if you tell it to generate the AI image with a black background you can visually see the synthid with a good enough monitor, it's just a repeating fuzzy pattern, nothing special.
I have found great success of getting rid of it by masking every 2nd pixel, regenerating missing pixels and then once again masking every 2nd pixel offset by 1.
Used an off the shelf model to fill in the pixels, but I also exported a depthmap first (before any alternations) and denoised it so generated masked pixels comform to the original content. The result was obviously not 100% perfect, but with more time and a model fine tuned for this specific use-case would be able to remove any kind of ai watermarking without too many issues.
I guess the SynthID-Image paper from Oct 2025[0] was an encoder-decoder for which they tested checking a flag or a 136 bit payload in 512x512 images and the watermark's robustness after various transformations.
Presumably the deployed version is meaningfully different.
You can stuff per-item database unique IDs, user IDs, geohashes, and other nefarious things inside.
We need to protest this LOUDLY.
Our devices are being locked down, we're having attestation and trusted computing forced on us, the internet all over the world is undergoing age verification with full ID verification.
Just because this is on "ai images" today doesn't mean it won't be on all images - screenshots, your camera reel, etc. - in the fullness of time.
This is scary.
These are the tools of 1984. They've been boiling the water slowly, but in the last year things have really started to pick up pace. Please push back. Loudly.
Everyone at Google and OpenAI working on this: WHAT THE FUCK ARE YOU DOING. STOP.
We have laws and mechanisms to prevent revenge porn, CSAM, defamation, etc. They are robust and can be made even stronger. We do not need to sacrifice the security of our privacy and our speech to fight imagined harms when the real danger is turning into an authoritarian society.
The paper references some threat models they considered. They suggest someone might "possess paired information (both original and watermarked content)" and therefore be able to undo watermarking. Presumably it's fairly easy to get identity operations out of image APIs that would result in this situation. I'm not sure that addresses echelon's main concerns though.
The metadata is kept separately from the original data, and is, by design, modifiable and removable.
Watermark, by design, irreversibly modifies the original data, and is, by design, hard to remove without producing detectable artifacts (or rendering the data useless altogether).
Just because something isn't perfect doesn't mean it's not useful. I've already seen posts online that were able to be proven as falsified because someone ran the images through Google for SynthID checks.
> How old is photoshop and why is it exempt?
For one, it's not developed by Google or OpenAI. The barrier to entry to making realistic but deceptive images with Photoshop is far higher than with AI, and there are already techniques that can, imperfectly, be used to detect the use of traditional image editing.
So 999 people that are just making an image need to be DRM'ed so that you might catch the 1 person making "realistic but deceptive" images... like this is some kind of special case of ... internet images.
Stable Diffusion with 10%~15% denoising strength. Done.
I tested the day 1 when Nano Banana Pro was released and it worked. It still works today for Nano Banana 2.
I didn't post this anywhere because I (arrogantly) thought saying it publicly would make the internet worse. But it was pure arrogancy: if I came up with this the first day then of course other millions of programmers did too.
That being said, it'll introduce the typical artifacts from SD models and that might be detected by other methods (or just by zooming in a lot and looking carefully).
Yup, OOC a while back I put together a ComfyUI node that took in a NB image and start with the smallest amount of denoise strength using Flux.1 (but works with any model), then run img2img with a synthid check incrementing denoise in a loop until it was defeated.
Never released it, but it was obvious to most people in the SD community that denoising using a diffusion model was a relatively trivial means to beat most steganographic watermarks.
Aren't these kinds of watermarks easy to remove or distort? Seems like they're only helpful as long as people are relying on them sparingly so it's not worth the effort to circumvent.
If social media platforms started banning images with these watermarks seems like they'd be stripped out overnight.
I’m surprised! I guess I’m being naive but I would imagine you could pass an image to an image model without synthid and have it reconstruct the image in a net new way without the markers. I guess I’m wrong? That’s cool if the watermarks are so deeply ingrained that they persist
As I understand it, they modify the image by applying a special Gaussian noise filter which affects each pixel in the image in subtle (possibly not reversible) ways. The detecting service will look for this noise pattern to flag it, so even a part of the image is enough to know it was generated by AI.
I still don't think there's a single GitHub repo that actually removes real SynthID watermarks from Nano Banana 2/NBPro outputs. Most of them are just some research projects that haven't achieved this. The only methods so far I've seen are weird tricks with transparency/overlaying the original image if you're using edits, and also using a diffusion model to regenerate the NB-generated image at low noise levels, but this also modifies the original.
Right I think that’s why you probably need to start with very low levels of denoising and experiment with different approaches.
Set up as a ComfyUI workflow that does a few things: it tries SDXL, Flux, and a couple of different denoising methods at the lowest possible strength (progressively incrementing) to avoid changing the image too much, while also running a SynthID check each time, and repeating this in a loop until the watermark is essentially gone.
At the same time, you’d probably want to add some kind of threshold based on a perceptual hash aka the maximum perceptual quality difference you’re willing to accept.
FWIW there are a few people in the issues saying that the tool is giving false negatives and the output image gets flagged by the actual Gemini API as having SynthID. Most recently 3 weeks ago without a response.
This one was released a few years ago and still seems unbroken. I'm sure it will be broken at some point, but if you have to wait a year or two from when you make a deepfake until you can post it on Facebook, maybe that's enough. Maybe even a month is enough.
That'd not work with today's technology. No open model's prompt adherence is anywhere remotely close to ChatGPT/NanoBanana. 'remotely' here is a funny understatement, as I don't have a strong enough word in my vocabulary to describe how far the open models are behind the closed ones.
Writing a more detailed description does not make the models stick to it more.
Definitely. I run an entire site built around a series of benchmarks that focus on prompts of increasingly difficult complexity with a focus on adherence, and even the state-of-the-art local models are probably only about thirty percent as good as proprietary models like Gemini 3.1 Flash Image and GPT Image 2.
Comparing Qwen-Image, Flux.2, ZiT, NB2, and gpt-image-2
First they verify whether a picture came from OpenAI, then they'll include subscriber data and geolocation.
Well, they'll finally find out that no one wants to look at AI generated pictures or text. Once they do that, the tool will fail for the public and only work for the government.
It helps significantly in the current moment. A lot of people are lazy and are getting caught quickly by SynthID.
Eventually it won’t matter when image generation is cheap. But few self-host today and few are willing to pay unsubsidized prices, so the vast majority are using the Gemini, OpenAI, and Midjourney. If all 3 adopted SynthID, only a small fraction would use something else.
There should be no way for anyone to track down who posted a political meme, anti-religious message, or any other legally protected speech. This will come back to bite us in the ass if we keep building it.
Soon every image or communication we make will be watermarked if we continue to let this shit seep into the commons. Everything from your phone photos, to your screenshots, to your social media posts.
One day soon Republicans or Democrats or whoever doesn't like your freedoms will use this tech to identify you and control you.
There are laws for harms - CSAM, revenge porn, etc. Social media platforms can identify, ban, and report abusers. The framework of the law can take care of the rest.
Our digital footprint should not be tracked and barcoded.
That's a lot of hyperbole, there's no cause/effect relationship I can think of here that could realistically produce your slippery slope.
Google or anyone else could start adding those unique tracking watermarks you're concerned about any time they want, regardless of whether they use this AI detection watermark, that to be clear can not track you in any way.
> That's a lot of hyperbole, there's no cause/effect relationship I can think of here that could realistically produce your slippery slope.
Have you been watching the headlines over the last year? It's like there's a global push towards locked down and verified computing (age verification, TPMs everywhere, Captchas that only work on non-rooted phones, ...).
You can look out the window and see movement in this direction happening right now. Governments and corporations around the world can't get enough of this shit. Privacy matters, advocating for it is not a "slippery slope."
> this AI detection watermark [...] that to be clear can not track you in any way.
Is that clear? We have no idea what metadata they are or aren't embedding in SynthID.
> Google or anyone else could start adding those unique tracking watermarks you're concerned about any time they want,
The point is that this is bad and should be denounced!
All you need to do is encode a database identifier, GeoIP, or other identifying information, and you've violated a person's privacy without their knowledge or consent.
Once these systems become popular, the intelligence agencies will "suggest" that Google adds it to their phone cameras. It will start seeping into everything.
The "slippery slope" is not a fallacy. We're on the verge of having device attestation and identity verification to use the internet. This is so beyond fucked.
I'd built an on-device app for detecting C2PA and IPTC metadata in images, amongst other things. I might be able to add support for SynthID detection once it's been reverse engineered.
I'm annoyed that Google is keeping it closed-sourced and limited to partners. Is there a negative externality about open-sourcing image watermark technology so anyone can use it and audit the watermarks independently? If not, then I may have a repository for an open-source invisible and tamper-resistant image watermarking approach that's feature complete...
I assume a selfish benefit is that OpenAI and Google don't want the models to train on their own data. There is just /so much/ AI generated content online that they definitely need to filter it out somehow when assembling the training data. This is a pretty effective way to do that, with the nice bonus of being mostly good from a PR standpoint.
if you tell it to generate the AI image with a black background you can visually see the synthid with a good enough monitor, it's just a repeating fuzzy pattern, nothing special.
I have found great success of getting rid of it by masking every 2nd pixel, regenerating missing pixels and then once again masking every 2nd pixel offset by 1.
Used an off the shelf model to fill in the pixels, but I also exported a depthmap first (before any alternations) and denoised it so generated masked pixels comform to the original content. The result was obviously not 100% perfect, but with more time and a model fine tuned for this specific use-case would be able to remove any kind of ai watermarking without too many issues.
But why
What information is included in the metadata or SynthID? How many bits can be encoded in a SynthID?
Can it be used to create something like nutritional labels for synthetic content? 10% synthetic text, 30 synthetic images.
Your reality was 15% synthetic today (75% mega corp, 25% open-weight neocloud).
I guess the SynthID-Image paper from Oct 2025[0] was an encoder-decoder for which they tested checking a flag or a 136 bit payload in 512x512 images and the watermark's robustness after various transformations.
Presumably the deployed version is meaningfully different.
[0]:https://arxiv.org/html/2510.09263v1
This is very similar to audiowmark
https://github.com/swesterfeld/audiowmark
You can stuff per-item database unique IDs, user IDs, geohashes, and other nefarious things inside.
We need to protest this LOUDLY.
Our devices are being locked down, we're having attestation and trusted computing forced on us, the internet all over the world is undergoing age verification with full ID verification.
Just because this is on "ai images" today doesn't mean it won't be on all images - screenshots, your camera reel, etc. - in the fullness of time.
This is scary.
These are the tools of 1984. They've been boiling the water slowly, but in the last year things have really started to pick up pace. Please push back. Loudly.
Everyone at Google and OpenAI working on this: WHAT THE FUCK ARE YOU DOING. STOP.
We have laws and mechanisms to prevent revenge porn, CSAM, defamation, etc. They are robust and can be made even stronger. We do not need to sacrifice the security of our privacy and our speech to fight imagined harms when the real danger is turning into an authoritarian society.
Most cameras already produce metadata. You can remove this metadata. Can you not also detect and remove watermarks?
The paper references some threat models they considered. They suggest someone might "possess paired information (both original and watermarked content)" and therefore be able to undo watermarking. Presumably it's fairly easy to get identity operations out of image APIs that would result in this situation. I'm not sure that addresses echelon's main concerns though.
The metadata is kept separately from the original data, and is, by design, modifiable and removable.
Watermark, by design, irreversibly modifies the original data, and is, by design, hard to remove without producing detectable artifacts (or rendering the data useless altogether).
In short, the answer is no.
Don’t think that would be possible. If I paste a synthetic piece into an otherwise organic image, the synth id isn’t going to know that.
Synth ID can detect parts of images with the watermark.
This is just performative nonsense.
As someone that creates things with tools with different media I would just hard avoid this tool that adds...
arbitrary metadata not of my choosing.
Should I seriously make a texture for a videogame with this weird DRM glorp in it?
How old is photoshop and why is it exempt?
Just because something isn't perfect doesn't mean it's not useful. I've already seen posts online that were able to be proven as falsified because someone ran the images through Google for SynthID checks.
> How old is photoshop and why is it exempt?
For one, it's not developed by Google or OpenAI. The barrier to entry to making realistic but deceptive images with Photoshop is far higher than with AI, and there are already techniques that can, imperfectly, be used to detect the use of traditional image editing.
So 999 people that are just making an image need to be DRM'ed so that you might catch the 1 person making "realistic but deceptive" images... like this is some kind of special case of ... internet images.
>How old is photoshop and why is it exempt?
I'm sure you can think of a couple things that differentiate gen AI from photoshop, I believe in you.
The main difference is we are in the middle of a moral panic and people have lost perspective.
Its a tool with different modalties and affordances.
Strictly speaking, DRM = digital rights management, which is related to intellectual property.
SynthID would only be DRM if Google/OpenAI were claiming IP rights over their images. I don’t even know if that’s legal though.
What value does "strictly speaking" bring to the discussion?
So that you don't have to address any of the issues?
> How old is photoshop and why is it exempt?
How does today’s maximum theoretical disinformation output per minute compare to 2021 Photoshop?
Its 2026... people are deliberately choosing to live in their own realities with no care about objective facts or moral choices.
So weird images are a big problem? No they don't matter at all.
Political deepfakes on the mind here more than weird stuff.
"I could stand in the middle of 5th Avenue and shoot somebody and wouldn't lose any voters, ok? It's, like, incredible." — Donald Trump
So what does a deepfake matter?
Good. Despite people saying it will be removed, I have seen no reproducible repo demonstrating it.
Stable Diffusion with 10%~15% denoising strength. Done.
I tested the day 1 when Nano Banana Pro was released and it worked. It still works today for Nano Banana 2.
I didn't post this anywhere because I (arrogantly) thought saying it publicly would make the internet worse. But it was pure arrogancy: if I came up with this the first day then of course other millions of programmers did too.
That being said, it'll introduce the typical artifacts from SD models and that might be detected by other methods (or just by zooming in a lot and looking carefully).
Yup, OOC a while back I put together a ComfyUI node that took in a NB image and start with the smallest amount of denoise strength using Flux.1 (but works with any model), then run img2img with a synthid check incrementing denoise in a loop until it was defeated.
Never released it, but it was obvious to most people in the SD community that denoising using a diffusion model was a relatively trivial means to beat most steganographic watermarks.
Yet is in itself fairly trivial to detect assuming you use some open-weight image model as a base.
> if I came up with this the first day then of course other millions of programmers did too.
Don't sell yourself short. I'm sure it was only hundreds of thousands.
Post a repro. I can do that too but then the similarity index is weak. The point is that it it looks indistinguishable then the integrity persists.
In my tests the image looks clearly distinct. In other words, if you can tell the difference then it isn’t a good test.
Probably a lot easier to use a different model in the first place
It will but many people won't as i've seen disinformation that could be detected by synth-id.
[flagged]
Aren't these kinds of watermarks easy to remove or distort? Seems like they're only helpful as long as people are relying on them sparingly so it's not worth the effort to circumvent.
If social media platforms started banning images with these watermarks seems like they'd be stripped out overnight.
No, they are very resistant to modification that can be done easily. That being said I doubt it is impossible
Yeah cropping, color shifting, resizing and compression don’t remove it. That said, there’s pretty well known workarounds:
https://github.com/wiltodelta/remove-ai-watermarks
I’m surprised! I guess I’m being naive but I would imagine you could pass an image to an image model without synthid and have it reconstruct the image in a net new way without the markers. I guess I’m wrong? That’s cool if the watermarks are so deeply ingrained that they persist
As I understand it, they modify the image by applying a special Gaussian noise filter which affects each pixel in the image in subtle (possibly not reversible) ways. The detecting service will look for this noise pattern to flag it, so even a part of the image is enough to know it was generated by AI.
Yes, Gemini can actually say how much of the image is AI generated.
I still don't think there's a single GitHub repo that actually removes real SynthID watermarks from Nano Banana 2/NBPro outputs. Most of them are just some research projects that haven't achieved this. The only methods so far I've seen are weird tricks with transparency/overlaying the original image if you're using edits, and also using a diffusion model to regenerate the NB-generated image at low noise levels, but this also modifies the original.
Right I think that’s why you probably need to start with very low levels of denoising and experiment with different approaches.
Set up as a ComfyUI workflow that does a few things: it tries SDXL, Flux, and a couple of different denoising methods at the lowest possible strength (progressively incrementing) to avoid changing the image too much, while also running a SynthID check each time, and repeating this in a loop until the watermark is essentially gone.
At the same time, you’d probably want to add some kind of threshold based on a perceptual hash aka the maximum perceptual quality difference you’re willing to accept.
Define easily. There is an approach that apparently works and is based on spectral analysis of the images.
https://github.com/aloshdenny/reverse-SynthID
FWIW there are a few people in the issues saying that the tool is giving false negatives and the output image gets flagged by the actual Gemini API as having SynthID. Most recently 3 weeks ago without a response.
This one was released a few years ago and still seems unbroken. I'm sure it will be broken at some point, but if you have to wait a year or two from when you make a deepfake until you can post it on Facebook, maybe that's enough. Maybe even a month is enough.
I imagine the technique of having AI recreate the image from scratch based on a very detailed description might work.
That'd not work with today's technology. No open model's prompt adherence is anywhere remotely close to ChatGPT/NanoBanana. 'remotely' here is a funny understatement, as I don't have a strong enough word in my vocabulary to describe how far the open models are behind the closed ones.
Writing a more detailed description does not make the models stick to it more.
Definitely. I run an entire site built around a series of benchmarks that focus on prompts of increasingly difficult complexity with a focus on adherence, and even the state-of-the-art local models are probably only about thirty percent as good as proprietary models like Gemini 3.1 Flash Image and GPT Image 2.
Comparing Qwen-Image, Flux.2, ZiT, NB2, and gpt-image-2
https://genai-showdown.specr.net/?models=qi,nbp3,f2d,g2,zt
First they verify whether a picture came from OpenAI, then they'll include subscriber data and geolocation.
Well, they'll finally find out that no one wants to look at AI generated pictures or text. Once they do that, the tool will fail for the public and only work for the government.
While these are great, isn’t the problem that malicious actors will create systems that do not use synthID
It helps significantly in the current moment. A lot of people are lazy and are getting caught quickly by SynthID.
Eventually it won’t matter when image generation is cheap. But few self-host today and few are willing to pay unsubsidized prices, so the vast majority are using the Gemini, OpenAI, and Midjourney. If all 3 adopted SynthID, only a small fraction would use something else.
These systems should be removed.
This is antithetical to freedom and privacy.
There should be no way for anyone to track down who posted a political meme, anti-religious message, or any other legally protected speech. This will come back to bite us in the ass if we keep building it.
Soon every image or communication we make will be watermarked if we continue to let this shit seep into the commons. Everything from your phone photos, to your screenshots, to your social media posts.
One day soon Republicans or Democrats or whoever doesn't like your freedoms will use this tech to identify you and control you.
There are laws for harms - CSAM, revenge porn, etc. Social media platforms can identify, ban, and report abusers. The framework of the law can take care of the rest.
Our digital footprint should not be tracked and barcoded.
That's a lot of hyperbole, there's no cause/effect relationship I can think of here that could realistically produce your slippery slope.
Google or anyone else could start adding those unique tracking watermarks you're concerned about any time they want, regardless of whether they use this AI detection watermark, that to be clear can not track you in any way.
> That's a lot of hyperbole, there's no cause/effect relationship I can think of here that could realistically produce your slippery slope.
Have you been watching the headlines over the last year? It's like there's a global push towards locked down and verified computing (age verification, TPMs everywhere, Captchas that only work on non-rooted phones, ...).
You can look out the window and see movement in this direction happening right now. Governments and corporations around the world can't get enough of this shit. Privacy matters, advocating for it is not a "slippery slope."
> this AI detection watermark [...] that to be clear can not track you in any way.
Is that clear? We have no idea what metadata they are or aren't embedding in SynthID.
> Google or anyone else could start adding those unique tracking watermarks you're concerned about any time they want,
The point is that this is bad and should be denounced!
I'm not going to mince words - what you're saying is dangerous and harmful.
> to be clear can not track you in any way
All they have to do is encode enough entropy for a database unique identifier. Systems like this have been used to do it for audio:
https://github.com/swesterfeld/audiowmark
SynthID payloads work the same way, and the paper discusses encoding a "user identifier":
https://arxiv.org/html/2510.09263v1#S5
All you need to do is encode a database identifier, GeoIP, or other identifying information, and you've violated a person's privacy without their knowledge or consent.
Once these systems become popular, the intelligence agencies will "suggest" that Google adds it to their phone cameras. It will start seeping into everything.
The "slippery slope" is not a fallacy. We're on the verge of having device attestation and identity verification to use the internet. This is so beyond fucked.
Stop defending this.
Saying this is okay is EVIL.
What if they use advanced evasion techniques like printing it out and scanning it or taking a photo with their phone?
SynthID is fairly resistant to this sort of thing, although not perfect.
Is there no way to do this without uploading it?
I'd built an on-device app for detecting C2PA and IPTC metadata in images, amongst other things. I might be able to add support for SynthID detection once it's been reverse engineered.
Currently, there is not. OpenAI has promised "public verification tooling" down the line, but I'll believe it when I see it.
What happens if you generate an image with only a single pixel color or say two colors?
This was done in the past, Google saw it, and now either refuses to generate or doesn't emit the SynthID watermark for those images
You waste a lot of compute on overhead?
I'm annoyed that Google is keeping it closed-sourced and limited to partners. Is there a negative externality about open-sourcing image watermark technology so anyone can use it and audit the watermarks independently? If not, then I may have a repository for an open-source invisible and tamper-resistant image watermarking approach that's feature complete...
potentially to stop bad actors from poisoning datasets by just adding the filter to real pictures?
might be easier to strip it?
so ? people wanting to make AI propaganda will just make tool to remove it. Possibly using AI to do it too
I assume a selfish benefit is that OpenAI and Google don't want the models to train on their own data. There is just /so much/ AI generated content online that they definitely need to filter it out somehow when assembling the training data. This is a pretty effective way to do that, with the nice bonus of being mostly good from a PR standpoint.
[dead]
always trust a vibe-coded website which uses a Discord bot as its backend
(i'm sure there are countless bypasses out there, but please don't use something like this)
You can tell the difference in the example with your bare eyes lol
Why does this matter?