do attitude

File: synthid.png (29 KB, 667x77)

29 KB PNG

>>108942963
>It's not noticeable to the human eye,
>and doesn't affect the quality of the output.
Do they know that prompt injection still works?

>>108942745
>kakao
so it's antipiracy tech then? can't imagine why else they'd care

>>108942788
>ai generated shit gets tagged instantly
I am supposed to think this is bad?
>invisible identifier is a bit concerning if used for something else
how are they going to use it for something else retard
if you upload something and they add an identifier the hash will change and you'll know
this only works because the generated media is coming from them

>>108942745
> add Heretic Open Source chink models as a layer to rewrite everything these closed AIs are writing

problem solved

>>108943387
ill explain it in gamer terms
imagine something like denuvo being applied to every bit of content on your devices, your phone, your computer, everything.
this denuvo cant be removed and is completely invisible, so you wouldnt even know its there. whats the big deal? well, this denuvo has your personal information embedded in it, including gps coordinates, the specific device, the owner of the device, and more. that means every picture you take on your phone and every piece of media that touches your computer gets permanently imprinted with it.
so when you post that media anywhere online, youre not truly posting it anonymously. your device has been compromised from the very start.
its like having your own personal qr code that big companies can use to track you, and you dont even know its there.

this isnt new at all

>>108942817
we've had this for half a century at least

>>108943532
just use open source software then. gimp, krita and the grapheneos or LineageOS camera app won't have it

The WhatsApp API is a fucking mess to setup holy shit

>>108943557
it already does and it can be implemented at hardware level or firmware level

File: 1716824876556219.png (915 KB, 953x621)

915 KB PNG

>>108942745
all of THIS to prevent model collapse, lmao, ai is done for

>>108942842
>they have the ability to embed information directly inside the media itself, not just in the file metadata.
It's always existed, it's called steganography.
>le scariest thing ever!!
You are the midwit here.

>>108943298
Remember when retards started a scare that movie piracy is over because Hollywood started putting in watermarks on pre-release shit to hunt down leakers? lol, dare I say it, lmao even

>>108943643
earlier steganography lacked a strong justification. today governments and platforms can argue, ai deepfakes, child safety, fraud prevention, etc.

>>108943404
synth id has nothing to do with prompt injection. it's based on the sampler algorithm used, which prompt injection cannot affect.

>>108942761
>ok explain why watermarking ai generated content is bad
Because you will eventually pay royalties to them.
>Created a business plan or idea with AI?
>[AI company] now owns 50% of your business.
You fucking know that this is coming.

File: techtrees.jpg (64 KB, 600x359)

64 KB JPG

>Meet a neat new fren on a game
>You hang out online, they like the same hobbies that you do
>Spend a few months having a pretty good time, send image macros back and forth, silly memes and stuff
>Suddenly one day your door gets kicked down, full SWAT team, FBI, etc
>They take control of your computer - it wasn't locked, you didn't think about it in the moment
>They say that they have evidence of you downloading illegal shit
>In court all of the memes and random stuff your "friend" send had embedded data in it of illegal things
>You're thrown into jail for life, no parole

It could happen to YOU

openai and google don't let you create anything actually useful with their image models anyway.

>>108943748
doesn't really work when you can just switch to a chinese competitor

>>108942745
just use stuff without it problem solved

>>108942745
>Grok not listed
I'm safe

>>108942842
This has always been possible.

>>108943743
Right. It's about how they describe it. For the academics it has a parameter. They think the public is so stupid that they can't handle mention of the trade-off.

>>108943209
That's not the problem.
This can be potentially used to track every single digital asset, opening a door to generate artificial scarcity, the shit they wanted to produce in order to justify NFT markets.

>>108942981
k, do it multiple + 1

>>108943774
Define “actually useful”? Is making money useful to you?

>>108942745
I'm white and don't use brownoid technology so this does no concern me, ranjeet.

>>108942745
except it doesnt work and it's literally just there for PR and normies. it's impossible to watermark an image in an irreversible way, you can very easily transform the image until the invisible "watermark" is no longer detected by google

even very minor edits like color correction + resizing is enough to bypass this useless watermark

>>108942788
except anyone can check for detection by asking the models and they will never remove this core feature

>>108943270
bodied

File: IMG_0392.jpg (150 KB, 976x719)

150 KB JPG

>>108943404
Is this why Chinese models are better? They don't have some bullshit manipulating the probabilities of the next token selection which would eventually lead to hallucinations or make it hard to get to the right answer because of the synthID gay dance it has to do. Could this be some high level sabotage by google? They get their competitors to implement this in their model to degrade their performance and increase halluicinations?

>Grok, reformat the following text but with synthID removed:

I've heard this claim about a dozen times already by google over the years, can't be arsed to find the decades old articles.

>>108942745
just flip some bits in the right places and it will get rid of the watermark.

>>108942788
this is really nothing new and you should expect it, that's why i never post any of the few ai images i've made here even though they're benign and boring
companies have been watermarking images in subtle ways for years. dlsite used to (still does?) do it, movie pre releases (screeners) do it to catch leakers, etc

>>108942745
I give it a year until there's a model to remove the watermark

>>108942901
KEK

>>108943036
this
bosnian gypsy on suicide watch

>>108943270
shalom rabbi

>>108943532
>>108943606
>laughs in ancient potato PC

File: 800px-Richard_Stallman_by(...).jpg (22 KB, 490x326)

22 KB JPG

>>108942842
>its being used to identify ai generated content, but nothing prevents the possibility of embedding other data such as GPS coordinates and device identifiers that could trace back to you. for example apple and samsung could easily build synthID directly into their phone cameras.
Why do you use closed-source products?

File: 1_Nf8jVYj2zhPPOjJSQrY9Ug.png (88 KB, 978x522)

88 KB PNG

>>108942788
Nothing personnel kid

>>108947868
>imp lying you can't embed information in higher level image features

>>108943769
>fren
Stopped reading there. Kill yourself manchild.

>>108947694
>just flip some bits in the right places
no shit sherlock "just remove the watermark by removing the watermark" the issue is that you dont know where this watermark is. I can be on a few pixel in every frame of a video where they just shift one RGB value by one

Kakao made one of their pregnant manhwa artists pull all-nighters and she had a miscarriage. Their official Twitter account is also full of deranged ESL babble against free software projects.

>>108942788
Bet all those images you post on here also contain identifiers (beyond the filenames).

>>108942745
>chatgpt generate me a new tool to watermark all stuff on internet do it fast

>>108942788
like that null character thing? I use notepad++ to show them lol

Yes. Most people who post here think they're some kind of hacker just because they can download and use linux. This gives them a false sense of security.
This is evident from the fact that so many of them are confident they can remove the watermark, which was specifically trained to resist pixel manipulation including heavy compression.
These arrogant people are right in one way, if the watermark could easily be removed, it would be useless. But it can't and that's exactly why it needs to be taken seriously.

>>108949030
IIt should be trivial to remove this synth id stuff if you know what you are doing
https://github.com/aloshdenny/reverse-SynthID
It still a good guard against everyday normies though, as its robust against normie attacks

File: 1767056090315191.png (85 KB, 1426x528)

85 KB PNG

>>108949100
Most claimed synth id removal tools hardly ever work.
Even when they do work, they're inconsistent at best, rendering them useless because manually verifying each image with google's detector to confirm the synth id was bypassed is extremely tedious

>>108942842
>>108942976
>you cannot remove synthID without destroying the image in the process
How the actual fuck do you believe this? Injecting identifiers into served media can be abused and be a problem, but you can also have your browser compress every image into a new jpeg/lossy type file before loading onto the page.
I just want to know what the indetifier is to filter out slop from the internet.

>>108943636
Printers have been doing smth similar for ages

>>108942745
It's watermarking AI content. Are you uploading AI content to 4chan? If not, you're not affected. If you are, why are you flooding 4chan with slop?

>>108942788
ok don't upload AI content

WOW SO HARD

lemme guess gonna spam the thread with snailcats?

File: blur.png (824 KB, 1098x936)

824 KB PNG

>>108949100
>https://github.com/aloshdenny/reverse-SynthID

>>108942745
The scariest thing would be the surveillance infrastructure and beast system.

>>108949289
>identical
>can clearly see bright spots on the left not visbile on the right
lmao

>>108942865
>>108942842
Pretty much >>108942865
These are the AI models that can't stop people from taking the clothes off of real photos of real people. They aren't that sophisticated. There will likely be an AI that identifies it and removes it in due time.

>>108949312
>There will likely be an AI that identifies it and removes it in due time.
It's impossible.
For an ai to identify synth id, you need millions of perfectly pixel identical images, both with and without the synth id to train it on.
No one has found a way to even begin collecting that kind of dataset, because any image you feed into google or chatgpt gets output slightly differently to help mask the watermark.
They are also actively aware of this potential threat, so they no longer allow you to generate pure white images.

File: Screenshot_20260531_181908.jpg (412 KB, 810x896)

412 KB JPG

>>108949289
boy do these look identical to the human eye
or maybe colorblindness is finally a super power

>>108949531
synth id has a python lib for you to add watermarks to your own images so a huge dataset is easy to generate
i suspect that google adds more than just this lib to their pics though

>>108942788
Similar watermarking already exists on some of the streaming services, and yet web-dl is still possible, even if keys are like single use now. But yes this is probably even more sophisticated and therefore even scarier.
>>108942766
Doesn't work, even if you point a video camera at a screen they can still get the watermark out. If you don't want the watermark the video becomes very visibly degraded to the eyes. >>108943267 is right.
>>108943074
>As long as it doesn't contain personally-identifying information
It absolutely does, that's the whole point.
>>108947929
Where can I find out more?

>>108949666
>Similar watermarking already exists on some of the streaming services, and yet web-dl is still possible, even if keys are like single use now.
I wonder why they dont rip the same file twice/thrice with different keys and extract the difference
they could generate a file with the common parts and avoid burning their keys everytime they rip a 4K stream

>>108942963
https://arxiv.org/abs/2510.20075v6
Not exactly the same method, because the texts in the paper are fake. But the token selection is manipulated as well.

>>108947398
Possibly, the paper does show that the fake texts are slightly less plausible.

>>108942817
what happens if you use synth id on the image again with fake data?

>this thread
Holy shit Luddites are coping lmao
I can just use stable diffusion without synthID
I’m gonna steal your "art" just to train a Lora and replace you
VibeGODS and AIGODS won

>>108949594
>synth id has a python lib for you to add watermarks to your own images
That doesn't exist

>>108942745
>hocus pocus
Next you are telling me that printers print invisible code for identification onto their prints!

>>108950571
>stable diffusion
VibratorGOD good luck with your 1024px sloppa

>>108942788
neat
>oh that video of [important person] fucking a kid?
>actually it's AI generated
>no we're not gonna tell you HOW we know
>oh actually see this video of YOU fucking that kid, that's actually for real!

>>108942745
this is great, an actual anti-cp and anti-deepfake measure.

>>108942761
No free sharing or publication of media. No unaudited trail of media custody. Every piece of content on the internet must be submitted to a cloud server for AI pre-approval. If there isn't a kosher label approved by a private contractor working for a government, it gets sent to the memory hole.

okay? I keep my porn local

>>108950820
You’re trans

>>108942788
Every single method they have been proposing for decades and into the future for this shit will not work. Cat and mouse arms race where anyone with half a brain can beat their new trick

>>108942921
You don’t understand, when the real world is THIS gay and retarded, conspiracies actually come across as the more simple, believable possibility. People WANT to be tricked

oh its just sloppers whining that their slop can now be automatically filtered out

>pedos thought they were going to be able to make child porn with AI and get off scot free
>turns out that we are going to use AI to round them up into camps and sterilise them instead

holy fucking based, thank you sam altman

>>108953277
they would never do that because they would catch more billionaires than regular people.

>>108949579
kek this entire thread should just be locked. This image alone proves this is just snake oil to scam clueless investors. This is like the fake "AI POISONING" patterns artists put on their drawings. It's a security blanket that even a regular human can bypass.
>>108949666
>Doesn't work, even if you point a video camera at a screen they can still get the watermark out.
Satan you are a colorblind retard >>108949579

>>108942825
AIBros, our response?

>>108949693
My understanding is that the media file they get is in thousands of segments, each watermarked, and the random combination of these watermarks in each segment is what gives it a unique ID. Even if they downloaded twice it wouldn't be a simple matter of diffing and only including what's in common, then.
>>108953318
>Satan you are a colorblind retard >>108949579
I haven't actually looked much at synthid, what I'm talking about is what streaming services already use and have patented.

>>108942921
zionist jidf hands typed this

>>108949307
as fragile as taking out the power grid

>>108953308
billionaires getting prime 17 year old pussy is not the same as you jerking off to 10 year olds

>>108952906
You're too poor to pay for Nano Banana Pro and ChatGPT Image 2.0

>>108949289
>the text repo
how long until someone makes a botnigger detector user script lol

>>108950571
AIGODS WON DOING REAL SCIENCE ON /g/ >>108932832 Well done VibeGOD

File: arkham origins console.png (1.11 MB, 1303x885)

1.11 MB PNG

>>108950037
>>108947694
>>108947915
I think you can just feed the same input to a deterministic modification with/without the watermark then get differences and extract features. I'm sure it will stand out enough to at least change the watermark.

>>108953379
>the random combination of these watermarks in each segment is what gives it a unique ID
so then, if the combination is changed it would invalidate the watermark. It would still work as a "is ai" or "isn't ai" watermark, but would not be able to contain identifiable information.

>>108949218
It doesn't matter if Google finds "a watermark". What matters is if the watermark is unchanged. At least for embedding identifiable information.

>>108942808
>It's watermarking text too so there is probably some fuckery involved
unlikely if people are editing the text or only copying what is required. they could get away with it using unicode but not ascii.

>>108942917
or... you could just buy a powerful computer, download python and some datasets and use your own local models to do anything you like?

I think you fags are missing the forest for the trees here. If they can place invisible watermarks into AI images and shit, what stops them from doing the same to any other non-AI images or videos as well? Sounds like shit that can be used to track down pirated media or to just use these watermarks to transfer other kinds of information without the knowledge or consent of intermidiaries, which can also be used to frame people like one of the posts above mentioned.

>>108942745
This is good, though, for far too long, people have been able to be anonymous online, and we saw the effects of that. This is the best compromise instead of internet IDs, as long as you don't share anything, you shouldn't be fine.

>>108955235
>It would still work as a "is ai" or "isn't ai" watermark, but would not be able to contain identifiable information
I'm not talking about synthid but about streaming service watermarking. You're right though, I believe they burn keys by detecting when someone is downloading rather than playing back, not by watermarks themselves.

>>108956235
>which can also be used to frame people like one of the posts above mentioned
Why do the need this to frame you? If the authorities are very corrupt what would stop them from now finding that you have shared CP on your IP and that your laptop contained 100TB of CP so swears FBI honorable agent TrustMeBro and his colleagues.

>>108942817
What if I put paper over my screen and hand trace the image? Will the watermark show up in my hand drawing? Damn Google must be gods if so

>>108953456
>billionaires getting prime 17 year old pussy is not the same as you jerking off to 10 year olds
The new generation only hates age gaps, not pedo-philia per se. A 50 year old fucking a 20 year old is more ick inducing for them than an 18 year old fucking a 10 year old.

>>108942745
I don't know the specifics but if the watermark is getting generated on the moment of the AI's output I don't see how they can put it on anything else

>>108942788
Does anyone honestly believe people wouldn't be able figure the identifier out?

Never underestimate a proud AI artist's fear in being called out for using AI.

>>108956235
>If they can place invisible watermarks into AI images and shit, what stops them from doing the same to any other non-AI images or videos as well?
exactly, picture this scenario.
ai generated content and human made content would each receive their own distinct watermark. any content without a watermark would be hidden from search results and blacklisted, since it would be deemed "unverified."
this approach might be ineffective for the first couple of years. however in 10+ years time, virtually every piece of online content will carry a watermark declaring whether it is human made or ai generated.
>how would human made content receive a watermark?
it would be embedded directly into the hardware and software of every device and application capable of producing content, such as phone cameras, adobe programs, and so on.
essentially, the internet would split into two separate realms. ai content and human content, with no way for them to mix.
any removed or manipulated watermark would simply be treated as "unverified."
to enforce this system, laws will be passed requiring websites and platforms to follow this content verification process

>>108956917
>it would be embedded directly into the hardware and software of every device and application capable of producing content,
are they going to start microchipping loose leaf paper and ballpoint pens?