Hacker News Reader

Comment by samiwami | Hacker News Reader

do they have anything similar to SynthID, or are they just pretending that problem doesn't exist?

I know this is probably mega cherry-picked to look more impressive, but some of the images are terrifyingly realistic. They seem to have put a lot of effort into the lighting.

View on HN

> Integrating an imperceptible, robust, and content-specific watermark

From the system card someone linked elsewhere in the discussion

Zhao et al. 2023 showed any imperceptible watermark is provably removable by generative regeneration: pass the image through an img2img or VAE, the model reconstructs it visually identical but starts from a different latent. Watermark gone. SynthID and similar schemes do hold up well against normal sharing: recompression, crops, color tweaks, Twitter's pipeline. That covers most users. But the asymmetry is stuck — normally a GPU and a bit of motivation should be enough to strip it. Right? Got a tool to share? ;-)

I think they do, and it's even scarier than SynthID. Look closely at the generated images and you'll start to notice some diamond-shaped high-frequency patterns in different parts of the image (tree crowns, vegetation, hair, textures, clouds). May be just a random bug, but for me it looks like the watermark is baked into the image details themselves and all the details follow this pattern. But it leads to unnatural results somewhere, i.e.:

https://x.com/rrnld_y/status/2047070630802006211/photo/1

https://x.com/Melothemyth777/status/2046963312357679540/phot...

https://x.com/kuroinu_ni/status/2047118826920440287/photo/1

https://x.com/jahflyx/status/2047109303140536703/photo/1

https://x.com/2016Suha/status/2047096528309760501/photo/3

> do they have anything similar to SynthID, or are they just pretending that problem doesn't exist?

At least they aren't pretending that a solution exists.

I think we are just going to have to accept that realistic images can be easily fabricated now.

Seeing is not believing anymore, and I don't think SynthID or anything like it can restore that trust in images.

Hopefully the arms race will balance out with improved AI image detection, but I can see how that will never be guaranteed to be reliable.

It's going to mess up accountability.

Some politician will be recorded doing something & he'll have his people release a thousand photos/videos of him doing crimes. And they'll say, look, it's a smear campaign.

This is just one stupid example, but people will have better schemes.

Also global coordinated releases of fake content and hypertargeted possibly abusive content. Virtual kidnappings will take off, automated & scaled.

Some politician will be recorded doing something & he'll have his people release a thousand photos/videos of him doing crimes. And they'll say, look, it's a smear campaign.

And his enemies will do the same, hopefully resulting in less blind trust for everyone in the population, which can only be a good thing.

I would’ve paused image models for now until we’ve better educated our less-savvy neighbors.

Maybe a stupid question, but does the SynthID still exist if you screenshot and crop your generated image? What if you screenshot, rotate _just_ a bit, and crop? Or apply some other effect to the image like adjusting the coloring a little bit, adding some blur, etc.

The paper they published last year goes over some of these transformations: https://arxiv.org/pdf/2510.09263

I feel like asking the image generators to mark AI images is the wrong way to go about it. It's like trying to maintain a blocklist. It seems better to me to have the major camera manufacturers or cell phones cryptographically sign their images as real.

I feel like this idea comes up often and in my opinion it doesn't solve anything. Take a picture of an AI image and you've made this approach useless. Which then goes to the argument of "well you'll see it's a picture of a picture" to which I will say there are plenty of ways to make this not appear so, and the ultimate form of this argument is that you can eventually project light directly into the photosensors, or otherwise hack the input between the photosensors and the rest of whatever digital magic that turns light into a JPG on your phone.

SynthID survives basic transforms including screenshots/photos, although it can of course be defeated. Even still it helps with the laziest fakes, which there seem to be a lot of - I've seen several quite widespread misinformative images over the past couple months that failed a synthID check.

Anyways I think approaching the problem from both directions is probably good.

Reply on news.ycombinator.com