SynthID seems interesting but in classic Google fashion, I haven't a clue on how to use it and the only button that exists is join a waitlist. Apparently it's been out since 2023? Also, does SynthID work only within gemini ecosystem? If so, is this the beginning of a slew of these products with no one standard way? i.e "Have you run that image through tool1, tool2, tool3, and tool4 before deciding this image is legit?"
edit: apparently people have been able to remove these watermarks with a high success rate so already this feels like a DOA product
I wouldn't trust any of the info in those images in the first carousel if I found them in the wild. It looks like AI image slop and I assume anyone who thinks those look good enough to share did not fact check any of the info and just prompted "make an image with a recipe for X"
> Generate better visuals with more accurate, legible text directly in the image in multiple languages
Assuming that this new model works as advertised, it's interesting to me that it took this long to get an image generation model that can reliably generate text. Why is text generation in images so hard?
Interesting they didn’t post any benchmark results - lmarena/artificial analysis etc. I would’ve thought they’d be testing it behind the scenes the same way they did with Gemini 3.
I wonder how hard it is to remove that SynthID watermark...
Looks like: "When tested on images marked with Google’s SynthID, the technique used in the example images above, Kassis says that UnMarker successfully removed 79 percent of watermarks." From https://spectrum.ieee.org/ai-watermark-remover
Adobe's stock is down 50% from last year's peak. It's humbling and scary that entire industries with millions of jobs evaporate in a matter of few years.
The interesting tidbit here is SynthID. While a good first step, it doesn't solve the problem of AI generated content NOT having any kind of watermark. So we can prove that something WITH the ID is AI generated but we can't prove that something without one ISN'T AI generated.
Like it would be nice if all photo and video generated by the big players would have some kind of standardized identifier on them - but now you're left with the bajillion other "grey market" models that won't give a damn about that.
> now you're left with the bajillion other "grey market" models that won't give a damn about that.
Exactly. When the barrier to entry for training a okay-ish AI model (not SOTA, obviously) is only a few thousand compute hours on H100s, you couldn't possibly hope to police the training of 100% of new models. Not to mention that lots of existing models are already out there are fully open-source. There will always be AI models that don't adhere to watermark regulations, especially if they were created a country that doesn't enforce your regulations.
You can't hope to solve the problem of non-watermarked AI completely. And by solving it partially by mandating that the big AI labs add a unified watermark, you condition people to be even more susceptible to AI images because "if it was AI, it would have a watermark". It's truly a no-win situation.
I tried the studio ghibli prompt on a photo my me and my wife in Japan and it was... not good. It looked more like a hand drawn sketch made with colored pencils, but none of the colors were correct. Everything was a weird shade of yellow/brown.
This has been an oddly difficult benchmark for Gemini's NB models. Googles images models have always been pretty bad at the studio ghibli prompt, but I'm shocked at how poorly it performs at this task still.
The SynthID check for fishy photos is a step in the right direction, but without tighter integration into everyday tooling its not going to move the needle much. Like when I hold the power button on my Pixel 9, It would be great if it could identify synthetic images on the screen before I think to ask about it. For what its worth it would be great if the power button shortcut on Pixel did a lot more things.
The rollout doesn't seem to have reached my userid yet. How successful are people at getting these things to actually produce useful images? I was trying recently with the (non-Pro) Nano Banana to see what the fuss was about. As a test case, I tried to get it to make a diagram of a zipper merge (in driving), using numbered arrows to indicate what the first, second, third, etc. cars should do.
I had trouble reliably getting it to...
* produce just two lanes of traffic
* have all the cars facing the same way—sometimes even within one lane they'd be facing in opposite directions.
* contain the construction within the blocked-off area. I think similarly it wouldn't understand which side was supposed to be blocked off. It'd also put the lane closure sign in lanes that were supposed to be open.
* have the cars be in proportion to the lane and road instead of two side-by-side within a lane.
* have the arrows go in the correct direction instead of veering into the shoulder or U-turning back into oncoming traffic
* use each number once, much less on the correct car
This is consistent with my understanding of how LLMs work, but I don't understand how you can "visualize real-time information like weather or sports" accurately with these failings.
Below is one of the prompts I tried to go from scratch to an image:
> You are an illustrator for a drivers' education handbook. You are an expert on US road signage and traffic laws. We need to prepare a diagram of a "zipper merge". It should clearly show what drivers are expected to do, without distracting elements.
> First, draw two lanes representing a single direction of travel from the bottom to the top of the image (not an entire two-way road), with a dotted white line dividing them. Make sure there's enough space for the several car-lengths approaching a construction site. Include only the illustration; no title or legend.
> Add the construction in the right lane only near the top (far side). It should have the correct signage for lane closure and merging to the left as drivers approach a demolished section. The left lane should be clear. The sign should be in the closed lane or right shoulder.
> Add cars in the unclosed sections of the road. Each car should be almost as wide as its lane.
> Add numbered arrows #1–#5 indicating the next cars to pass to the left of the "lane closed" sign. They should be in the direction the cars will move: from the bottom of the illustration to the top. One car should proceed straight in the left lane, then one should merge from the right to the left (indicate this with a curved arrow), another should proceed straight in the left, another should merge, and so on.
I did have a bit better luck starting from a simple image and adding an element to it with each prompt. But on the other hand, when I did that it wouldn't do as well at keeping space for things. And sometimes it just didn't make any changes to the image at all. A lot of dead ends.
I also tried sketching myself and having it change the illustration style. But it didn't do it completely. It turned some of my boxes into cars but not necessarily all of them. It drew a "proper" lane divider over my thin dotted line but still kept the original line. etc.
I've tried to repaint the exterior of my house. More than 20 times with very detailed prompts. I even tried to optimize it with Claude. No matter what, every time it added one, two or three extra windows to the same wall.
I'll be running it through my GenAI Comparison benchmark shortly - but so far it seems to be failing on the same tests that the original Nano Banana struggled with (such as SHRDLU).
119 comments
[ 566 ms ] story [ 414 ms ] threadDeepMind Page: https://deepmind.google/models/gemini-image/pro/
Model Card: https://storage.googleapis.com/deepmind-media/Model-Cards/Ge...
SynthID in Gemini: https://blog.google/technology/ai/ai-image-verification-gemi...
edit: apparently people have been able to remove these watermarks with a high success rate so already this feels like a DOA product
(The Gemini 3 post has a million comments too many to ask this now)
Assuming that this new model works as advertised, it's interesting to me that it took this long to get an image generation model that can reliably generate text. Why is text generation in images so hard?
Looks like: "When tested on images marked with Google’s SynthID, the technique used in the example images above, Kassis says that UnMarker successfully removed 79 percent of watermarks." From https://spectrum.ieee.org/ai-watermark-remover
> Rolling out globally in the Gemini app
wanna be any more vague? is it out or not? where? when?
Like it would be nice if all photo and video generated by the big players would have some kind of standardized identifier on them - but now you're left with the bajillion other "grey market" models that won't give a damn about that.
And if it can be seen like that, it should be removeable too. There are more examples in that thread.
Exactly. When the barrier to entry for training a okay-ish AI model (not SOTA, obviously) is only a few thousand compute hours on H100s, you couldn't possibly hope to police the training of 100% of new models. Not to mention that lots of existing models are already out there are fully open-source. There will always be AI models that don't adhere to watermark regulations, especially if they were created a country that doesn't enforce your regulations.
You can't hope to solve the problem of non-watermarked AI completely. And by solving it partially by mandating that the big AI labs add a unified watermark, you condition people to be even more susceptible to AI images because "if it was AI, it would have a watermark". It's truly a no-win situation.
The inline verification of images following the prompt is awesome, and you can do some _amazing_ stuff with it.
It's probably not as fun anymore though (in the early access program, it doesn't have censoring!)
This has been an oddly difficult benchmark for Gemini's NB models. Googles images models have always been pretty bad at the studio ghibli prompt, but I'm shocked at how poorly it performs at this task still.
I had trouble reliably getting it to...
* produce just two lanes of traffic
* have all the cars facing the same way—sometimes even within one lane they'd be facing in opposite directions.
* contain the construction within the blocked-off area. I think similarly it wouldn't understand which side was supposed to be blocked off. It'd also put the lane closure sign in lanes that were supposed to be open.
* have the cars be in proportion to the lane and road instead of two side-by-side within a lane.
* have the arrows go in the correct direction instead of veering into the shoulder or U-turning back into oncoming traffic
* use each number once, much less on the correct car
This is consistent with my understanding of how LLMs work, but I don't understand how you can "visualize real-time information like weather or sports" accurately with these failings.
Below is one of the prompts I tried to go from scratch to an image:
> You are an illustrator for a drivers' education handbook. You are an expert on US road signage and traffic laws. We need to prepare a diagram of a "zipper merge". It should clearly show what drivers are expected to do, without distracting elements.
> First, draw two lanes representing a single direction of travel from the bottom to the top of the image (not an entire two-way road), with a dotted white line dividing them. Make sure there's enough space for the several car-lengths approaching a construction site. Include only the illustration; no title or legend.
> Add the construction in the right lane only near the top (far side). It should have the correct signage for lane closure and merging to the left as drivers approach a demolished section. The left lane should be clear. The sign should be in the closed lane or right shoulder.
> Add cars in the unclosed sections of the road. Each car should be almost as wide as its lane.
> Add numbered arrows #1–#5 indicating the next cars to pass to the left of the "lane closed" sign. They should be in the direction the cars will move: from the bottom of the illustration to the top. One car should proceed straight in the left lane, then one should merge from the right to the left (indicate this with a curved arrow), another should proceed straight in the left, another should merge, and so on.
I did have a bit better luck starting from a simple image and adding an element to it with each prompt. But on the other hand, when I did that it wouldn't do as well at keeping space for things. And sometimes it just didn't make any changes to the image at all. A lot of dead ends.
I also tried sketching myself and having it change the illustration style. But it didn't do it completely. It turned some of my boxes into cars but not necessarily all of them. It drew a "proper" lane divider over my thin dotted line but still kept the original line. etc.
Not just are they making slop machines, they seem to be run by them.
I am too old for this shit.
https://genai-showdown.specr.net/image-editing