Google has been stomping around like Godzilla this week, and this is the first time I decided to link my card to their AI studio.
I had seen people saying that they gave up and went to another platform because it was "impossible to pay". I thought this was strange, but after trying to get a working API key for the past half hour, I see what they mean.
Everything is set up, I see a message that says "You're using Paid API key [NanoBanano] as part of [NanoBanano]. All requests sent in this session will be charged." Go to prompt, and I get a "permission denied" error.
There is no point in having impressive models if you make it a chore for me to -give you my money-
First off, apologies for the bad first impression, the team is pushing super hard to make sure it is easy to access these models.
- On permission issue, not sure I follow the flow that got you there, pls email me more details if you are able too and happy to debug: Lkilpatrick@google.com
- On overall friction for billing: we are working on a new billing experience built right into AI Studio that will make it super easy to add a CC and go build. This will also come along with things like hard billing caps and such. The expected ETA for global rollout is January!
If it's just the API you're interested in, Fal.ai has put Nano-Banana-Pro up for both generative and editing. A great deal less annoying to sign up for them since they're a pretty generalized provider of lots of AI related models.
In general a better option, in the early days of AI video I tried to generate a video of a golden retriever using Google's AI Studio. It generated 4 in the highest quality and charged me 36 bucks. Not a crazy amount but definitely an unwelcome suprise.
Fal.ai is pay as you go and has the cost right upfront.
There's the solution right there. Google is still growing its AI "sea legs". They've turned the ship around on a dime and things are still a little janky. Truly a "startup mode" pivot.
While we're on this subject of "Google has been stomping around like Godzilla", this is a nice place to state that I think the tide of AI is turning and the new battle lines are starting to appear. Google looks like it's going to lay waste to OpenAI and Anthropic and claim most of the market for itself. These companies do not have the cash flow and will have to train and build their asses off to keep up with where Google already is.
gpt-image-1 is 1/1000th of Nano Banana Pro and takes 80 seconds to generate outputs.
Two years ago Google looked weak. Now I really want to move a lot of my investments over to Google stock.
How are we feeling about Google putting everyone out of work and owning the future? It's starting to feel that way to me.
(FWIW, I really don't like how much power this one company has and how much of a monopoly it already was and is becoming.)
100% this. I am using the pro/max plans on both claude and openai. Would love to experiment with gemini but paying is next to impossible. Why do i need the risk of a full blown gcp project just to test gemini. No thx.
It wasn't there when I first went to Gemini after the announcement, but upon revisiting it gave me the prompt to try Nano Banana Pro. It failed at my niche (rare palm trees).
Incredible technology, don't get me wrong, but still shocked at the cumbersome payment interface and annoyed that enabling Drive is the only way to save.
I hate that they kinda try to hide the model version. Like if you click the dropdown in the chat box, you can see that "Thinking" means 3 Pro. When you select the "Create images" tool, it doesn't tell you it's using Nano Banana Pro until it actually starts generating the image.
Tell me the model it's using. It's as if Google is trying to unburden me with the knowledge of what model does what but it's just making things more confusing.
Oh, and setting up AI Studio is a mess. First I have to create a project. Then an API key. Then I have to link the API key to the project. Then I have to link the project to the chat session... Come on, Google.
Alright results are in! I've re-run all my editing based adherence related prompts through Nano Banana Pro. NB Pro managed to successfully pass SHRDLU, the M&M Van Halen test (as verified independently by Simon), and the Scorpio street test - all of which the original NB failed.
Definitely! Even though NB's predominant use case seems to be editing, it's still producing surprisingly decent text-to-image results. Imagen4 currently still comes out ahead in terms of image fidelity, but I think NB Pro will close the gap even further.
I'll try to have the generative comparisons for NB Pro up later this afternoon once I catch my breath.
I'll add the new output resolutions and other features ASAP. However, looking at the pricing (https://ai.google.dev/gemini-api/docs/pricing#standard_1), I'm definitely not changing the default model to Pro as $0.13 per 1k/2k output will make it a tougher sell.
> The model generates up to two interim images to test composition and logic. The last image within Thinking is also the final rendered image.
Maybe that's partially why the cost is higher: it's hard to tell if intermediate images are billed in addition to the output. However, this could cause an issue with the base gemimg and have it return an intermediate image instead of the final image depending on how the output is constructed, so will need to double-check.
>> - Put a strawberry in the left eye socket.
>>- Put a blackberry in the right eye socket.
>> All five of the edits are implemented correctly
This is a GREAT example of the (not so) subtle mistakes AI will make in image generation, or code creation, or your future knee surgery. The model placed the specified items in the eye sockets based on the viewers left/right; when we talk relative in this scenario we usually (always?) mean from the perspective of the target or "owner". Doctors make this mistake too (they typically mark the correct side with a sharpie while the patient is still alert) but I'd be more concerned if we're "outsourcing" decision making without adequate oversight.
There's a classic well-illustrated book, _How to Keep Your Volkswagen Alive_, which spends a whole illustrated page at the beginning building up a reference frame for working on the vehicle. Up is sky, down is ground, front is always vehicle's front, left is always vehicle's left.
Sounds a bit silly to write it out, but the diagram did a great job removing ambiguity when you expect someone to be laying on the ground in a tight place looking backwards, upside down.
Also feels important to note that in the theatre, there is stage-right and stage-left, jargon to disambiguate even though the jargon expects you to know the meaning to understand it.
It doesn't affect your point but technically since the IAU are insane, exoplanets aren't technically planets and Jupiter is the largest planet in the universe.
Seems like you're making a judgment based on your own experience, but as another commenter pointed out, it was wrong. There are plenty of us out there who would confirm, because people are too flawed to trust. Humans double/triple check, especially under higher stakes conditions (surgery).
Heck, humans are so flawed, they'll put the things in the wrong eye socket even knowing full well exactly where they should go - something a computer literally couldn't do.
Intelligence in my book includes error correction. Questioning possible mistakes is part of wisdom.
So the understanding that AI and HI are different entities altogether with only a subset of communication protocols between them will become more and more obvious, like some comments here are already implicitly telling.
If the instructions were actually specific, e.g. Put a blackberry in its right eye socket, then yes, most humans would know what that meant. But the instructions were not that specific: in the right eye socket
If you asked me right now what the biggest known planet was, I'd think Jupiter. I'd assume you were talking about our solar system ("known" here implying there might be more planets out in the distant reaches).
I don't know if that's so much a mistake as it is ambiguity though? To me, using the viewer's perspective in this case seems totally reasonable.
Does it still use the viewer's perspective if the prompt specifies "Put a strawberry in the _patient's left eye_"? If it does, then you're onto something. Otherwise I completely disagree with this.
“The right socket” can only be implied one way when talking about a body just like you only have one right hand despite the fact that it is on my left when looking at you.
If you are facing a wall-plate with two power sockets on it side by side and you are telling someone to plug something in, which one would be "the right socket", and which would be "the left socket"?
If above the wall-plate is a photo of a person and you are someone to draw a tattoo on the photo, which is "the right arm" and which is "the left arm"?
I think "the left eye" in this particular case (a photo of a skull made of pancake batter) is still very slightly ambiguous. "The skull's left eye" would not be.
In case anyone missed Max's Nano Banana prompting guide, it's absolutely the definitive manual for prompting the original Nano Banana... and I tried some of the prompts in there against Nano Banana Pro and found it to be very applicable to the new model as well.
In my experience multimodal models like gpt-image-1/nano/etc. don't really require a lot of prompt trickery [1] like the good ol' days of SD 1.5.
To be clear, that's a good thing though. It's also one of the reasons why "prompt engineering" will become less relevant as model understanding goes up.
[1] - Unless you're trying to circumvent guardrails
How do you know Simon? It's certainly a blog post, with content about prompting in it. If your goal is to make generative art that uses specific IP, I wouldn't use it.
I was going off the footnote of "Image input is set at 560 tokens or $0.067 per image" but 560 * 2 / 1_000_000 is indeed $0.0011 so I have no idea where the $0.067 came from. Fixed, and this is why I typically don't read docs without coffee.
I just pushed gemimg 0.3.2 which adds image_size support for Nano Banana Pro, and I ran a few tests on some of the images in the blog. In my testing, Nano Banana Pro correctly handled most of the image generation errors noted in my blog post: https://x.com/minimaxir/status/1991580127587921971
- Fibonacci magnets: code is correctly indented and the syntax highlighting atleast tries giving variables, numbers, and keywords different colors.
- Make me a Studio Ghibli: actually does style transfer correctly, and does it better than ChatGPT ever did.
- Rendering a webpage from HTML: near-perfect recreation of the HTML, including text layout and element sizing.
That said, there may be regressions where even with prompt engineering, the generated images which are more photorealistic look too good and land back into the uncanny valley. I haven't decided if I'm going to write a follow up blog post yet.
The system prompt hacking trick doesn't work with Nano Banana Pro unfortunately.
> "I...worked on the detailed Nano Banana prompt engineering analysis for months"
Early in four decades of tech innovation I wasted time layering on fixes for clear deficiencies in a snowballing trend's tech offerings. If it's a big enough trend to have well funded competitors, just wait. The concern is likely not unique, and will likely be solved tomorrow.
I realized it's better to learn adaptive/defensive techniques, giving your product resilience to change. Your goal is that when surfing the change waves you can pick a point you like between rock solid and cutting edge and surf there safely.
Invest that "remediate their thing" time in "change resilience" instead – pays dividends from then on. It can be argued your tool is in this camp!
// Getting better at this also helps you with zero days.
> The model generates up to two interim images to test composition and logic. The last image within Thinking is also the final rendered image.
I've been using a bespoke Generative Model -> VLM Validator -> LLM Prompt Modifier REPL as part of my benchmarks for a while now so I'd be curious to see how this stacks up. From some preliminary testing (9 pointed star, 5 leaf clover, etc) - NB Pro seems slightly better than NB though it still seems to get them wrong. It's hard to tell what's happening under the covers.
yes they are pricey but the price will go down over time and then you can switch. vlm.run got access as early customers and are releasing it for free with unlimited generations(till they are bottlenecked by google). some results here combining image gen(Nano Banana pro) with video gen(Veo 3.1) in a single chat https://chat.vlm.run/c/1c726fab-04ef-47cc-923d-cb3b005d6262. This combined the synth generation of a person and made the puppet dance. Quite impressive
It's written to mimic that style but without meaning that the work has been done for them, just that there is new work to be done, making it an odd perhaps unconscious reference
this is pretty cool!
have you found success with image editing in nano banana - i mean photoshop-like stuff.
from your article i seem to wonder if nano banana is good for editing versus generating new images.
That IS the use-case for Nano Banana (as opposed to pure generative like Imagen4).
In my benchmarks, Nano-Banana scores a 7 out of 12. Seedream4 managed to outpace it, but Seedream can also introduce slight tone mapping variations. NB is the gold standard for highly localized edits.
Comparisons of Seedream4, NanoBanana, gpt-image-1, etc.
I tried your "Remove all the brown pieces of candy from the glass bowl." prompt against Nano Banana Pro and it converted them to green, which I think is a pass by your criteria. Original Nano Banana had failed that test because it changed the composition of the M&Ms.
Thanks Simon - I'm in the middle of re-running all my prompts through NB Pro at the moment. Nice to know it's already edged out the original. It also passed the SHRDLU test (swapping colored blocks) without cheating and just changing the colors. I'll have an update to the site shortly!
EDIT: Finished the comparisons. NB Pro scored a few more points than NB which was already super impressive.
Even generating a standard piano with 7 full octaves that are consistent is pretty hard. If you ask it to invert the colors of the naturals and sharps/flats you'll completely break them.
It even worked really well at creating an infographic for one of my quirkier projects which doesn't have that much information online (other than its repo).
Then the question becomes, can it incorporate targeted feedback, or is it a oneshot-or-bust affair?
My experience is that ChatGPT is very good at iterating on text (prose, code) but fairly bad at iterating on images. It struggles to integrate small changes, choosing instead to start over from scratch, with wildly different results. Thinking especially here of architectural stuff, where it does a great job laying out furniture in a room, but when I ask it to keep everything the same but change the colour of one piece, it goes completely off the rails.
You can use targeted feedback - but it's on the user to verify whether the edits were completely localized. In my experience NB mostly tends to make relatively surgical edits but if you're not careful it'll introduce other minute changes.
And that point you can either start over or just feather/mask with the original in any Photoshop type application.
I would assume it depends on how it generates the images.
I've used Claude to generate fairly simple icons and launch images for an iOS game and I make sure to have it start with SVG files since those can be defined as code first. This way it's easier to iterate on specific elements of the image (certain shapes need to be moved to a different position, color needs to be changed, text needs an update, etc.).
Claude does image generation in surprising ways - we did a small evaluation [1] of different frontier models for image generation and understanding, and Claude is by far the most surprising in results.
I’ve been really excited for you infographic generation. Previous models from Google and openAI had very low detail/resolution for these things.
I’ve found in general that the first generation may not be accurate but a few rolls of the dice and you should have enough to pick a style and format that works, which you can iterate on.
Something I find weird about AI image generation models is that even though they no longer produce weird "artifacts" that give away that the fact that it was AI generated, you can still recognize that it's AI due to stylistic choices.
Not all examples they gave were like this. The example they gave of the word "Typography" would have fooled me as human-made. The infographics stood out though. I would have immediately noticed that the String of Turtles infographic was AI generated because of the stylistic choices. Same for the guide on how to make chai. I would be "suspicious" of the example they gave of the weather forecast but wouldn't immediately flag at as AI generated.
Similar note, earlier I was able to tell if something was AI generated right off the bat by noticing that it had a "Deviant Art" quality to it. My immediate guess is that certain sources of training data are over-represented.
We are just very sharp when it comes to seeing small differences in images.
I'm reminded of when the air force decided to create a pilot seat that worked for everyone. They took the average body dimensions of all their recruits and designed a seat to fit the average. It turned out, the seat fit none of their recruits. [1]
I think AI image generation is a lot like this. When you train on all images, you get to this weird sort of average space. AI images look like that, and we recognize it immediately. You can prompt or fine tune image models to get away from this, though -- the features are there it's a matter of getting them out. Lots of people trying stuff like this: https://www.reddit.com/r/StableDiffusion/comments/1euqwhr/re..., the results are nearly impossible to distinguish from real images.
What determines which “average” AI models latch onto? At a pixel level, the average of every image is a grayish rectangle; that's obviously not what we mean and AI does not produce that. At a slightly higher level, the average of every image is the average of every subject every photographed or drawn (human, tree, house, plate of food, ...) in concept space; but AI still doesn't generate a human with branches or a house with spaghetti on it. At a still higher level there are things we recognize as sensible scenes, e.g., barista pouring a cup of coffee, anime scene of a guy fighting a robot, watercolor of a boat on a lake, which AI still does not (by default) average into, say, an equal parts watercolor/anime/photorealistic image of a barista fighting a robot on a boat while pouring a cup of coffee.
But it is undeniable that AI images do have an “average” feel to them. What causes this? What is the space over which AI is taking an average to produce its output? One possible answer is that a finite model size means that the model can only explore image space with a limited resolution, and as models get bigger/better they can average over a smaller and smaller portion of this space, but it is always limited.
But that raises the question of why models don't just naturally land on a point in image space. Is this just a limitation of training, which punishes big failures more strongly than it rewards perfection? Or is there something else at play here that's preventing models from landing directly on a “real” image?
> At a pixel level, the average of every image is a grayish rectangle; that's obviously not what we mean and AI does not produce that.
That isn't correct since images in the real world aren't uniformly distributed color-wise. Take, for example, the famous ImageNet normalization magic numbers:
If it were actually uniformly distributed, the mean for each channel would be 0.5 and the standard deviation would be 0.289. Also due to z-normalization, the "image" most image models see is not how humans typically see images.
I think it's because they're all trained on the same data (everything they could possibly scrape from the open web). The models tend to learn some kind of distribution of what is most likely for a given prompt. It tends to produce things that are very average looking, very "likely", but as a result also predictable and unoriginal.
If you want something that looks original, you have to come up with a more original prompt. Or we have to find a way to train these models to sample things that are less likely from their distribution? Find a way to mathematically describe what it means to be original.
It still has some artifacts more often than not, they are a lot subtler in nature but they still come out, whether it's texture, proportion, lighting, or perspective. Now some things are easier to fix on second pass edits, some are not. I guess it's why they consider image editing to be the next challenge.
It's a bit odd to say, but another big clue identifying something as AI-generated is that it simply looks "too good" for what it is being used for. If I see a little info graphic demonstrating something relatively mundane, and it has nice 3D rendered characters or graphical elements, at this point it's basically guaranteed to be AI, because you just sort of intuitively know when something would've justified the human labor necessary to produce that.
Funny enough that had crossed my mind with the woodchuck example, because at a glance I can't see any weird artifacts, but I felt confident I could tell it was AI generated immediately if I saw it in the wild, and I couldn't really explain why. My immediate guess was "well, who the hell would actually bother to make something like this?"
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.
Some days it feels like I'm the only hacker left who doesn't want government mandated watermarking in creative tools. Were politicians 20 years ago as overreative they'd have demanded Photoshop leave a trace on anything it edited. The amount of moral panic is off the charts. It's still a computer, and we still shouldn't trust everything we see. The fundamentals haven't changed.
> It's still a computer, and we still shouldn't trust everything we see. The fundamentals haven't changed.
I think that by now it should be crystal clear to everyone that it matters a lot the sheer scale a new technology permits for $nefarious_intent.
Knives (under a certain size) are not regulated. Guns are regulated in most countries. Atomic bombs are definitely regulated. They can all kill people if used badly, though.
When a photo was faked/composed with old tech, it was relatively easy to spot. With photoshop, it became more complicated to spot it but at the same time it wasn't easy to mass-produce altered images. Large models are changing the rules here as well.
I think we're overreacting. Digital fakes will proliferate, and we'll freak out bc it's new. But after a certain amount of time, we'll just get used to it and realize that the world goes on, and whatever major adverse effects actually aren't that difficult to deal with. Which is not the case with nuclear proliferation or things like that.
The story of human history is newer generations freaking about progress and novel changes that have never been seen before. And later generations being perfectly okay with it and adapting to a new style of life.
In general I concur but the adaptation doesn't come out of the blue or just only because people get used to it but also because countermeasures are taken, regulations are written and adjustments are made to reduce the negative impact. Also the hyperconnected society is still relatively new and I'm not sure we have adapted for it yet.
I think the long term effect will be that photos and videos no longer have any evidentiary value legally or socially, absent a trusted chain of custody.
It shouldn’t be that we panic about it and regulate the hell out.
We could use the opportunity to deploy robust systems of verification and validation to all digital works. One that allows for proving authenticity while respecting privacy if desired. For example… it’s insane in the US we revolve around a paper social security number that we know damn well isn’t unique. Or that it’s a massive pain in the ass for most people to even check the hash of a download.
But people with actual nefarious intent will easily be able to remove these watermarks, however they're implemented. This is copy protection and key escrow all over again - it hurts honest people and doesn't even slow down bad people.
> Knives (under a certain size) are not regulated. Guns are regulated in most countries. Atomic bombs are definitely regulated
I don’t think this is a good comparison: knives are easy to produce, guns a bit harder, atomic bombs definitely harder. You should find something that is as easy to produce as a knife, but regulated.
Politicians absolutely were doing this 20-30 years ago. Plenty of folks here are old enough to remember debates on Slashdot around the Communications Decency Act, Child Online Protection Act, Children's Online Privacy Protection Act, Children's Internet Protection Act, et al.
> “My wife and I have been together for over 30 years, and she has my voice everywhere,” Schlegel said. “She could easily clone my voice on free or inexpensive software to create a threatening message that sounds like it’s from me and walk into any courthouse around the country with that recording.”
> “The judge will sign that restraining order. They will sign every single time,” said Schlegel, referring to the hypothetical recording. “So you lose your cat, dog, guns, house, you lose everything.”
At the moment, the only alternative is courts simply never accept photo/video/audio as evidence. I know if I were a juror I wouldn't.
At the same time, yeah, watermarks won't work. Sure, Google can add a watermark/fingerprint that is impossible to remove, but there will be tools that won't put such watermarks/fingerprints.
I suspect watermarking ends up being a net negative, as people learn to trust that lack of a watermark indicates authenticity. Propaganda won’t have the watermark.
You do know that every color copier comes with the ability to identify US currency and would refuse to copy it? And that every color printer leaves a pattern of faint yellow dots on every printout that uniquely identifies the printer?
I generally don't think that's it's good or just for a government to collude with manufacturers to track/trace it's citizens without consent or notice. And even if notice was given, I'd still be against it
The arguments put forward by people generally I don't find compelling -- for example, in this thread around protecting against counterfeit.
The "force" applied to address these concerns is totally out of proportion. Whenever these discussions happen, I feel like they descend into a general viewpoint, "if we could technically solve any possible crime, we should do everything in our power to solve it."
I'm against this viewpoint, and acknowledge that that means _some crime_ occurs. That's acceptable to me. I don't feel that society is correctly structured to "treat" crime appropriately, and technology has outpaced our ability to holistically address it.
Generally, I don't see (speaking for the US) the highest incarceration rate in the world to be a good thing, or being generally effective, and I don't believe that increasing that number will change outcomes.
I'm sure Apple will roll something out in the coming years. Now that just anyone can easily AI themselves into a picture in front of the Eiffel tower, they'll want a feature that will let their users prove that they _really_ took that photo in front of the Eiffel tower (since to a lot of people sharing that you're on a Paris vacation is the point, more than the particular photo).
I bet it will be called "Real Photos" or something like that, and the pictures will be signed by the camera hardware. Then iMessage will put a special border around it or something, so that when people share the photos with other Apple users they can prove that it was a real photo taken with their phone's camera.
How "real" are iPhone photos? They're also computationally generated, not just the light that came through the lens.
Even without any other post-processing, iPhones generate gibberish text when attempting to sharpen blurry images, they delete actual textures and replace them with smooth, smeared surfaces that look like a watercolor or oil paintings, and combine data from multiple frames to give dogs five legs.
For example, it's trivial to post an advertisement without disclosure. Yet it's illegal, so large players mostly comply and harm is less likely on the whole.
> I don't think it will be easy to just remove it.
No, but model training technology is out in the open, so it will continue to be possible to train models and build model toolchains that just don't incorporate watermarking at all, which is what any motivated actor seeking to mislead will do; the only thing watermarking will do is train people to accept its absence as a sign of reliability, increasing the effectiveness of fakes by motivated bad actors.
It's an image. There's simply no way to add a watermark to an image that's both imperceptible to the user and non-trivial to remove. You'd have to pick one of those options.
I'm not sure that's correct. I'm not an expert, but there's a lot of literature on digital watermarks that are robust to manipulation.
It may be easier if you have an oracle on your end to say "yes, this image has/does not have the watermark," which could be the case for some proposed implementations of an AI watermark. (Often the use-case for digital watermarks assumes that the watermarker keeps the evaluation tool secret - this lets them find, e.g, people who leak early screenings of movies.)
Exactly, a diffusion model can denoise the watermark out of the image. If you wanted to be doubly sure you could add noise first and then denoise which should completely overwrite any encoded data. Those are trivial operations so it would be easy to create a tool or service explicitly for that purpose.
The incentive for commercial providers to apply watermarks is so that they can safely route and classify generated content when it gets piped back in as training or reference data from the wild. That it's something that some users want is mostly secondary, although it is something they can earn some social credit for by advertising.
You're right that there will existed generated content without these watermarks, but you can bet that all the commercial providers burning $$$$ on state of the art models will gradually coalesce around some means of widespread by-default/non-optional watermarking for content they let the public generate so that they can all avoid drowning in their own filth.
I don't understand why there isn't an obvious, visible watermark at all. Yes, one could remove it but let's assume 95% of people don't bother removing the visible watermark. It would really help with seeing instantly when an image was AI generated.
Regardless of how you feel about this kind of steganography, it seems clear that outside of a courtroom, deepfakes still have the potential to do massive damage.
Unless the watermark randomly replaces objects in the scene with bananas, these images/videos will still spread like wildfire on platforms like TikTok, where the average netizen's idea of due diligence is checking for a six‑fingered hand... at best.
It solves some problems! For example, if you want to run a camgirl website based on AI models and want to also prove that you're not exploiting real people
> It solves some problems! For example, if you want to run a camgirl website based on AI models and want to also prove that you're not exploiting real people
So, you exploit real people, but run your images through a realtime AI video transformation model doing either a close-to-noop transformation or something like changing the background so that it can't be used to identify the actual location if people do figure out you are exploiting real people, and then you have your real exploitation watermarked as AI fakery.
I don't think this is solving a problem, unless you mean a problem for the would-be exploiter.
Your use case doesn't even make sense. What customers are clamoring for that feature? I doubt any paying customer in the market for (that product) cares. If the law cares, the law has tools to inquire.
All of this is trivially easy to circumvent ceremony.
Google is doing this to deflect litigation and to preserve their brand in the face of negative press.
They'll do this (1) as long as they're the market leader, (2) as long as there aren't dozens of other similar products - especially ones available as open source, (3) as long as the public is still freaked out / new to the idea anyone can make images and video of whatever, and (4) as long as the signing compute doesn't eat into the bottom line once everyone in the world has uniform access to the tech.
The idea here is that {law enforcement, lawyers, journalists} find a deep fake {illegal, porn, libelous, controversial} image and goes to Google to ask who made it. That only works for so long, if at all. Once everyone can do this and the lookup hit rates (or even inquiries) are < 0.01%, it'll go away.
It's really so you can tell journalists "we did our very best" so that they shut up and stop writing bad articles about "Google causing harm" and "Google enabling the bad guys".
We're just in the awkward phase where everyone is freaking out that you can make images of Trump wearing a bikini, Tim Cook saying he hates Apple and loves Samsung, or the South Park kids deep faking each other into silly circumstances. In ten years, this will be normal for everyone.
Writing the sentence "Dr. Phil eats a bagel" is no different than writing the prompt "Dr. Phil eats a bagel". The former has been easy to do for centuries and required the brain to do some work to visualize. Now we have tools that previsualize and get those ideas as pixels into the brain a little faster than ASCII/UTF-8 graphemes. At the end of the day, it's the same thing.
And you'll recall that various forms of written text - and indeed, speech itself - have been illegal in various times, places, and jurisdictions throughout history. You didn't insult Caesar, you didn't blaspheme the medieval church, and you don't libel in America today.
> How can they distinguish from real people exploited to AI models autogenerating everything?
Watermarking by compliant models doesn't help this much because (1) models without watermarking exist and can continue to be developed (especially if absence of a watermark is treated as a sign of authenticity), so you cannot rely on AI fakery being watermarked, and (2) AI models can be used for video-to-video generation without changing much of the source, so you can't rely on something accurately watermarked as "AI-generated" not being based in actual exploitation.
Now, if the watermarking includes provenance information, and you require certain types of content to be watermarked not just as AI using a known watermarking system, but by a registered AI provider with regulated input data safety guardrails and/or retention requirements, and be traceable to a registered user, and...
Well, then it does something when it is present, largely by creating a new content gatekeepiing cartel.
> How can they distinguish from real people exploited to AI models autogenerating everything?
The people who care don't consume content which even just plausibly looks like real people exploited. They wouldn't consume the content even if you pinky promised that the exploited looking people are not real people. Even if you digitally signed that promise.
It would be more productive for camera manufacturers to embed a per-device digital signature. Those care to prove their image is genuine could publish both pre and post processed images for transparency.
If watermarking becomes a legal mandate, it will inevitably include a prohibition on distributing (and using and maybe even possessing, but the distribution ban is the thing that will have the most impact, since it is the part that is most policable, and most people aren't going to be training their own models, except, of course, the most motivated bad actors) open models that do not include watermarking as a baked-in model feature. So, for most users, it'll be much less accessible (and, at the same time, it won't solve the problem.)
It is terrifying, but inevitable. Perhaps AI companies flooding the commons with excrement wasn't the best idea, now we all have to suffer the consequences.
Reminder that even in the hypothetical world where every AI image is digitally watermarked, and all cameras have a TPM that writes a hash of every photo to the blockchain, there’s nothing to stop you from pointing that perfectly-verified camera at a screen showing your perfectly-watermarked AI image and taking a picture.
Image verification has never been easy. People have been airbrushed out of and pasted into photos for over a century; AI just makes it easier and more accessible. Expecting a “click to verify” workflow is unreasonable as it has ever been; only media literacy and a bit of legwork can accomplish this task.
Competent digital watermarks usually survive the 'analog hole'. Screen-cam resistant watermarks have been in use since at least 2020, and if memory serves, back to 2010 when I first starting reading about them, but I don't recall what it was called back then.
I just tried asking Gemini about a photo I took of my screen showing an image I edited with Nano Banana Pro... and it said "All or part of the content was generated with Google AI. SynthID detected in less than 25% of the image".
We need to be super careful with how legislation around this is passed and implemented. As it currently stands, I can totally see this as a backdoor to surveillance and government overreach.
If social media platforms are required by law to categorize content as AI generated, this means they need to check with the public "AI generation" providers. And since there is no agreed upon (public) standard for imperceptible watermarks hashing that means the content (image, video, audio) in its entirety needs to be uploaded to the various providers to check if it's AI generated.
Yes, it sounds crazy, but that's the plan; imagine every image you post on Facebook/X/Reddit/Whatsapp/whatever gets uploaded to Google / Microsoft / OpenAI / UnnamedGovernmentEntity / etc. to "check if it's AI". That's what the current law in Korea and the upcoming laws in California and EU (for August 2026) require :(
I don't believe that you can do this for photography. For AI-images, if the embedded data has enough information (model identification and random seed), one can prove that it was AI by recreating it on the fly and comparing. How do you prove that a photographic image was created by a CCD? If your AI-generated image were good enough to pass, then hacking hardware (or stealing some crypto key to sign it) would "prove" that it was a real photograph.
Hell, it might even be possible for some arbitrary photographs to come up with an AI prompt that produces them or something similar enough to be indistinguishable to the human eye, opening up the possibility of "proving" something is fake even when it was actually real.
What you want just can't work, not even from a theoretical or practical standpoint, let alone the other concerns mentioned in this thread.
It solves a real problem - if you have something sketchy, the big players can repudiate it, the authorities can more formally define the black market, and we can have a ‘war on deepfakes’ to further enable the authorities in their attempts to control the narratives.
Every model is "grey market". They're all trained on data without complying with any licensing terms that may exist, be they proprietary or copyleft. Every major AI model is an instance of IP theft.
I feel like I am going crazy or missed something simple but when I use the Gemini app and I ask it to edit a photo that I upload, 2.5 flash works really well but 2.5 pro or 3.0 pro do a very poor job. I uploaded an image of me and asked it to make me bald and flash did a great job of just changing me in the photo but 3.0 pro took me out of the photo completely and just created a headshot of a bald man that only sort of resembled me. Am I missing something or does paying for the pro version not give you anything over the 2.5 flash model?
This is the first image model I’ve used that passed my piano test. It actually generated an image of a keyboard with the proper pattern of black keys repeated per octave – every other model I’ve tried this with since the first Dall-E has struggled to render more than a single octave, usually clumping groups of two black keys or grouping them four at a time. Very impressive grasp of recursive patterns.
If you ask it for anything outside of the standard 88 key set it falls short. For instance
"Generate a piano, but have the left most key start at middle C, and the notes continue in the standard order up (D, E, F, G, ...) to the right most key"
The above prompt will be wrong, seemingly every time. The model has no understanding of the keys or where they belong, and it is not able to intuit creating something within the actual confines of how piano notes are patterned.
"Generate a piano but color every other D key red"
This also wrong, every time, with seemingly random keys being colored.
I would imagine that a keyboard is difficult to render (to some extent) but I also don't think its particularly interesting since it is a fully standardized object with millions of pictures from all angles in existence to learn from right?
You can try it out for free on LMArena [0]: New Chat -> Battle dropdown -> Direct Chat -> Click on Generate Image in the chat box -> Click dropdown from hunyuan-image-3.0 -> gemini-3-pro-image-preview (nano-banana-pro).
I've only managed to get a few prompts to go through, if it takes longer than 30 seconds it seems to just time out. Image quality seems to vary wildly; the first image I tried looked really good but then I tried to refresh a few times and it kept getting worse.
Thanks - this worked for me (some errors, some success).
Last week I was making a birthday card for my son with the old model. The new model is dramatically better - I'm asking for an image in comic book style, prompted with some images of him.
With the previous model, the boy was descriptively similar (e.g. hair colour and style) but looked nothing like him. With this model it's recognisably him.
When I do that, I get two (very similar but not identical) responses side-by-side in one image (I guess as if the model is battling itself?). Is that normal for lmarena?
I'd be curious about how well the inline verification works - an easy example is to have it generate a 9-pointed star, a classic example that many SOTA models have difficulties with.
In the past, I've deliberately stuck a Vision-language model in a REPL with a loop running against generative models to try to have it verify/try again because of this exact issue.
EDIT: Just tested it in Gemini - it either didn't use a VLM to actually look at the finished image or the VLM itself failed.
Output:
I have finished cross-referencing the image against the user's specific requests. The primary focus was on confirming that the number of points on the star precisely matched the requested nine. I observed a clear visual representation of a gold-colored star with the exact point count that the user specified, confirming a complete and precise match.
"Inline verification of images following the prompt is awesome, and you can do some _amazing_ stuff with it." - could you elaborate on this? sounds fascinating but I couldn't grok it via the blog post (like, it this synthid?)
LLMs might be a dead end, but we're going to have amazing images, video, and 3D.
To me the AI revolution is making visual media (and music) catch up with the text-based revolution we've had since the dawn of computing.
Computers accelerated typing and text almost immediately, but we've had really crude tools for images, video, and 3D despite graphics and image processing algorithms.
AI really pushes the envelope here.
I think images/media alone could save AI from "the bubble" as these tools enable everyone to make incredible content if you put the work into it.
Everyone now has the ingredients of Pixar and a music production studio in their hands. You just need to learn the tools and put the hours in and you can make chart-topping songs and Hollywood grade VFX. The models won't get you there by themselves, but using them in conjunction with other tools and understanding as to what makes good art - that can and will do it.
Screw ChatGPT, Claude, Gemini, and the rest. This is the exciting part of AI.
LLMs are useful, but they've hit a wall on the path to automating our jobs. Benchmark scores are just getting better at test taking. I don't see them replacing software engineers without overcoming obstacles.
AI for images, video, music - these tools can already make movies, games, and music today with just a little bit of effort by domain experts. They're 10,000x time and cost savers. The models and tools are continuing to get better on an obvious trend line.
I'm literally a software engineer, and a business owner. I don't think about this in binary terms (replacement or not), but just like CMS's replaced the jobs of people that write HTML by hand to build websites, I think whole classes of software development will get democratized.
For example, I'm currently vibe coding an app that will be specific to our company, that helps me run all the aspects of our business and integrates with our systems (so it'll integrate with quickbooks for invoicing, etc), and help us track whether we have the right insurance across multiple contracts, will remind me about contract deadlines coming up, etc.
It's going to combine the information that's currently in about 10 different slightly out of sync spreadsheets, about 2 dozen google docs/drive files, and multiple external systems (Gusto, Quickbooks, email, etc).
Even though I could build all this manually (as a software developer), I'd never take the time to do it, because it takes away from client work. But now I can actually do it because the pace is 100x faster, and in the background while I'm doing client work.
Doesn’t seem like a dead end at all. Once we can apply LLMs to the physical world and its outputs control robot movements it’s essentially game over for 90% of the things humans do, AGI or not.
It's crazy how good these models are at text now. Remember when text was literally impossible? Now the models can diagetically render any text. It's so good now that it seems like a weird blip that it _wasn't_ possible before.
I agree, it's improving by leaps. I'm still patiently awaiting for my niche use of creating new icons though, one that can match the existing curvature, weight, spacing, and balance. It seems AI is struggling in the overlap of visuals <-> code, or perhaps there's less business incentive to train on that front. I know the pelican on bicycle svg is getting better, but still really rough looking and hard to modify with prompt versus just spending some time upfront to do it yourself in an editor.
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
> 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
No, its not the beginning, multiple different watermarking standards, watermark checking systems, and, of course, published countermeasures of various effectiveness for most of them, have been around for a while.
I don't understand the excitement around generating and/or watching AI-produced videos. To me it's probably the single most uninteresting and boring thing related to AI that I can think of. What is the appeal?
Nonetheless, ask it to “create an infographic on how Google works”. Do you not see any excitement in the result? I think it’s pretty impressive and has a lot of utility.
As a general content I agree it's a bit off putting, but I find it a lot of fun when generating content among friends like internal jokes and educational content. I got my kid to drink some meds by generating an image of a hero telling him it's important to take.
Does anyone know if this is predicting the entire image at once, or if it's breaking it into constituent steps i.e. "draw text in this font at this location" and then composing it from those "tools"? It would be really interesting if they've solved the garbled text problem within the constraint of predicting the entire image at once.
The previous nano banana was using composing tools. It was really obvious by some of the janky outputs it made. Not sure about this one, but presumably they built off it.
There still is some garbled text sometimes so it can't be the latter (try to get it to generate a map of 48 us states labeled - the ones that are too small to write on and need arrows were garbled (1 attempt))
I’m pretty sure, but no expert on the matter, that correct text rendering was solved by feeding in bitmaps of rasterized fonts as supplemental context to the image generation models.
Google needs to pace themselves. AI studio, Antigravity, Banana, Banana Pro, Grape Ultra, Gemini 3, etc. This information overload don't do them any good whatsoever.
Why? They're mostly different markets. Most people using Nano Banana Pro aren't using Antigravity.
A cluster of launches reinforces the idea that Google is growing and leading in a bunch of areas.
In other words, if it's having so many successes it feels like overload, that's an excellent narrative. It's not like it's going to prevent people from using the tools.
Powell Doctrine, but for AI. No one should dispute that Google is the leader in every(?) category of AI: LLM, image gen, video editing, world models, etc.
This cluster of launches might not be intentional. It could just be a bunch of independent teams all trying to get their launches out before the EOY deadline.
I was at a tech conference yesterday, and I asked someone if they had tried nano banana. They looked at me like I was crazy. These names aren't helping! (But honestly I love it, easier to remember than Gemini-2.whatever.
Honestly I give Google credit for realizing that they had something that people were talking about and running with it instead of just calling it gemini-image-large-with-text-pro
They tried calling it gemini-2.5-whatever, but social media obsessed over the name "Nano Banana", which was just its codename that got teased on Twitter for a few weeks prior to launch.
After launch, Google's public branding for the product was "Gemini" until Google just decided to lean in and fully adopt the vastly more popular "Nano Banana" label.
The public named this product, not Google. Google's internal codename went virally popular and outstaged the official name.
Branding matters for distribution. When you install yourself into the public consciousness with a name, you'd better use the name. It's free distribution. You own human wetware market share for free. You're alive in the minds of the public.
Renaming things every human has brand recognition of, eg. HBO -> Max, is stupid. It doesn't matter if the name sucks. ChatGPT as a name sucks. But everyone in the world knows it.
This will forever be Nano Banana unless they deprecate the product.
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.
Here, it mostly poisons your test, because that exact photo probably exists in the underlying training data and the trained network will be more or less optimized on working with it. It's really the same consideration you'd want to make when testing classifiers or other ML techs 10 years ago.
Most people taking to a task like this will be using an original photo -- missing entirely from any training date, poorly framed, unevenly lit, etc -- and you need to be careful to capture as much of that as possible when trying to evaluate how a model will work in that kind of use case.
The failure and stress points for AI tools are generally kind of alien and unfamiliar because the way they operate is totally different than the way a human operates, and if you're not especially attentive to their weird failure shapes and biases when you want to test them, or you'll easily get false positives (and false negatives) that lead you to misleading conclusions.
I also tried that in the past with poor results. I just tried it this morning with nano banana pro and it nailed it with a very short prompt: "Repaint the house white with black trim. Do not paint over brick."
I don't know what it is with Gemini (and even other models) but I swear they must be doing some kind of active load-dependant quanitization or a/b/c/d testing behind the scenes, because sometimes the model is stellar and hitting everything, and other times it's tripping all over itself.
The most effective fix I have found is that when the model is acting dumb, just turn it off and come back in the few hours to a new chat and try again.
Maybe somewhere in the original comment it would have been fair to mention you can barely see the house in the original photo. This is actually a hilarious complaint
That cannot be a valid excuse. Other than adding extra windows to the clearly visible wall, it's obvious that model perfectly capable to "see" the house. It just cannot "believe" that there can be a big empty wall on a garden house.
Everyone who worked on this is a traitor to the human race. Why do we need to make it impossible to make a living as an artist? Who thinks an endless tsunami of garbage “content” churned out by machines dropping the bottom out of all artistic disciplines is a good idea?
Capitalism, at work. Wherever there is a cost, there will be attempts made at cost efficiency. Google understands that hiring designers or artists is expensive, and they want to offer a cheaper, more effective alternative so that they can capture the market.
In a coffee shop this morning I saw a lady drawing tulips with a paper and pencil. It was beautiful, and I let her know... But as I walked away I felt sad that I don't feel that when browsing online anymore- because I remember how impressive it used to feel to see an epic render, or an oil painting, etc... I've been turned cynical.
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.
Wow, that top image is actually quite good! Interestingly, I just got into Pro and got a worse result than yours. https://imgur.com/a/ENNk68B ... and it really seems to just vary by attempt even with the exact same prompt.
Much better than previous attempts. Still has an extra lane with the cars on the right cutting off the cars in the middle. Still has the numbers in the wrong order.
I'd try a some more if I were you. I saw an example of generated infographic that was greatly improved over anything I've seen an image generator do before. What you desire seems in the realm of possibility.
Imagen4 did no better. edit: example https://imgur.com/Dl8PWgm with a so-so result: four lanes, cars at least facing the same way, lane block looks good, weird extra division in the center, some numbers repeated, one arrow going straight into construction, one arrow going backwards
edit: or Imagen4 Ultra. https://imgur.com/a/xr2ElXj cars facing opposite directions within a lane, 2-way (4 lanes total), double-ended arrows, confused disaster. pretty though.
One of the things I've always been curious about is how effective diffusion models can be for web and app design. They're generally trained on more organic photos, but post-training on SDXL and Flux have given me good results here in the past (with the exception of text).
It's been interesting seeing the results of Nano Banana Pro in this domain. Here are a few examples:
Prompt: "A travel planner for an elegant Swiss website for luxury hiking tours. An interactive map with trail difficulty and booking management. Should have a theme that is alpine green, granite grey, glacier white"
Prompt: "a landing page for a saas crypto website, purple gradient dark theme. Include multiple sections, including one for coin prices, and some graphs of value over time for coins, plus a footer"
Note that this is with a lora I built for flux specifically for website generation. Overall, nbp seems to have less creative / inspired outputs, but the text is FAR better than the fever dream Flux is producing. I'm really excited to see how this changes design. At the very least it proved it can get close to a production quality for output, now it's just about tuning it.
In my limited testing, at least in terms of maintaining consistency between input and output for Asian faces, it has even regressed.
Actually, Gemini 3 is about the same, and doesn't feel as good as Claude 4.5. I have a feeling it's been fine-tuned for a cool front-end marketing effect.
Furthermore, I really don't understand why AI Studio, now requiring me to use its own API for payment, still adds a watermark.
There's some really impressive things about this (the speed, the lack of typical AI image gen artifacts) but it also seems less creative than other models I've tried?
"mountain dew themed pokemon" is the first search prompt I always try with new image models and Nano Banna Pro just gave me a green pikachu.
Other models do a much better job of creating something new.
IMHO I'd rather them focus on strong literal prompt adherence so that more detailed prompts produce more accurate results.
That way you can stick your choice of any number of LLM preprocessors in front of a generic prompt like "mountain dew themed pokemon" and push the responsibility of creating a more detailed prompt upstream.
Just last night I was using Gemini "Fast" to test its output for a unique image we would have used in some consumer research if there had been a good stock image back in the day. I have been testing this prompt since the early days of AI images. The improvement in quality has been pretty remarkable for the same prompt. Composition across this time has been consistent. What I initially thought was "good enough" now is... fantastic. Just so many little details got more life-like w/ each new generation. Funnily enough, our images must be 3:2 aspect ratio. I kept asking GFast to change its square Fast output to 3:2. It kept saying it would, but each image was square or nearly square. GFast in the end was very apologetic, and said it would alert about this issue. Today I read that GPro does aspect ratios. Tried the same prompt again burning up some "Thinking" credits, and got another fantastically life-like image in 3:2. We have a new project coming up. We have relied entirely on stock or in some cases custom shot images to date. Now, apart from the time needed to get the prompts right whilst meeting with the client, I cannot see how stock or custom images can compete. I mean the GPro images -- again which is very specific to an unusual prompt -- is just "Wow". Want to emphasize again -- we are looking for specific details that many would not. So the thoughts above are specific to this. Still, while many faults can be found with AI, Nano Banana is certainly proven itself to me.
edit: I was thinking about this, and am not sure I even saw Pro3 as my image option last night. Today it was clearly there.
It’s interesting, I’m trying to use it to create a themed collage by providing a few images and it does that wonderfully, but in the process it is also hallucinating the images I use so I end up with weird distorted faces. Other tools can do this without issue, but something about faces in images this model just has to modify them every time. Ask it to remove background objects and the faces get distorted as well.
Using it for non-people involved images and it’s pretty good although I haven’t done much and it isn’t doing anything 2.5-flash wasn’t already doing in the same amount of requests.
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.
Could be they are specifically training against it. There was some controversy about "studio ghibli style". Similarly how in the early days of Stable Diffusion "Greg Rutkowski style" was a very popular prompt to get a specific look. These days modern Stable Diffusion based models like SD 3 or FLUX mostly removed references to specific artists from their datasets.
Slightly off topic, but how are people creating long videos like 30 second videos that I often see on Instagram? It I try to use Veo to make split videos, it simply cannot maintain the style or weird quirks get into the subsequent videos. Is there anything else that's the best video generation model currently other than Veo?
I really hope Google reads these HN posts. They've had some big "product" wins but the pricing, packaging, and user system is a severe blocker to growth. If developers can't or won't figure it out -- how the heck are consumers?
And both their consumer apps are slow. You can replicate this yourself. Go to AI Studio, paste in 80K tokens of text, then type something on your keyboard, and see what happens. The Gemini web app is even worse somehow. A horrifically slow and buggy app. Not new problems either, barely any improvement on this over more than 1 year.
First model I've seen that was consistently compositional, easily handling requests like
“Generate an image of an african elephant painted in the New England flag, doing a backflip in front of the russian federal assembly.”
OpenAI made the biggest step change towards compositionality in image generation when they started directly generating image tokens for decoders from foundation llms, and it worked very well (openais images were better in this regard than nano banana 1, but struggled with some OOD images like elephants doing backflips), but banana 2 nails this stuff in a way I haven't seen anywhere else
if video follows the same trends as images in terms of prompt adherence, that will be very valuable... and interesting
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
To expand, it comes from the stealth name it was given on LMArena I believe. The model made news while still in "stealth mode" and so Google capitalised on the PR they'd already built around that and just launched it officially with the same name.
This is really impressive. As a former designer, I'm equally excited that people will be able to generate images like this with a prompt, and sad that there will be much less incentive for people to explore design / "photoshopping" as a craft or a career.
At the end of the day, a tool is a tool, and the computer had the same effect on the creative industry when people started using them in place of illustrating by hand, typesetting by hand, etc. I don't want my personal bias to get in the way too much, but every nail that AI hammers into the creative industry's coffin is hard to witness.
I feel you. Infact, IMO, SWE1 level coding industry seems to be a couple years lagging on this aspect.
The trouble is that learning fundamentals now is a large trough to go past, just the way grade 3-10 children learn their math fundamentals despite there being calculators. It's no longer "easy mode" in creative careers.
If Nano-Banana-pro with Veo 3.1 existed during my PhD, I would’ve finished a 6-year dissertation in a single year — it’s generating ideas today that used to take me 18 months just to convince people were possible.
I tried the same prompt as one of the examples (https://i.imgur.com/iQTPJzz.png), in the two ways they say you can run it, via Google Gemini and Google AI Studio (I suppose they're different somehow?). The prompt was "Create an infographic that shows hot to make elaichi chai" and Google Gemini created a infographic (https://i.imgur.com/aXlRzTR.png), but it was all different from what the example showed. Google AI Studio instead created a interactive website, again with different directions: https://i.imgur.com/OjBKTkJ.png
There is not a single mention about accuracy, risks or anything else in the blogpost, just how awesome the thing is. It's clearly not meant to be reliable just yet, but not making this clear up front. Isn't this almost intentionally misleading people, something that should be illegal?
Whoever said there was a universal recipe for Elaichi Chai? It makes sense that there would be different recipes. If you are more stringent with the prompt and give it the proper context of what you want the steps to be, you'll arrive at that consistency.
I was just playing with the non-pro version of this and it seems to add both a Gemini and Disney watermark. Presumably this was because I referenced beauty and the beast.
Anyone know if this is an hallucination or if they have some kind of deal with content owners to add branding?
Neat use-case, though the sword literally telescopically inverts itself at the beginning of the scene like a light saber where you would have expected it to be drawn from its scabbard.
I'd be interested to see how Wan 2.2 First/Last frame handles those images though...
That is an interesting error actually. It happened because both orientations of the sword are visually plausible, but not abrupt transitions from one to the other; there needs to be physical continuity.
Here is a reproduction of the Matrix bullet time shot with and without pose guidance to illustrate the problem: https://youtu.be/iq5JaG53dho?t=1125
yeah sadly veo 3.1 has not caught up to the image generation capabilities. May be we need to work on how to make video generation more physically consistent. but the image generation results from banana pro are great.
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.
1. Trigger Circle to Search with long holding the home button/bar
2. Select the image
3. Navigate to About this image on the Google search top bar all the way to the right - check if it says "Made by Google AI" - which means it detected the SynthID watermark.
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).
My experience with Nano Banana is to constantly get consistent image when dealing with muliple objects in a image, I mean creating consistent sequence etc.
We spent a lot of money trying but eventully gave up. If it is easier in Pro, then probably it stands a chance.
What can nano-banana do that chatGPT made images can't? Or is it only better for image editing from what I can gather from these comments so far. I haven't used it so genuinely curious.
I made some direct comparisons my Nano Banana post (https://news.ycombinator.com/item?id=45917875) but Nano Banana can handle photorealistic photos with nuanced prompts much better. And there is no yellow filter.
> 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?
It’s not necessarily harder than other aspects. However:
- It requires an AI that actually understands English, I.e. an LLM. Older, diffusion-only models were naturally terrible at that, because they weren’t trained on it.
- It requires the AI to make no mistakes on image rendering, and that’s a high bar. Mistakes in image generation are so common we have memes about it, and for all that hands generally work fine now, the rest of the picture is full of mistakes you can’t tell are mistakes. Entirely impossible with text.
Nano Banana Pro seems to somewhat reliably produce entire pictures without any mistakes at all.
As a complete layman, it seems obvious that it should be hard? Like, text is a type of graphic that needs to be coherent both in its detail and its large structure, and there’s a very small amount of variation that we don’t immediately notice as strange or flat out incorrect. That’s not true of most types of imagery.
Super important for Google as a search engine so they can filter out and downrank AI generated results. However I expect there are many models out there which don’t do this, that everyone could use instead. So in the end a “feature” like this makes me less likely to use their model because I don’t know how Google will end up treating my blog post if I decide to include an AI generated or AI edited image.
The EU didn't define any specific method of watermarking nor does it need to be tamper resistant. Even if they had specified it though, it's easy to remove watermarks like SynthID.
> Today, we are putting a powerful verification tool directly in consumers’ hands: you can now upload an image into the Gemini app and simply ask if it was generated by Google AI, thanks to SynthID technology. We are starting with images, but will expand to audio and video soon.
Re-rolling a few times got it to mention trying SynthID, but as a false negative, assuming it actually did the check and isn't just bullshitting.
> No Digital Watermark Detected: I was unable to detect any digital watermarks (such as Google's SynthID) that would definitively label it as being generated by a specific AI tool.
This would be a lot simpler if they just exposed the detector directly, but apparently the future is coaxing an LLM into doing a tool call and then second guessing whether it actually ran the tool.
By anybody's AI using SynthID watermarking, not just Google's AI using SynthID watermarking (it looks like partnership is not open to just anyone though, you have to apply).
Nano Banana has been the only model I’ve really loved. As a small businesses who makes products, it’s been a game changer on the marketing side. Now when I’ve got something new I need to advertise in a hurry, I take a crappy pic and fix it in that. Don’t have a perfect model ready yet? That’s ok, I can just alter to look exactly like it will.
What used to cost money and involve wait time is now free and instant.
My most regular use-case is generating silly memes in group chats. If someone posts something meme-worthy or I come up with a creative response, image generation is good for one-off throwaway memes. A recent example was an "official license to opine on sociology", following someone arguing about credentialism.
Recently I also started using image generation models to explore ideas for what changes to make in my paintings. Although generally I don't like the suggestions it makes, sometimes it provides me with creative ideas of techniques that are worth experimenting with.
One way to approach thinking about it is that it's good for exploring permutations in an idea-space.
1) I have a tricep tendon injury and ChatGPT wants me to check my tricep reflex. I have no idea where on the elbow you're supposed to tap to trigger the reflex.
2) I'm measuring my body fat using skin fold calipers. Show me were the measurement sites are.
3) I'm going hiking. Remind me how to identify poison ivy and dangerous snakes.
I think that's a fair assessment. I write a lot of bizarre fiction in my spare time, so Text2Image tools are a fun way to see my visions visualized.
Like this one:
A piano where the keyboard is wrapped in a circular interface surrounding a drummer's stool connected to a motor that spins the seat, with a foot-operated pedal to control rotation speed for endless glissandos.
Nano Banana is more of an image editing model, which probably has more broad use cases for non-generative applications: interior decorating, architecture, picking wardrobes, etc.
Definitely, but don't sleep on its generative capacities either. You can give it a image and instruct it "Use the attached image purely as a stylistic reference" and then proceed to use it as a regular generative model.
In my tests it does outscore Imagen3 and Imagen4 even in the generative capacity, but my benchmark is more focused around prompt adherence. I'd wager that for certain photorealistic tests Imagen4 is probably better.
Yeah... For some reason none of these are use cases in my day to day life. That said, I also don't open Photoshop very often. And maybe that's what this is meant to replace.
Not for everyone everyday, but a good tool to have in the toolbox. I recently was very easily able to mock up what a certain Christmas decoration would look like on the house. By next year, I'm sure that feature will be part of the product page.
I'm creating a team T-shirt from a bunch of kids drawings. The model has synthesize a bunch of disparate drawings into a cohesive concept, incorporate the team's name in the appropriate color and font, and make it simple enough for a T-shirt.
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 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"
I would do the same. But the reason for that is because I’m terrible at drawing and digital art, so I would need some help with the graphics in an infographics anyways. I don’t really need help with writing text or typesetting the text. I feel like if I were better at creating art I would not want AI involved at all.
Honestly I think this is exactly how we're all feeling right now. Racing towards an unknown horizon in a nitrous powered dragster surrounded by fire tornadoes.
But ... it comes from Google. My goal is to eventually degoogle completely. I am not going to add any more dependency - I am way too annoyed at having to use the search engine (getting constantly worse though), google chrome (long story ...) and youtube.
Currently, it’s rolling out in the Gemini app. When you use the “Create image” option, you’ll see a tooltip saying “Generating image with Nano Banana Pro.”
And in AI Studio, you need to connect a paid API key to use it:
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.
On the contrary, it's encouraging to know that maliciously greedy companies like Adobe are getting screwed for being so malicious and greedy :thumbsup:
I had second thoughts about this comment, but if I stopped typing in the middle of it, I would've had to pay a cancellation fee.
Adobe makes money by renting software, not selling it. There are many creatives that would disagree with your ranking of who is more malicious or greedy.
Nano Banana Pro sounds like classic Google branding: quirky name, serious tech underneath. I’m curious whether the “Pro” here is about actual professional‑grade features or just marketing polish. Either way, it’s another reminder that naming can shape expectations as much as specs.
Google has been stomping around like Godzilla this week, and this is the first time I decided to link my card to their AI studio.
I had seen people saying that they gave up and went to another platform because it was "impossible to pay". I thought this was strange, but after trying to get a working API key for the past half hour, I see what they mean.
Everything is set up, I see a message that says "You're using Paid API key [NanoBanano] as part of [NanoBanano]. All requests sent in this session will be charged." Go to prompt, and I get a "permission denied" error.
There is no point in having impressive models if you make it a chore for me to -give you my money-
First off, apologies for the bad first impression, the team is pushing super hard to make sure it is easy to access these models.
- On permission issue, not sure I follow the flow that got you there, pls email me more details if you are able too and happy to debug: Lkilpatrick@google.com
- On overall friction for billing: we are working on a new billing experience built right into AI Studio that will make it super easy to add a CC and go build. This will also come along with things like hard billing caps and such. The expected ETA for global rollout is January!
If it's just the API you're interested in, Fal.ai has put Nano-Banana-Pro up for both generative and editing. A great deal less annoying to sign up for them since they're a pretty generalized provider of lots of AI related models.
https://fal.ai/models/fal-ai/nano-banana-pro
In general a better option, in the early days of AI video I tried to generate a video of a golden retriever using Google's AI Studio. It generated 4 in the highest quality and charged me 36 bucks. Not a crazy amount but definitely an unwelcome suprise.
Fal.ai is pay as you go and has the cost right upfront.
100% agreed. Same reason that I use the OpenRouter API for most LLM usage.
There's the solution right there. Google is still growing its AI "sea legs". They've turned the ship around on a dime and things are still a little janky. Truly a "startup mode" pivot.
While we're on this subject of "Google has been stomping around like Godzilla", this is a nice place to state that I think the tide of AI is turning and the new battle lines are starting to appear. Google looks like it's going to lay waste to OpenAI and Anthropic and claim most of the market for itself. These companies do not have the cash flow and will have to train and build their asses off to keep up with where Google already is.
gpt-image-1 is 1/1000th of Nano Banana Pro and takes 80 seconds to generate outputs.
Two years ago Google looked weak. Now I really want to move a lot of my investments over to Google stock.
How are we feeling about Google putting everyone out of work and owning the future? It's starting to feel that way to me.
(FWIW, I really don't like how much power this one company has and how much of a monopoly it already was and is becoming.)
100% this. I am using the pro/max plans on both claude and openai. Would love to experiment with gemini but paying is next to impossible. Why do i need the risk of a full blown gcp project just to test gemini. No thx.
I had to write a post request to try it when it launched
Oh my, you should have tried to integrate with Google Prism. That was a madness! Nano Banana was just a little tricky to set up in comparison!
Yeah I was confused. I guess I’ll stick with nano plum for now.
You can use it also in Gemini.
It wasn't there when I first went to Gemini after the announcement, but upon revisiting it gave me the prompt to try Nano Banana Pro. It failed at my niche (rare palm trees).
Incredible technology, don't get me wrong, but still shocked at the cumbersome payment interface and annoyed that enabling Drive is the only way to save.
I hate that they kinda try to hide the model version. Like if you click the dropdown in the chat box, you can see that "Thinking" means 3 Pro. When you select the "Create images" tool, it doesn't tell you it's using Nano Banana Pro until it actually starts generating the image.
Tell me the model it's using. It's as if Google is trying to unburden me with the knowledge of what model does what but it's just making things more confusing.
Oh, and setting up AI Studio is a mess. First I have to create a project. Then an API key. Then I have to link the API key to the project. Then I have to link the project to the chat session... Come on, Google.
It's amazing that the "hard problems" are turning out to be "not creating a completely broken user experience".
Is that going to need AGI? Or maybe it will always be out of reach of our silicon overlords and require human input.
Alright results are in! I've re-run all my editing based adherence related prompts through Nano Banana Pro. NB Pro managed to successfully pass SHRDLU, the M&M Van Halen test (as verified independently by Simon), and the Scorpio street test - all of which the original NB failed.
https://genai-showdown.specr.net/image-editingIf you just want to see how NB and NB Pro compare against each other:
https://genai-showdown.specr.net/image-editing?models=nb,nbp
thanks, I love your website. Are you planning to do NB Pro for the text-to-image benchmark too?
Definitely! Even though NB's predominant use case seems to be editing, it's still producing surprisingly decent text-to-image results. Imagen4 currently still comes out ahead in terms of image fidelity, but I think NB Pro will close the gap even further.
I'll try to have the generative comparisons for NB Pro up later this afternoon once I catch my breath.
I...worked on the detailed Nano Banana prompt engineering analysis for months (https://news.ycombinator.com/item?id=45917875)...and...Google just...Google released a new version.
Nano Banana Pro should work with my gemimg package (https://github.com/minimaxir/gemimg) without pushing a new version by passing:
I'll add the new output resolutions and other features ASAP. However, looking at the pricing (https://ai.google.dev/gemini-api/docs/pricing#standard_1), I'm definitely not changing the default model to Pro as $0.13 per 1k/2k output will make it a tougher sell.EDIT: Something interesting in the docs: https://ai.google.dev/gemini-api/docs/image-generation#think...
> The model generates up to two interim images to test composition and logic. The last image within Thinking is also the final rendered image.
Maybe that's partially why the cost is higher: it's hard to tell if intermediate images are billed in addition to the output. However, this could cause an issue with the base gemimg and have it return an intermediate image instead of the final image depending on how the output is constructed, so will need to double-check.
>> - Put a strawberry in the left eye socket. >>- Put a blackberry in the right eye socket.
>> All five of the edits are implemented correctly
This is a GREAT example of the (not so) subtle mistakes AI will make in image generation, or code creation, or your future knee surgery. The model placed the specified items in the eye sockets based on the viewers left/right; when we talk relative in this scenario we usually (always?) mean from the perspective of the target or "owner". Doctors make this mistake too (they typically mark the correct side with a sharpie while the patient is still alert) but I'd be more concerned if we're "outsourcing" decision making without adequate oversight.
https://minimaxir.com/2025/11/nano-banana-prompts/#hello-nan...
There's a classic well-illustrated book, _How to Keep Your Volkswagen Alive_, which spends a whole illustrated page at the beginning building up a reference frame for working on the vehicle. Up is sky, down is ground, front is always vehicle's front, left is always vehicle's left.
Sounds a bit silly to write it out, but the diagram did a great job removing ambiguity when you expect someone to be laying on the ground in a tight place looking backwards, upside down.
Also feels important to note that in the theatre, there is stage-right and stage-left, jargon to disambiguate even though the jargon expects you to know the meaning to understand it.
>This is a GREAT example of the (not so) subtle mistakes AI will make in image generation, or code creation, or your future knee surgery.
The mistake is in the prompting (not enough information). The AI did the best it could
"What's the biggest known planet" "Jupiter" "NO I MEANT IN THE UNIVERSE!"
It doesn't affect your point but technically since the IAU are insane, exoplanets aren't technically planets and Jupiter is the largest planet in the universe.
I suppose it was too much to hope that chatbots could be trained to avoid pointless pedantry.
They've been trained on every web forum on the Internet. How could it be possible for them to avoid that?
asking "x-most known y" and not expecting a global answer is odd
No, this is squarely on the AI. A human would know what you mean without specific instructions.
Seems like you're making a judgment based on your own experience, but as another commenter pointed out, it was wrong. There are plenty of us out there who would confirm, because people are too flawed to trust. Humans double/triple check, especially under higher stakes conditions (surgery).
Heck, humans are so flawed, they'll put the things in the wrong eye socket even knowing full well exactly where they should go - something a computer literally couldn't do.
Intelligence in my book includes error correction. Questioning possible mistakes is part of wisdom.
So the understanding that AI and HI are different entities altogether with only a subset of communication protocols between them will become more and more obvious, like some comments here are already implicitly telling.
Why on earth would the fallback when a prompt is under specified be to do something no human expects?
If the instructions were actually specific, e.g. Put a blackberry in its right eye socket, then yes, most humans would know what that meant. But the instructions were not that specific: in the right eye socket
Or be even more explicit: Put a strawberry in the person’s right eye socket.
If you asked me right now what the biggest known planet was, I'd think Jupiter. I'd assume you were talking about our solar system ("known" here implying there might be more planets out in the distant reaches).
Yeah, just like humans always know what you mean.
I would not, I would clarify, and I think I'm a human.
But different humans would know what you meant differently. Some would have known it the same way the AI did.
I would be amused to see you test this theory with 100 men on the street
Right, that's why one should use "put a strawberry in the portside eye socket" and "put a strawberry in the starboard side socket"
When it doubt, always use nautical terminology
I don't know if that's so much a mistake as it is ambiguity though? To me, using the viewer's perspective in this case seems totally reasonable.
Does it still use the viewer's perspective if the prompt specifies "Put a strawberry in the _patient's left eye_"? If it does, then you're onto something. Otherwise I completely disagree with this.
“The right socket” can only be implied one way when talking about a body just like you only have one right hand despite the fact that it is on my left when looking at you.
"Plug into right power socket"
Same language, opposite meaning because of a particular noun + context.
I think the only thing obvious here is that there is no obvious solution other than adding lots of clarification to your prompt.
I think you missed the entire point?
No, they just disagree with you.
How do you disagree with having a right and a left hand?
GP is using right as in “correct”, not directionality.
No, I don't think they are.
If you are facing a wall-plate with two power sockets on it side by side and you are telling someone to plug something in, which one would be "the right socket", and which would be "the left socket"?
If above the wall-plate is a photo of a person and you are someone to draw a tattoo on the photo, which is "the right arm" and which is "the left arm"?
Same wording, different expectation.
“Eye on the left” is different from “the left eye”. First can be ambiguous, second really isn’t.
I think "the left eye" in this particular case (a photo of a skull made of pancake batter) is still very slightly ambiguous. "The skull's left eye" would not be.
I guess there's some ambiguity regarding whether or not this can be ambiguous. Because it seems like it can to me.
I meant to add a clarification to that point (because the ambiguity is a valid counterpoint), thanks for the reminder.
In case anyone missed Max's Nano Banana prompting guide, it's absolutely the definitive manual for prompting the original Nano Banana... and I tried some of the prompts in there against Nano Banana Pro and found it to be very applicable to the new model as well.
https://minimaxir.com/2025/11/nano-banana-prompts/#hello-nan...
My recreations of those pancake batter skulls using Nano Banana Pro: https://simonwillison.net/2025/Nov/20/nano-banana-pro/#tryin...
In my experience multimodal models like gpt-image-1/nano/etc. don't really require a lot of prompt trickery [1] like the good ol' days of SD 1.5.
To be clear, that's a good thing though. It's also one of the reasons why "prompt engineering" will become less relevant as model understanding goes up.
[1] - Unless you're trying to circumvent guardrails
Does the refrigerator magnet system prompt leak [1] still work?
[1] https://minimaxir.com/2025/11/nano-banana-prompts/#hello-nan....
Good call, I hadn't tried that. Here's what I got in AI Studio for:
It did NOT leak any system prompt: https://static.simonwillison.net/static/2025/nano-banana-fri...No, interestingly. (got a similar result as Simon did)
There may be more clever tricks to try and surface it though.
> it's absolutely the definitive manual
How do you know Simon? It's certainly a blog post, with content about prompting in it. If your goal is to make generative art that uses specific IP, I wouldn't use it.
Do you know of a better document specifically about prompting Nano Banana?
Why don't you just ask Gemini? It will tell you! There's no mystery.
You implied that Max's Nano Banana prompting guide wasn't the best available, so I think it's on you to provide a link to a better one.
Why would Gemini have any more insight than anyone else, let alone someone who's done hands on testing?
Minor clarification, the cost for every input image is $0.0011, not $0.06.
I was going off the footnote of "Image input is set at 560 tokens or $0.067 per image" but 560 * 2 / 1_000_000 is indeed $0.0011 so I have no idea where the $0.067 came from. Fixed, and this is why I typically don't read docs without coffee.
I would consider that a major clarification
I just pushed gemimg 0.3.2 which adds image_size support for Nano Banana Pro, and I ran a few tests on some of the images in the blog. In my testing, Nano Banana Pro correctly handled most of the image generation errors noted in my blog post: https://x.com/minimaxir/status/1991580127587921971
- Fibonacci magnets: code is correctly indented and the syntax highlighting atleast tries giving variables, numbers, and keywords different colors.
- Make me a Studio Ghibli: actually does style transfer correctly, and does it better than ChatGPT ever did.
- Rendering a webpage from HTML: near-perfect recreation of the HTML, including text layout and element sizing.
That said, there may be regressions where even with prompt engineering, the generated images which are more photorealistic look too good and land back into the uncanny valley. I haven't decided if I'm going to write a follow up blog post yet.
The system prompt hacking trick doesn't work with Nano Banana Pro unfortunately.
That result for rendering HTML to an image (the Counter Info one) is pretty impressive.
https://github.com/minimaxir/gemimg/blob/main/docs/files/cou... to this: https://x.com/minimaxir/status/1991580127587921971 - see also https://minimaxir.com/2025/11/nano-banana-prompts/#image-pro...
Your wrapper is awesome and still relevant.
> "I...worked on the detailed Nano Banana prompt engineering analysis for months"
Early in four decades of tech innovation I wasted time layering on fixes for clear deficiencies in a snowballing trend's tech offerings. If it's a big enough trend to have well funded competitors, just wait. The concern is likely not unique, and will likely be solved tomorrow.
I realized it's better to learn adaptive/defensive techniques, giving your product resilience to change. Your goal is that when surfing the change waves you can pick a point you like between rock solid and cutting edge and surf there safely.
Invest that "remediate their thing" time in "change resilience" instead – pays dividends from then on. It can be argued your tool is in this camp!
// Getting better at this also helps you with zero days.
btw you should get on their Trusted Testers program, they do give early heads up
GDM folks, get Max on!
> The model generates up to two interim images to test composition and logic. The last image within Thinking is also the final rendered image.
I've been using a bespoke Generative Model -> VLM Validator -> LLM Prompt Modifier REPL as part of my benchmarks for a while now so I'd be curious to see how this stacks up. From some preliminary testing (9 pointed star, 5 leaf clover, etc) - NB Pro seems slightly better than NB though it still seems to get them wrong. It's hard to tell what's happening under the covers.
yes they are pricey but the price will go down over time and then you can switch. vlm.run got access as early customers and are releasing it for free with unlimited generations(till they are bottlenecked by google). some results here combining image gen(Nano Banana pro) with video gen(Veo 3.1) in a single chat https://chat.vlm.run/c/1c726fab-04ef-47cc-923d-cb3b005d6262. This combined the synth generation of a person and made the puppet dance. Quite impressive
This reminds me of the journalist working for months on uncovering Trump's dirty business just for Trump himself to admit the entire thing in a tweet.
It's written to mimic that style but without meaning that the work has been done for them, just that there is new work to be done, making it an odd perhaps unconscious reference
this is pretty cool! have you found success with image editing in nano banana - i mean photoshop-like stuff. from your article i seem to wonder if nano banana is good for editing versus generating new images.
That IS the use-case for Nano Banana (as opposed to pure generative like Imagen4).
In my benchmarks, Nano-Banana scores a 7 out of 12. Seedream4 managed to outpace it, but Seedream can also introduce slight tone mapping variations. NB is the gold standard for highly localized edits.
Comparisons of Seedream4, NanoBanana, gpt-image-1, etc.
https://genai-showdown.specr.net/image-editing
I tried your "Remove all the brown pieces of candy from the glass bowl." prompt against Nano Banana Pro and it converted them to green, which I think is a pass by your criteria. Original Nano Banana had failed that test because it changed the composition of the M&Ms.
https://static.simonwillison.net/static/2025/brown-mms-remov...
Thanks Simon - I'm in the middle of re-running all my prompts through NB Pro at the moment. Nice to know it's already edged out the original. It also passed the SHRDLU test (swapping colored blocks) without cheating and just changing the colors. I'll have an update to the site shortly!
EDIT: Finished the comparisons. NB Pro scored a few more points than NB which was already super impressive.
https://genai-showdown.specr.net/image-editing?models=nb,nbp
It looks nice, what are people using the package for?
This thing's ability to produce entire infographics from a short prompt is really impressive, especially since it can run extra Google searches first.
I tried this prompt:
Here's the result: https://simonwillison.net/2025/Nov/20/nano-banana-pro/#creat...It didn’t do so well at finding middle C on a piano keyboard:
https://gemini.google.com/share/c9af8de05628
I did manage to get one image of a piano keyboard where the black keys were correct, but not consistently.
I've tried similar stuff such as: "Show a piano with an outstretched hand playing a Emaj triad on the E, G#, and B keys".
https://imgur.com/ogPnHcO
Even generating a standard piano with 7 full octaves that are consistent is pretty hard. If you ask it to invert the colors of the naturals and sharps/flats you'll completely break them.
Fooled me because it was locally correct!
It even worked really well at creating an infographic for one of my quirkier projects which doesn't have that much information online (other than its repo).
"An infographic explaining how player.html works (from the player.html project on Github). https://github.com/pseudosavant/player.html"
And then it made one formatted for social: "Change it to be an infographic formatted to fit on Instagram as a 1:1 square image."
Is the infographic accurate in terms of the way datasette wprks?
Almost entirely. I called out the one discrepancy in my post:
> “Data Ingestion (Read-Only)” is a bit off.
It’s subtly incorrect. R/w permissions for example are described incorrectly on some nodes.
Then the question becomes, can it incorporate targeted feedback, or is it a oneshot-or-bust affair?
My experience is that ChatGPT is very good at iterating on text (prose, code) but fairly bad at iterating on images. It struggles to integrate small changes, choosing instead to start over from scratch, with wildly different results. Thinking especially here of architectural stuff, where it does a great job laying out furniture in a room, but when I ask it to keep everything the same but change the colour of one piece, it goes completely off the rails.
Nano Banana is really good at iterating on images, as shown by the pancake skull example I borrowed from Max Woolf: https://simonwillison.net/2025/Nov/20/nano-banana-pro/#tryin...
I've tried iterating on slides with test on them a bit and it seems to be competent at that too.
You can use targeted feedback - but it's on the user to verify whether the edits were completely localized. In my experience NB mostly tends to make relatively surgical edits but if you're not careful it'll introduce other minute changes.
And that point you can either start over or just feather/mask with the original in any Photoshop type application.
I would assume it depends on how it generates the images.
I've used Claude to generate fairly simple icons and launch images for an iOS game and I make sure to have it start with SVG files since those can be defined as code first. This way it's easier to iterate on specific elements of the image (certain shapes need to be moved to a different position, color needs to be changed, text needs an update, etc.).
FWIW not sure how Nano Banana Pro works though.
Claude does image generation in surprising ways - we did a small evaluation [1] of different frontier models for image generation and understanding, and Claude is by far the most surprising in results.
[1] https://chat.vlm.run/showdown
[2] https://news.ycombinator.com/item?id=45996392
None of it was accurate.
But boy was it beautiful.
Funny thing to say considering the author of Datasette himself says it's accurate.
I’ve been really excited for you infographic generation. Previous models from Google and openAI had very low detail/resolution for these things.
I’ve found in general that the first generation may not be accurate but a few rolls of the dice and you should have enough to pick a style and format that works, which you can iterate on.
Did you check if the SynthID works when you edit the photos with filters like GrayScale?
Something I find weird about AI image generation models is that even though they no longer produce weird "artifacts" that give away that the fact that it was AI generated, you can still recognize that it's AI due to stylistic choices.
Not all examples they gave were like this. The example they gave of the word "Typography" would have fooled me as human-made. The infographics stood out though. I would have immediately noticed that the String of Turtles infographic was AI generated because of the stylistic choices. Same for the guide on how to make chai. I would be "suspicious" of the example they gave of the weather forecast but wouldn't immediately flag at as AI generated.
Similar note, earlier I was able to tell if something was AI generated right off the bat by noticing that it had a "Deviant Art" quality to it. My immediate guess is that certain sources of training data are over-represented.
We are just very sharp when it comes to seeing small differences in images.
I'm reminded of when the air force decided to create a pilot seat that worked for everyone. They took the average body dimensions of all their recruits and designed a seat to fit the average. It turned out, the seat fit none of their recruits. [1]
I think AI image generation is a lot like this. When you train on all images, you get to this weird sort of average space. AI images look like that, and we recognize it immediately. You can prompt or fine tune image models to get away from this, though -- the features are there it's a matter of getting them out. Lots of people trying stuff like this: https://www.reddit.com/r/StableDiffusion/comments/1euqwhr/re..., the results are nearly impossible to distinguish from real images.
[1] https://www.thestar.com/news/insight/when-u-s-air-force-disc...
What determines which “average” AI models latch onto? At a pixel level, the average of every image is a grayish rectangle; that's obviously not what we mean and AI does not produce that. At a slightly higher level, the average of every image is the average of every subject every photographed or drawn (human, tree, house, plate of food, ...) in concept space; but AI still doesn't generate a human with branches or a house with spaghetti on it. At a still higher level there are things we recognize as sensible scenes, e.g., barista pouring a cup of coffee, anime scene of a guy fighting a robot, watercolor of a boat on a lake, which AI still does not (by default) average into, say, an equal parts watercolor/anime/photorealistic image of a barista fighting a robot on a boat while pouring a cup of coffee.
But it is undeniable that AI images do have an “average” feel to them. What causes this? What is the space over which AI is taking an average to produce its output? One possible answer is that a finite model size means that the model can only explore image space with a limited resolution, and as models get bigger/better they can average over a smaller and smaller portion of this space, but it is always limited.
But that raises the question of why models don't just naturally land on a point in image space. Is this just a limitation of training, which punishes big failures more strongly than it rewards perfection? Or is there something else at play here that's preventing models from landing directly on a “real” image?
> At a pixel level, the average of every image is a grayish rectangle; that's obviously not what we mean and AI does not produce that.
That isn't correct since images in the real world aren't uniformly distributed color-wise. Take, for example, the famous ImageNet normalization magic numbers:
If it were actually uniformly distributed, the mean for each channel would be 0.5 and the standard deviation would be 0.289. Also due to z-normalization, the "image" most image models see is not how humans typically see images.I think it's because they're all trained on the same data (everything they could possibly scrape from the open web). The models tend to learn some kind of distribution of what is most likely for a given prompt. It tends to produce things that are very average looking, very "likely", but as a result also predictable and unoriginal.
If you want something that looks original, you have to come up with a more original prompt. Or we have to find a way to train these models to sample things that are less likely from their distribution? Find a way to mathematically describe what it means to be original.
Do you know of some tools with a parameter that asks it to be "weird" and increase diversity of outputs?
If you ever had a pinterest account and a deviant art account, all becomes clear.
It still has some artifacts more often than not, they are a lot subtler in nature but they still come out, whether it's texture, proportion, lighting, or perspective. Now some things are easier to fix on second pass edits, some are not. I guess it's why they consider image editing to be the next challenge.
It's a bit odd to say, but another big clue identifying something as AI-generated is that it simply looks "too good" for what it is being used for. If I see a little info graphic demonstrating something relatively mundane, and it has nice 3D rendered characters or graphical elements, at this point it's basically guaranteed to be AI, because you just sort of intuitively know when something would've justified the human labor necessary to produce that.
Funny enough that had crossed my mind with the woodchuck example, because at a glance I can't see any weird artifacts, but I felt confident I could tell it was AI generated immediately if I saw it in the wild, and I couldn't really explain why. My immediate guess was "well, who the hell would actually bother to make something like this?"
It's not odd to say. It was one of the first telling signs to identify AI artists[0] on Twitter: overly detailed backgrounds.
Of course now a lot of them have learned the lesson and it's much harder to tell.
[0]: I know, I know...
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.
Some days it feels like I'm the only hacker left who doesn't want government mandated watermarking in creative tools. Were politicians 20 years ago as overreative they'd have demanded Photoshop leave a trace on anything it edited. The amount of moral panic is off the charts. It's still a computer, and we still shouldn't trust everything we see. The fundamentals haven't changed.
> It's still a computer, and we still shouldn't trust everything we see. The fundamentals haven't changed.
I think that by now it should be crystal clear to everyone that it matters a lot the sheer scale a new technology permits for $nefarious_intent.
Knives (under a certain size) are not regulated. Guns are regulated in most countries. Atomic bombs are definitely regulated. They can all kill people if used badly, though.
When a photo was faked/composed with old tech, it was relatively easy to spot. With photoshop, it became more complicated to spot it but at the same time it wasn't easy to mass-produce altered images. Large models are changing the rules here as well.
I think we're overreacting. Digital fakes will proliferate, and we'll freak out bc it's new. But after a certain amount of time, we'll just get used to it and realize that the world goes on, and whatever major adverse effects actually aren't that difficult to deal with. Which is not the case with nuclear proliferation or things like that.
The story of human history is newer generations freaking about progress and novel changes that have never been seen before. And later generations being perfectly okay with it and adapting to a new style of life.
In general I concur but the adaptation doesn't come out of the blue or just only because people get used to it but also because countermeasures are taken, regulations are written and adjustments are made to reduce the negative impact. Also the hyperconnected society is still relatively new and I'm not sure we have adapted for it yet.
I think the long term effect will be that photos and videos no longer have any evidentiary value legally or socially, absent a trusted chain of custody.
It shouldn’t be that we panic about it and regulate the hell out.
We could use the opportunity to deploy robust systems of verification and validation to all digital works. One that allows for proving authenticity while respecting privacy if desired. For example… it’s insane in the US we revolve around a paper social security number that we know damn well isn’t unique. Or that it’s a massive pain in the ass for most people to even check the hash of a download.
Guess which we’ll do!
> a new technology permits for $nefarious_intent
But people with actual nefarious intent will easily be able to remove these watermarks, however they're implemented. This is copy protection and key escrow all over again - it hurts honest people and doesn't even slow down bad people.
> Knives (under a certain size) are not regulated. Guns are regulated in most countries. Atomic bombs are definitely regulated
I don’t think this is a good comparison: knives are easy to produce, guns a bit harder, atomic bombs definitely harder. You should find something that is as easy to produce as a knife, but regulated.
The "product" to be regulated here is the LLM/model itself, not its output.
Or, if you see the altered photo as the "product", then the "product" of the knife/gun/bomb is the damage it creates to a human body.
>You should find something that is as easy to produce as a knife, but regulated.
The DEA and ATF have entered the chat
They can leave, plain water fits this bill.
Politicians absolutely were doing this 20-30 years ago. Plenty of folks here are old enough to remember debates on Slashdot around the Communications Decency Act, Child Online Protection Act, Children's Online Privacy Protection Act, Children's Internet Protection Act, et al.
https://en.wikipedia.org/wiki/Communications_Decency_Act
It’s annoying how effective “for the children” is. That peiole really just turn off their brains for that.
Easy to say until it impacts you in a bad way:
https://www.nbcnews.com/tech/tech-news/ai-generated-evidence...
> “My wife and I have been together for over 30 years, and she has my voice everywhere,” Schlegel said. “She could easily clone my voice on free or inexpensive software to create a threatening message that sounds like it’s from me and walk into any courthouse around the country with that recording.”
> “The judge will sign that restraining order. They will sign every single time,” said Schlegel, referring to the hypothetical recording. “So you lose your cat, dog, guns, house, you lose everything.”
At the moment, the only alternative is courts simply never accept photo/video/audio as evidence. I know if I were a juror I wouldn't.
At the same time, yeah, watermarks won't work. Sure, Google can add a watermark/fingerprint that is impossible to remove, but there will be tools that won't put such watermarks/fingerprints.
Testimony is evidence. I don't think most cases have any physical evidence.
A lot of cases rely heavily on security camera footage.
I suspect watermarking ends up being a net negative, as people learn to trust that lack of a watermark indicates authenticity. Propaganda won’t have the watermark.
Unless they've recently changed it, Photoshop will actually refuse to open or edit images of at least US banknotes.
You do know that every color copier comes with the ability to identify US currency and would refuse to copy it? And that every color printer leaves a pattern of faint yellow dots on every printout that uniquely identifies the printer?
Is this something strictly with the US currency notes or is the same true for other countries currency as well?
It's most notes, and for EU and US notes (as well as some others), it's based on a certain pattern on the bills: https://en.wikipedia.org/wiki/EURion_constellation
And that's not a good thing.
Nope, having a stable, trusted currency trumps whatever productive use one could have for a anonymous, currency reproducing color printer
I'm just responding to this by OP:
> Were politicians 20 years ago as overreative they'd have demanded Photoshop leave a trace on anything it edited.
Why not? Like, genuinely.
I generally don't think that's it's good or just for a government to collude with manufacturers to track/trace it's citizens without consent or notice. And even if notice was given, I'd still be against it
The arguments put forward by people generally I don't find compelling -- for example, in this thread around protecting against counterfeit.
The "force" applied to address these concerns is totally out of proportion. Whenever these discussions happen, I feel like they descend into a general viewpoint, "if we could technically solve any possible crime, we should do everything in our power to solve it."
I'm against this viewpoint, and acknowledge that that means _some crime_ occurs. That's acceptable to me. I don't feel that society is correctly structured to "treat" crime appropriately, and technology has outpaced our ability to holistically address it.
Generally, I don't see (speaking for the US) the highest incarceration rate in the world to be a good thing, or being generally effective, and I don't believe that increasing that number will change outcomes.
It depends on how you're looking at it. For the people not getting handed counterfeit currency, it's probably a good thing.
Also probably good for the people trying to counterfeit money with a printer, better not to end up in jail for that.
Try photocopying some US dollar bills.
HN is full of authoritarian bootlickers who can't imagine that people can exist without a paternalistic force to keep them from doing bad things.
I'm sure Apple will roll something out in the coming years. Now that just anyone can easily AI themselves into a picture in front of the Eiffel tower, they'll want a feature that will let their users prove that they _really_ took that photo in front of the Eiffel tower (since to a lot of people sharing that you're on a Paris vacation is the point, more than the particular photo).
I bet it will be called "Real Photos" or something like that, and the pictures will be signed by the camera hardware. Then iMessage will put a special border around it or something, so that when people share the photos with other Apple users they can prove that it was a real photo taken with their phone's camera.
Does anyone other than you actually care about your vacation photos?
There used to be a joke about people who did slideshows (on an actual slide projector) of their vacation photos at parties.
> a real photo taken with their phone's camera
How "real" are iPhone photos? They're also computationally generated, not just the light that came through the lens.
Even without any other post-processing, iPhones generate gibberish text when attempting to sharpen blurry images, they delete actual textures and replace them with smooth, smeared surfaces that look like a watercolor or oil paintings, and combine data from multiple frames to give dogs five legs.
Don’t be a pedant. You know very well there is a big different between a photo taken on an iPhone and a photo edited with Nano Banana.
If there was a standardized identifier, there would be software dedicated to just removing it.
I don't see how it would defeat the cat and mouse game.
It doesn't have to be perfect to be helpful.
For example, it's trivial to post an advertisement without disclosure. Yet it's illegal, so large players mostly comply and harm is less likely on the whole.
You'd need a similar law around posting AI photos/videos without disclosure. Which maybe is where we're heading.
It still won't prevent it, but it would prevent large players from doing it.
I don't think it will be easy to just remove it. It's built into the image and thus won't be the same every time.
Plus, any service good at reverse-image search (like Google) can basically apply that to determine whether they generated it.
There will always be a way to defeat anything, but I don't see why this won't work for like 90% of cases.
> I don't think it will be easy to just remove it.
No, but model training technology is out in the open, so it will continue to be possible to train models and build model toolchains that just don't incorporate watermarking at all, which is what any motivated actor seeking to mislead will do; the only thing watermarking will do is train people to accept its absence as a sign of reliability, increasing the effectiveness of fakes by motivated bad actors.
It's an image. There's simply no way to add a watermark to an image that's both imperceptible to the user and non-trivial to remove. You'd have to pick one of those options.
That is patently false.
So, uh... do you know of an implementation that has both those properties? I'd be quite interested in that.
https://arxiv.org/html/2502.10465v1
I'm not sure that's correct. I'm not an expert, but there's a lot of literature on digital watermarks that are robust to manipulation.
It may be easier if you have an oracle on your end to say "yes, this image has/does not have the watermark," which could be the case for some proposed implementations of an AI watermark. (Often the use-case for digital watermarks assumes that the watermarker keeps the evaluation tool secret - this lets them find, e.g, people who leak early screenings of movies.)
> I don't think it will be easy to just remove it.
Always has been so far. You add noise until the signal gets swamped. In order to remain imperceptible it's a tiny signal, so it's easy to swamp.
You could probably just stick your image in another model or tool that didn't watermark and have it regenerate the image as accurately as possible.
Exactly, a diffusion model can denoise the watermark out of the image. If you wanted to be doubly sure you could add noise first and then denoise which should completely overwrite any encoded data. Those are trivial operations so it would be easy to create a tool or service explicitly for that purpose.
It would be like standardizing a captcha, you make a single target to defeat. Whether it is easy or hard is irrelevant.
There will be a model trained to remove synthids from graphics generated by other models
The incentive for commercial providers to apply watermarks is so that they can safely route and classify generated content when it gets piped back in as training or reference data from the wild. That it's something that some users want is mostly secondary, although it is something they can earn some social credit for by advertising.
You're right that there will existed generated content without these watermarks, but you can bet that all the commercial providers burning $$$$ on state of the art models will gradually coalesce around some means of widespread by-default/non-optional watermarking for content they let the public generate so that they can all avoid drowning in their own filth.
This is what C2PA is trying to do: https://c2pa.org/
I don't understand why there isn't an obvious, visible watermark at all. Yes, one could remove it but let's assume 95% of people don't bother removing the visible watermark. It would really help with seeing instantly when an image was AI generated.
SynthID has been in use for over 2 years.
Regardless of how you feel about this kind of steganography, it seems clear that outside of a courtroom, deepfakes still have the potential to do massive damage.
Unless the watermark randomly replaces objects in the scene with bananas, these images/videos will still spread like wildfire on platforms like TikTok, where the average netizen's idea of due diligence is checking for a six‑fingered hand... at best.
It solves some problems! For example, if you want to run a camgirl website based on AI models and want to also prove that you're not exploiting real people
> It solves some problems! For example, if you want to run a camgirl website based on AI models and want to also prove that you're not exploiting real people
So, you exploit real people, but run your images through a realtime AI video transformation model doing either a close-to-noop transformation or something like changing the background so that it can't be used to identify the actual location if people do figure out you are exploiting real people, and then you have your real exploitation watermarked as AI fakery.
I don't think this is solving a problem, unless you mean a problem for the would-be exploiter.
Your use case doesn't even make sense. What customers are clamoring for that feature? I doubt any paying customer in the market for (that product) cares. If the law cares, the law has tools to inquire.
All of this is trivially easy to circumvent ceremony.
Google is doing this to deflect litigation and to preserve their brand in the face of negative press.
They'll do this (1) as long as they're the market leader, (2) as long as there aren't dozens of other similar products - especially ones available as open source, (3) as long as the public is still freaked out / new to the idea anyone can make images and video of whatever, and (4) as long as the signing compute doesn't eat into the bottom line once everyone in the world has uniform access to the tech.
The idea here is that {law enforcement, lawyers, journalists} find a deep fake {illegal, porn, libelous, controversial} image and goes to Google to ask who made it. That only works for so long, if at all. Once everyone can do this and the lookup hit rates (or even inquiries) are < 0.01%, it'll go away.
It's really so you can tell journalists "we did our very best" so that they shut up and stop writing bad articles about "Google causing harm" and "Google enabling the bad guys".
We're just in the awkward phase where everyone is freaking out that you can make images of Trump wearing a bikini, Tim Cook saying he hates Apple and loves Samsung, or the South Park kids deep faking each other into silly circumstances. In ten years, this will be normal for everyone.
Writing the sentence "Dr. Phil eats a bagel" is no different than writing the prompt "Dr. Phil eats a bagel". The former has been easy to do for centuries and required the brain to do some work to visualize. Now we have tools that previsualize and get those ideas as pixels into the brain a little faster than ASCII/UTF-8 graphemes. At the end of the day, it's the same thing.
And you'll recall that various forms of written text - and indeed, speech itself - have been illegal in various times, places, and jurisdictions throughout history. You didn't insult Caesar, you didn't blaspheme the medieval church, and you don't libel in America today.
> What customers are clamoring for that feature? If the law cares, the law has tools to inquire.
How can they distinguish from real people exploited to AI models autogenerating everything?
I mean right now this is possible, largely because a lot of the AI videos have shortcomings. But imagine in 5 years from now on ...
> How can they distinguish from real people exploited to AI models autogenerating everything?
Watermarking by compliant models doesn't help this much because (1) models without watermarking exist and can continue to be developed (especially if absence of a watermark is treated as a sign of authenticity), so you cannot rely on AI fakery being watermarked, and (2) AI models can be used for video-to-video generation without changing much of the source, so you can't rely on something accurately watermarked as "AI-generated" not being based in actual exploitation.
Now, if the watermarking includes provenance information, and you require certain types of content to be watermarked not just as AI using a known watermarking system, but by a registered AI provider with regulated input data safety guardrails and/or retention requirements, and be traceable to a registered user, and...
Well, then it does something when it is present, largely by creating a new content gatekeepiing cartel.
> How can they distinguish from real people exploited to AI models autogenerating everything?
The people who care don't consume content which even just plausibly looks like real people exploited. They wouldn't consume the content even if you pinky promised that the exploited looking people are not real people. Even if you digitally signed that promise.
The people who don't care don't care.
It would be more productive for camera manufacturers to embed a per-device digital signature. Those care to prove their image is genuine could publish both pre and post processed images for transparency.
This watermarking ceremony is useless.
We will always have local models. Eventually the Chinese will release a Nano Banana equivalent as open source.
Qwen-Image-Edit is pretty good already: https://simonwillison.net/2025/Aug/19/qwen-image-edit/
Qwen won the latest models round last month…
https://generative-ai.review/2025/09/september-2025-image-ge... (non-pro Nano Banana)
> We will always have local models.
If watermarking becomes a legal mandate, it will inevitably include a prohibition on distributing (and using and maybe even possessing, but the distribution ban is the thing that will have the most impact, since it is the part that is most policable, and most people aren't going to be training their own models, except, of course, the most motivated bad actors) open models that do not include watermarking as a baked-in model feature. So, for most users, it'll be much less accessible (and, at the same time, it won't solve the problem.)
yeah, personally identifying undetectable watermarks are kindof a terrifying prospect
It is terrifying, but inevitable. Perhaps AI companies flooding the commons with excrement wasn't the best idea, now we all have to suffer the consequences.
Reminder that even in the hypothetical world where every AI image is digitally watermarked, and all cameras have a TPM that writes a hash of every photo to the blockchain, there’s nothing to stop you from pointing that perfectly-verified camera at a screen showing your perfectly-watermarked AI image and taking a picture.
Image verification has never been easy. People have been airbrushed out of and pasted into photos for over a century; AI just makes it easier and more accessible. Expecting a “click to verify” workflow is unreasonable as it has ever been; only media literacy and a bit of legwork can accomplish this task.
Competent digital watermarks usually survive the 'analog hole'. Screen-cam resistant watermarks have been in use since at least 2020, and if memory serves, back to 2010 when I first starting reading about them, but I don't recall what it was called back then.
I just tried asking Gemini about a photo I took of my screen showing an image I edited with Nano Banana Pro... and it said "All or part of the content was generated with Google AI. SynthID detected in less than 25% of the image".
Photo-of-a-screen: https://gemini.google.com/share/ab587bdcd03e
It reported 25-50% for the image without having been through that analog hole: https://gemini.google.com/share/022e486fd6bf
We need to be super careful with how legislation around this is passed and implemented. As it currently stands, I can totally see this as a backdoor to surveillance and government overreach.
If social media platforms are required by law to categorize content as AI generated, this means they need to check with the public "AI generation" providers. And since there is no agreed upon (public) standard for imperceptible watermarks hashing that means the content (image, video, audio) in its entirety needs to be uploaded to the various providers to check if it's AI generated.
Yes, it sounds crazy, but that's the plan; imagine every image you post on Facebook/X/Reddit/Whatsapp/whatever gets uploaded to Google / Microsoft / OpenAI / UnnamedGovernmentEntity / etc. to "check if it's AI". That's what the current law in Korea and the upcoming laws in California and EU (for August 2026) require :(
I don't believe that you can do this for photography. For AI-images, if the embedded data has enough information (model identification and random seed), one can prove that it was AI by recreating it on the fly and comparing. How do you prove that a photographic image was created by a CCD? If your AI-generated image were good enough to pass, then hacking hardware (or stealing some crypto key to sign it) would "prove" that it was a real photograph.
Hell, it might even be possible for some arbitrary photographs to come up with an AI prompt that produces them or something similar enough to be indistinguishable to the human eye, opening up the possibility of "proving" something is fake even when it was actually real.
What you want just can't work, not even from a theoretical or practical standpoint, let alone the other concerns mentioned in this thread.
It solves a real problem - if you have something sketchy, the big players can repudiate it, the authorities can more formally define the black market, and we can have a ‘war on deepfakes’ to further enable the authorities in their attempts to control the narratives.
Labelling open source models as "grey market" is a heck of a presumption
Every model is "grey market". They're all trained on data without complying with any licensing terms that may exist, be they proprietary or copyleft. Every major AI model is an instance of IP theft.
It's why I used "scare quotes".
I asked Gemini "dymamic view" how SynthID works: https://gemini.google.com/share/62fb0eb38e6b
I feel like I am going crazy or missed something simple but when I use the Gemini app and I ask it to edit a photo that I upload, 2.5 flash works really well but 2.5 pro or 3.0 pro do a very poor job. I uploaded an image of me and asked it to make me bald and flash did a great job of just changing me in the photo but 3.0 pro took me out of the photo completely and just created a headshot of a bald man that only sort of resembled me. Am I missing something or does paying for the pro version not give you anything over the 2.5 flash model?
The code name “nano banana” model is based on the Flash 2.5 foundation. Until today it was the “latest and greatest”.
Developer Blog: https://blog.google/technology/developers/gemini-3-pro-image...
DeepMind 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...
This is the first image model I’ve used that passed my piano test. It actually generated an image of a keyboard with the proper pattern of black keys repeated per octave – every other model I’ve tried this with since the first Dall-E has struggled to render more than a single octave, usually clumping groups of two black keys or grouping them four at a time. Very impressive grasp of recursive patterns.
If you ask it for anything outside of the standard 88 key set it falls short. For instance
"Generate a piano, but have the left most key start at middle C, and the notes continue in the standard order up (D, E, F, G, ...) to the right most key"
The above prompt will be wrong, seemingly every time. The model has no understanding of the keys or where they belong, and it is not able to intuit creating something within the actual confines of how piano notes are patterned.
"Generate a piano but color every other D key red"
This also wrong, every time, with seemingly random keys being colored.
I would imagine that a keyboard is difficult to render (to some extent) but I also don't think its particularly interesting since it is a fully standardized object with millions of pictures from all angles in existence to learn from right?
Yep - one of my goto bench marks is a "historical piano" - meaning the naturals are black and the sharps/flats are white.
https://imgur.com/a/SZbzsYv
Periodic motion (groups of repeating patterns) always tend to degrade at some point. Maintaining coherence over 88 keys is impressive.
You can try it out for free on LMArena [0]: New Chat -> Battle dropdown -> Direct Chat -> Click on Generate Image in the chat box -> Click dropdown from hunyuan-image-3.0 -> gemini-3-pro-image-preview (nano-banana-pro).
I've only managed to get a few prompts to go through, if it takes longer than 30 seconds it seems to just time out. Image quality seems to vary wildly; the first image I tried looked really good but then I tried to refresh a few times and it kept getting worse.
[0] lmarena.ai/
Thanks - this worked for me (some errors, some success).
Last week I was making a birthday card for my son with the old model. The new model is dramatically better - I'm asking for an image in comic book style, prompted with some images of him.
With the previous model, the boy was descriptively similar (e.g. hair colour and style) but looked nothing like him. With this model it's recognisably him.
When I do that, I get two (very similar but not identical) responses side-by-side in one image (I guess as if the model is battling itself?). Is that normal for lmarena?
https://imgur.com/a/h0ncCFN
I've had nano banana pro for a few weeks now, and it's the most impressive AI model I've ever seen
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!)
Genuinely believe that images are 99.5% solved now and unless you’re extremely keen eyed, you won’t be able to tell AI images from real images now
I'd be curious about how well the inline verification works - an easy example is to have it generate a 9-pointed star, a classic example that many SOTA models have difficulties with.
In the past, I've deliberately stuck a Vision-language model in a REPL with a loop running against generative models to try to have it verify/try again because of this exact issue.
EDIT: Just tested it in Gemini - it either didn't use a VLM to actually look at the finished image or the VLM itself failed.
Output:
Result:How did you get early access?!
"Inline verification of images following the prompt is awesome, and you can do some _amazing_ stuff with it." - could you elaborate on this? sounds fascinating but I couldn't grok it via the blog post (like, it this synthid?)
It uses Gemini 3 inline with the reasoning to make sure it followed the instructions before giving you the output image
LLMs might be a dead end, but we're going to have amazing images, video, and 3D.
To me the AI revolution is making visual media (and music) catch up with the text-based revolution we've had since the dawn of computing.
Computers accelerated typing and text almost immediately, but we've had really crude tools for images, video, and 3D despite graphics and image processing algorithms.
AI really pushes the envelope here.
I think images/media alone could save AI from "the bubble" as these tools enable everyone to make incredible content if you put the work into it.
Everyone now has the ingredients of Pixar and a music production studio in their hands. You just need to learn the tools and put the hours in and you can make chart-topping songs and Hollywood grade VFX. The models won't get you there by themselves, but using them in conjunction with other tools and understanding as to what makes good art - that can and will do it.
Screw ChatGPT, Claude, Gemini, and the rest. This is the exciting part of AI.
How can LLMs be a dead end? The last improvement in LLMs came out this week.
I wouldn’t call LLMs a dead end, they’re so useful as-is
LLMs are useful, but they've hit a wall on the path to automating our jobs. Benchmark scores are just getting better at test taking. I don't see them replacing software engineers without overcoming obstacles.
AI for images, video, music - these tools can already make movies, games, and music today with just a little bit of effort by domain experts. They're 10,000x time and cost savers. The models and tools are continuing to get better on an obvious trend line.
I'm literally a software engineer, and a business owner. I don't think about this in binary terms (replacement or not), but just like CMS's replaced the jobs of people that write HTML by hand to build websites, I think whole classes of software development will get democratized.
For example, I'm currently vibe coding an app that will be specific to our company, that helps me run all the aspects of our business and integrates with our systems (so it'll integrate with quickbooks for invoicing, etc), and help us track whether we have the right insurance across multiple contracts, will remind me about contract deadlines coming up, etc.
It's going to combine the information that's currently in about 10 different slightly out of sync spreadsheets, about 2 dozen google docs/drive files, and multiple external systems (Gusto, Quickbooks, email, etc).
Even though I could build all this manually (as a software developer), I'd never take the time to do it, because it takes away from client work. But now I can actually do it because the pace is 100x faster, and in the background while I'm doing client work.
Doesn’t seem like a dead end at all. Once we can apply LLMs to the physical world and its outputs control robot movements it’s essentially game over for 90% of the things humans do, AGI or not.
It's crazy how good these models are at text now. Remember when text was literally impossible? Now the models can diagetically render any text. It's so good now that it seems like a weird blip that it _wasn't_ possible before.
Not to mention all the other stuff.
I agree, it's improving by leaps. I'm still patiently awaiting for my niche use of creating new icons though, one that can match the existing curvature, weight, spacing, and balance. It seems AI is struggling in the overlap of visuals <-> code, or perhaps there's less business incentive to train on that front. I know the pelican on bicycle svg is getting better, but still really rough looking and hard to modify with prompt versus just spending some time upfront to do it yourself in an editor.
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
> 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
No, its not the beginning, multiple different watermarking standards, watermark checking systems, and, of course, published countermeasures of various effectiveness for most of them, have been around for a while.
I don't understand the excitement around generating and/or watching AI-produced videos. To me it's probably the single most uninteresting and boring thing related to AI that I can think of. What is the appeal?
Pretty sure Nano Banana only produces images.
Nonetheless, ask it to “create an infographic on how Google works”. Do you not see any excitement in the result? I think it’s pretty impressive and has a lot of utility.
As a general content I agree it's a bit off putting, but I find it a lot of fun when generating content among friends like internal jokes and educational content. I got my kid to drink some meds by generating an image of a hero telling him it's important to take.
Do you feel the same way about VFX (marvel etc) or animated movies (pixar etc)
Sometimes, an animation is the best way to convey information.
Does anyone know if this is predicting the entire image at once, or if it's breaking it into constituent steps i.e. "draw text in this font at this location" and then composing it from those "tools"? It would be really interesting if they've solved the garbled text problem within the constraint of predicting the entire image at once.
I strongly suspect it's the latter, though someone please chime in if I'm wrong.
Even so, this is a real advancement. It's impressive to see existing techniques combined to meaningfully improve on SOTA image generation.
The previous nano banana was using composing tools. It was really obvious by some of the janky outputs it made. Not sure about this one, but presumably they built off it.
There still is some garbled text sometimes so it can't be the latter (try to get it to generate a map of 48 us states labeled - the ones that are too small to write on and need arrows were garbled (1 attempt))
I’m pretty sure, but no expert on the matter, that correct text rendering was solved by feeding in bitmaps of rasterized fonts as supplemental context to the image generation models.
Google needs to pace themselves. AI studio, Antigravity, Banana, Banana Pro, Grape Ultra, Gemini 3, etc. This information overload don't do them any good whatsoever.
Why? They're mostly different markets. Most people using Nano Banana Pro aren't using Antigravity.
A cluster of launches reinforces the idea that Google is growing and leading in a bunch of areas.
In other words, if it's having so many successes it feels like overload, that's an excellent narrative. It's not like it's going to prevent people from using the tools.
Google will never beat the "sunset after 2 years" allegations on all products that don't have "Google __" in the name
It reminds me of AWS services: I can't tell what they are because they've been named by a monkey with a typewriter.
Powell Doctrine, but for AI. No one should dispute that Google is the leader in every(?) category of AI: LLM, image gen, video editing, world models, etc.
This cluster of launches might not be intentional. It could just be a bunch of independent teams all trying to get their launches out before the EOY deadline.
Stock market seems to agree with their strategy....
Maybe? or lemmings following BH purchase of $4B in Google stock this week assuming "Buffett only buys value stocks; it must be ready to grow!"
https://finance.yahoo.com/news/warren-buffetts-berkshire-hat...
... and has a tendency to disagree past the Peak of Inflated Expectations.
I feel it's strategic, like a massive DDoS/"shock and awe" style attack on competitors. Gotta love it as PROsumers though!
Agree. I can't keep up with it, it's hard to grasp my head around them, where to go to actually use them, etc
Grape Ultra?
That part was a joke to illustrate the point.
Jules, Vertex...
They are riding the current buzzword wave. It'll eventually subside. And 80% of it will end up on Google's impressive software graveyard:
https://killedbygoogle.com/
I guess the true endgame of AI products is naming them. We still have quite a way to go.
We just need a new AI for that.
Need a name for something? Try our new Mini Skibidi model!
Also introducing the amazing 6-7 pro model
This has always been the hardest problem in computer science besides “Assume a lightweight J2EE distribution…”
I was at a tech conference yesterday, and I asked someone if they had tried nano banana. They looked at me like I was crazy. These names aren't helping! (But honestly I love it, easier to remember than Gemini-2.whatever.
Honestly I give Google credit for realizing that they had something that people were talking about and running with it instead of just calling it gemini-image-large-with-text-pro
They tried calling it gemini-2.5-whatever, but social media obsessed over the name "Nano Banana", which was just its codename that got teased on Twitter for a few weeks prior to launch.
After launch, Google's public branding for the product was "Gemini" until Google just decided to lean in and fully adopt the vastly more popular "Nano Banana" label.
The public named this product, not Google. Google's internal codename went virally popular and outstaged the official name.
Branding matters for distribution. When you install yourself into the public consciousness with a name, you'd better use the name. It's free distribution. You own human wetware market share for free. You're alive in the minds of the public.
Renaming things every human has brand recognition of, eg. HBO -> Max, is stupid. It doesn't matter if the name sucks. ChatGPT as a name sucks. But everyone in the world knows it.
This will forever be Nano Banana unless they deprecate the product.
There are only 2 hard problems in computer science: cache coherency, naming things and off by 1 errors...
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 have this problem selecting Pro, but if I use 2.5 Flash it does a great job at these things. I am not sure why Pro does not work as well.
I tried this in AI studio just now with nano banana.
Results: https://imgur.com/a/9II0Aip
The white house was the original (random photo from Google). The prompt was "What paint color would look nice? Paint the house."
> (random photo from Google)
Careful with that kind of thing.
Here, it mostly poisons your test, because that exact photo probably exists in the underlying training data and the trained network will be more or less optimized on working with it. It's really the same consideration you'd want to make when testing classifiers or other ML techs 10 years ago.
Most people taking to a task like this will be using an original photo -- missing entirely from any training date, poorly framed, unevenly lit, etc -- and you need to be careful to capture as much of that as possible when trying to evaluate how a model will work in that kind of use case.
The failure and stress points for AI tools are generally kind of alien and unfamiliar because the way they operate is totally different than the way a human operates, and if you're not especially attentive to their weird failure shapes and biases when you want to test them, or you'll easily get false positives (and false negatives) that lead you to misleading conclusions.
Yea, the base image was the first google image result for the search term "house". So definitely in the training set.
> The prompt was "What paint color would look nice? Paint the house."
At some point, this is probably gonna result in you coming home to a painted house and a big bill, lol.
Guess they ran out of paint - notice the upper window.
Oops. Original link wasn't using the Pro version. Edited the comment with an updated link.
I also tried that in the past with poor results. I just tried it this morning with nano banana pro and it nailed it with a very short prompt: "Repaint the house white with black trim. Do not paint over brick."
I don't know what it is with Gemini (and even other models) but I swear they must be doing some kind of active load-dependant quanitization or a/b/c/d testing behind the scenes, because sometimes the model is stellar and hitting everything, and other times it's tripping all over itself.
The most effective fix I have found is that when the model is acting dumb, just turn it off and come back in the few hours to a new chat and try again.
Yeah I think they all shed under heavy load as part of some scaling strategy.
Huh, can you share a link? I tried here: https://gemini.google.com/share/e753745dfc5d
https://gemini.google.com/share/79fe1a38e440
https://gemini.google.com/share/3b4d2cd55778
Maybe somewhere in the original comment it would have been fair to mention you can barely see the house in the original photo. This is actually a hilarious complaint
That cannot be a valid excuse. Other than adding extra windows to the clearly visible wall, it's obvious that model perfectly capable to "see" the house. It just cannot "believe" that there can be a big empty wall on a garden house.
Maybe. But this is not an edge case. I consider this genuine use of the marketed tool.
Nano Banana Pro is a chatGPT 3.5 to 4 tier leap.
Everyone who worked on this is a traitor to the human race. Why do we need to make it impossible to make a living as an artist? Who thinks an endless tsunami of garbage “content” churned out by machines dropping the bottom out of all artistic disciplines is a good idea?
Capitalism, at work. Wherever there is a cost, there will be attempts made at cost efficiency. Google understands that hiring designers or artists is expensive, and they want to offer a cheaper, more effective alternative so that they can capture the market.
In a coffee shop this morning I saw a lady drawing tulips with a paper and pencil. It was beautiful, and I let her know... But as I walked away I felt sad that I don't feel that when browsing online anymore- because I remember how impressive it used to feel to see an epic render, or an oil painting, etc... I've been turned cynical.
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.
Nano Banana is focused on editing. But the Pro version handles your prompt much better. First image is Pro, second is 2.5
https://imgur.com/a/3PDUIQP
Wow, that top image is actually quite good! Interestingly, I just got into Pro and got a worse result than yours. https://imgur.com/a/ENNk68B ... and it really seems to just vary by attempt even with the exact same prompt.
Ooh, I just got offered the new version on https://gemini.google.com/. Plugged in that exact prompt, got this:
https://imgur.com/a/ENNk68B
Much better than previous attempts. Still has an extra lane with the cars on the right cutting off the cars in the middle. Still has the numbers in the wrong order.
I'd try a some more if I were you. I saw an example of generated infographic that was greatly improved over anything I've seen an image generator do before. What you desire seems in the realm of possibility.
I think you tried using the wrong tool. Nano Banana is for editing, not generating (there's Imagen for that).
Imagen4 did no better. edit: example https://imgur.com/Dl8PWgm with a so-so result: four lanes, cars at least facing the same way, lane block looks good, weird extra division in the center, some numbers repeated, one arrow going straight into construction, one arrow going backwards
edit: or Imagen4 Ultra. https://imgur.com/a/xr2ElXj cars facing opposite directions within a lane, 2-way (4 lanes total), double-ended arrows, confused disaster. pretty though.
One of the things I've always been curious about is how effective diffusion models can be for web and app design. They're generally trained on more organic photos, but post-training on SDXL and Flux have given me good results here in the past (with the exception of text).
It's been interesting seeing the results of Nano Banana Pro in this domain. Here are a few examples:
Prompt: "A travel planner for an elegant Swiss website for luxury hiking tours. An interactive map with trail difficulty and booking management. Should have a theme that is alpine green, granite grey, glacier white"
Flux output: https://fal.media/files/rabbit/uPiqDsARrFhUJV01XADLw_11cb4d2...
NBP output: https://v3b.fal.media/files/b/panda/h9auGbrvUkW4Zpav1CnBy.pn...
---
Prompt: "a landing page for a saas crypto website, purple gradient dark theme. Include multiple sections, including one for coin prices, and some graphs of value over time for coins, plus a footer"
Flux output: https://fal.media/files/elephant/zSirai8mvJxTM7uNfU8CJ_109b0...
NBP output: https://v3b.fal.media/files/b/rabbit/1f3jHbxo4BwU6nL1-w6RI.p...
---
Prompt: "product launch website for a development tool, dark background with aqua blue and neon gold highlights, gradients"
Flux output: https://fal.media/files/zebra/aXg29QaVRbXe391pPBmLQ_4bfa61cc...
NBP output: https://v3b.fal.media/files/b/lion/Rj48BxO2Hg2IoxRrnSs0r.png
---
Note that this is with a lora I built for flux specifically for website generation. Overall, nbp seems to have less creative / inspired outputs, but the text is FAR better than the fever dream Flux is producing. I'm really excited to see how this changes design. At the very least it proved it can get close to a production quality for output, now it's just about tuning it.
In my limited testing, at least in terms of maintaining consistency between input and output for Asian faces, it has even regressed.
Actually, Gemini 3 is about the same, and doesn't feel as good as Claude 4.5. I have a feeling it's been fine-tuned for a cool front-end marketing effect.
Furthermore, I really don't understand why AI Studio, now requiring me to use its own API for payment, still adds a watermark.
There's some really impressive things about this (the speed, the lack of typical AI image gen artifacts) but it also seems less creative than other models I've tried?
"mountain dew themed pokemon" is the first search prompt I always try with new image models and Nano Banna Pro just gave me a green pikachu.
Other models do a much better job of creating something new.
IMHO I'd rather them focus on strong literal prompt adherence so that more detailed prompts produce more accurate results.
That way you can stick your choice of any number of LLM preprocessors in front of a generic prompt like "mountain dew themed pokemon" and push the responsibility of creating a more detailed prompt upstream.
https://imgur.com/a/s5zfxS5
Note: I'm not particularly impressed with either of the results - this is more a demonstration.
This is what the SynthID signature looks like on Nano Banana images https://www.reddit.com/r/nanobanana/comments/1o1tvbm/nano_ba...
And if it can be seen like that, it should be removeable too. There are more examples in that thread.
Just last night I was using Gemini "Fast" to test its output for a unique image we would have used in some consumer research if there had been a good stock image back in the day. I have been testing this prompt since the early days of AI images. The improvement in quality has been pretty remarkable for the same prompt. Composition across this time has been consistent. What I initially thought was "good enough" now is... fantastic. Just so many little details got more life-like w/ each new generation. Funnily enough, our images must be 3:2 aspect ratio. I kept asking GFast to change its square Fast output to 3:2. It kept saying it would, but each image was square or nearly square. GFast in the end was very apologetic, and said it would alert about this issue. Today I read that GPro does aspect ratios. Tried the same prompt again burning up some "Thinking" credits, and got another fantastically life-like image in 3:2. We have a new project coming up. We have relied entirely on stock or in some cases custom shot images to date. Now, apart from the time needed to get the prompts right whilst meeting with the client, I cannot see how stock or custom images can compete. I mean the GPro images -- again which is very specific to an unusual prompt -- is just "Wow". Want to emphasize again -- we are looking for specific details that many would not. So the thoughts above are specific to this. Still, while many faults can be found with AI, Nano Banana is certainly proven itself to me.
edit: I was thinking about this, and am not sure I even saw Pro3 as my image option last night. Today it was clearly there.
It’s interesting, I’m trying to use it to create a themed collage by providing a few images and it does that wonderfully, but in the process it is also hallucinating the images I use so I end up with weird distorted faces. Other tools can do this without issue, but something about faces in images this model just has to modify them every time. Ask it to remove background objects and the faces get distorted as well.
Using it for non-people involved images and it’s pretty good although I haven’t done much and it isn’t doing anything 2.5-flash wasn’t already doing in the same amount of requests.
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.
Could be they are specifically training against it. There was some controversy about "studio ghibli style". Similarly how in the early days of Stable Diffusion "Greg Rutkowski style" was a very popular prompt to get a specific look. These days modern Stable Diffusion based models like SD 3 or FLUX mostly removed references to specific artists from their datasets.
You might try it again with style transfer: 1 image of style to apply to 1 target image
This is a good idea, will give it a try!
I wonder ... do you think they might not be chasing that particular metric?
Sure! But it's weird how far off it is in terms of capability.
Slightly off topic, but how are people creating long videos like 30 second videos that I often see on Instagram? It I try to use Veo to make split videos, it simply cannot maintain the style or weird quirks get into the subsequent videos. Is there anything else that's the best video generation model currently other than Veo?
Longer videos without cuts are usually made from the first/last frame feature available in Veo 3.1 and other video models like Kling 2.5
I really hope Google reads these HN posts. They've had some big "product" wins but the pricing, packaging, and user system is a severe blocker to growth. If developers can't or won't figure it out -- how the heck are consumers?
And both their consumer apps are slow. You can replicate this yourself. Go to AI Studio, paste in 80K tokens of text, then type something on your keyboard, and see what happens. The Gemini web app is even worse somehow. A horrifically slow and buggy app. Not new problems either, barely any improvement on this over more than 1 year.
First model I've seen that was consistently compositional, easily handling requests like
“Generate an image of an african elephant painted in the New England flag, doing a backflip in front of the russian federal assembly.”
OpenAI made the biggest step change towards compositionality in image generation when they started directly generating image tokens for decoders from foundation llms, and it worked very well (openais images were better in this regard than nano banana 1, but struggled with some OOD images like elephants doing backflips), but banana 2 nails this stuff in a way I haven't seen anywhere else
if video follows the same trends as images in terms of prompt adherence, that will be very valuable... and interesting
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
We know what it looks like at least https://www.reddit.com/r/nanobanana/comments/1o1tvbm/nano_ba...
Is there an "in joke" to this name that I am too old to get? Or it's just a whimsically random name?
I believe it’s an internal code name that stuck.
To expand, it comes from the stealth name it was given on LMArena I believe. The model made news while still in "stealth mode" and so Google capitalised on the PR they'd already built around that and just launched it officially with the same name.
I see, naturally this is the first I've heard of it ;)
nano banano pronano.
Nani Banani, Nanu Bananu, Nano Banano...
be fi fo famo nano
Will be interesting to see how this model performs in real-world creative tasks. https://creativearena.ai/
This is really impressive. As a former designer, I'm equally excited that people will be able to generate images like this with a prompt, and sad that there will be much less incentive for people to explore design / "photoshopping" as a craft or a career.
At the end of the day, a tool is a tool, and the computer had the same effect on the creative industry when people started using them in place of illustrating by hand, typesetting by hand, etc. I don't want my personal bias to get in the way too much, but every nail that AI hammers into the creative industry's coffin is hard to witness.
I feel you. Infact, IMO, SWE1 level coding industry seems to be a couple years lagging on this aspect.
The trouble is that learning fundamentals now is a large trough to go past, just the way grade 3-10 children learn their math fundamentals despite there being calculators. It's no longer "easy mode" in creative careers.
If Nano-Banana-pro with Veo 3.1 existed during my PhD, I would’ve finished a 6-year dissertation in a single year — it’s generating ideas today that used to take me 18 months just to convince people were possible.
The person in the background's face is odd haha
I tried the same prompt as one of the examples (https://i.imgur.com/iQTPJzz.png), in the two ways they say you can run it, via Google Gemini and Google AI Studio (I suppose they're different somehow?). The prompt was "Create an infographic that shows hot to make elaichi chai" and Google Gemini created a infographic (https://i.imgur.com/aXlRzTR.png), but it was all different from what the example showed. Google AI Studio instead created a interactive website, again with different directions: https://i.imgur.com/OjBKTkJ.png
There is not a single mention about accuracy, risks or anything else in the blogpost, just how awesome the thing is. It's clearly not meant to be reliable just yet, but not making this clear up front. Isn't this almost intentionally misleading people, something that should be illegal?
Whoever said there was a universal recipe for Elaichi Chai? It makes sense that there would be different recipes. If you are more stringent with the prompt and give it the proper context of what you want the steps to be, you'll arrive at that consistency.
If it were illegal to intentionally mislead people, many magicians would be out of a job :)
I was just playing with the non-pro version of this and it seems to add both a Gemini and Disney watermark. Presumably this was because I referenced beauty and the beast.
Anyone know if this is an hallucination or if they have some kind of deal with content owners to add branding?
Wow! I was able to combine Nano Banana Pro and Veo 3.1 video generation in a single chat and it produced great results. https://chat.vlm.run/c/38b99710-560c-4967-839b-4578a4146956. Really cool model
Neat use-case, though the sword literally telescopically inverts itself at the beginning of the scene like a light saber where you would have expected it to be drawn from its scabbard.
I'd be interested to see how Wan 2.2 First/Last frame handles those images though...
That is an interesting error actually. It happened because both orientations of the sword are visually plausible, but not abrupt transitions from one to the other; there needs to be physical continuity.
Here is a reproduction of the Matrix bullet time shot with and without pose guidance to illustrate the problem: https://youtu.be/iq5JaG53dho?t=1125
yeah sadly veo 3.1 has not caught up to the image generation capabilities. May be we need to work on how to make video generation more physically consistent. but the image generation results from banana pro are great.
another interesting use case with synth https://chat.vlm.run/c/1c726fab-04ef-47cc-923d-cb3b005d6262. made a puppet from a image of a model and made the puppet dance.
The feet are doing unusual movements. Reminds me of leaf node cumulative error in overcompressed hierarchical animation.
I see many recent accounts posting vlm.run links and if this is what I suspect it is, that's normally not allowed here.
If you have concerns about spam, the right thing to do is to email the mods at hn@ycombinator.com with examples.
Really interesting. Curious what the main design motivation behind this project was and what gaps it fills compared to existing tools?
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.
You sort of can on Android, but it's a few steps:
1. Trigger Circle to Search with long holding the home button/bar
2. Select the image
3. Navigate to About this image on the Google search top bar all the way to the right - check if it says "Made by Google AI" - which means it detected the SynthID watermark.
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).
https://genai-showdown.specr.net/image-editing
My experience with Nano Banana is to constantly get consistent image when dealing with muliple objects in a image, I mean creating consistent sequence etc.
We spent a lot of money trying but eventully gave up. If it is easier in Pro, then probably it stands a chance.
What can nano-banana do that chatGPT made images can't? Or is it only better for image editing from what I can gather from these comments so far. I haven't used it so genuinely curious.
I made some direct comparisons my Nano Banana post (https://news.ycombinator.com/item?id=45917875) but Nano Banana can handle photorealistic photos with nuanced prompts much better. And there is no yellow filter.
https://news.ycombinator.com/item?id=45890186
> Nano Banana Pro is the best model for creating images with correctly rendered and legible text directly in the image
> 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?
It’s not necessarily harder than other aspects. However:
- It requires an AI that actually understands English, I.e. an LLM. Older, diffusion-only models were naturally terrible at that, because they weren’t trained on it.
- It requires the AI to make no mistakes on image rendering, and that’s a high bar. Mistakes in image generation are so common we have memes about it, and for all that hands generally work fine now, the rest of the picture is full of mistakes you can’t tell are mistakes. Entirely impossible with text.
Nano Banana Pro seems to somewhat reliably produce entire pictures without any mistakes at all.
As a complete layman, it seems obvious that it should be hard? Like, text is a type of graphic that needs to be coherent both in its detail and its large structure, and there’s a very small amount of variation that we don’t immediately notice as strange or flat out incorrect. That’s not true of most types of imagery.
Can anyone please explain me the invisible watermarking mentioned in the said promo?
It's called Synth ID. It's a watermark that proves an image was generated by AI.
https://deepmind.google/models/synthid/
Super important for Google as a search engine so they can filter out and downrank AI generated results. However I expect there are many models out there which don’t do this, that everyone could use instead. So in the end a “feature” like this makes me less likely to use their model because I don’t know how Google will end up treating my blog post if I decide to include an AI generated or AI edited image.
It’s required by EU regulations. Any public generator that doesn’t do it, is in violation of that unless it’s entirely inaccessible from the EU…
But of course there’s no way to enforce it on local generation.
The EU didn't define any specific method of watermarking nor does it need to be tamper resistant. Even if they had specified it though, it's easy to remove watermarks like SynthID.
So whoever creates AI content needs to voluntarily adopt this so that Google can sell "technology" for identifying said content?
Not sure how that makes any sense
In theory, at least. In practice maybe not.
https://i.imgur.com/WKckRmi.png
?
Google doesn't claim that Gemini would call SynthID detector at this point.
Edit: well they actually do. I guess it is not rolled out yet.
From the OP:
> Today, we are putting a powerful verification tool directly in consumers’ hands: you can now upload an image into the Gemini app and simply ask if it was generated by Google AI, thanks to SynthID technology. We are starting with images, but will expand to audio and video soon.
Re-rolling a few times got it to mention trying SynthID, but as a false negative, assuming it actually did the check and isn't just bullshitting.
> No Digital Watermark Detected: I was unable to detect any digital watermarks (such as Google's SynthID) that would definitively label it as being generated by a specific AI tool.
This would be a lot simpler if they just exposed the detector directly, but apparently the future is coaxing an LLM into doing a tool call and then second guessing whether it actually ran the tool.
*by Google's AI.
By anybody's AI using SynthID watermarking, not just Google's AI using SynthID watermarking (it looks like partnership is not open to just anyone though, you have to apply).
Has anyone found out how to use Synth ID? If I want to if some images are AI, how can I do?
It's a funny juxtaposition to slap the "Pro" label on it which makes it sound more enterprisey but leave the name as Nano Banana.
Nano Banana has been the only model I’ve really loved. As a small businesses who makes products, it’s been a game changer on the marketing side. Now when I’ve got something new I need to advertise in a hurry, I take a crappy pic and fix it in that. Don’t have a perfect model ready yet? That’s ok, I can just alter to look exactly like it will.
What used to cost money and involve wait time is now free and instant.
Maybe I'm an obscure case, but I'm just not sure what I'd use an image generation model for.
For people that use them (regularly or not), what do you use them for?
My most regular use-case is generating silly memes in group chats. If someone posts something meme-worthy or I come up with a creative response, image generation is good for one-off throwaway memes. A recent example was an "official license to opine on sociology", following someone arguing about credentialism.
Recently I also started using image generation models to explore ideas for what changes to make in my paintings. Although generally I don't like the suggestions it makes, sometimes it provides me with creative ideas of techniques that are worth experimenting with.
One way to approach thinking about it is that it's good for exploring permutations in an idea-space.
Random examples:
1) I have a tricep tendon injury and ChatGPT wants me to check my tricep reflex. I have no idea where on the elbow you're supposed to tap to trigger the reflex.
2) I'm measuring my body fat using skin fold calipers. Show me were the measurement sites are.
3) I'm going hiking. Remind me how to identify poison ivy and dangerous snakes.
4) What would I look like with a buzz cut?
You should never rely on AI to do 1, 2 or 3, especially a sloppy model like this.
First three are interesting - all question / knowledge based where the answer is a picture. Hadn't really considered this.
The answer is a picture that almost certainly already exists.
Why would you want a program that just makes one up instead?
Mostly highly specific images in blog posts but I also use it for occasional comics.
https://mordenstar.com/portfolio/gorgonzo
https://mordenstar.com/portfolio/brawny-tortillas
https://mordenstar.com/portfolio/ms-frizzle-lava
I'm kind of reading between the lines, but sounds like "for fun" which makes sense / what I generally expected for why people use it
I think that's a fair assessment. I write a lot of bizarre fiction in my spare time, so Text2Image tools are a fun way to see my visions visualized.
Like this one:
A piano where the keyboard is wrapped in a circular interface surrounding a drummer's stool connected to a motor that spins the seat, with a foot-operated pedal to control rotation speed for endless glissandos.
Nano Banana is more of an image editing model, which probably has more broad use cases for non-generative applications: interior decorating, architecture, picking wardrobes, etc.
Definitely, but don't sleep on its generative capacities either. You can give it a image and instruct it "Use the attached image purely as a stylistic reference" and then proceed to use it as a regular generative model.
Indeed. Is Nano Banana now Google flagship image gen model (over Imagen 4)?
In my tests it does outscore Imagen3 and Imagen4 even in the generative capacity, but my benchmark is more focused around prompt adherence. I'd wager that for certain photorealistic tests Imagen4 is probably better.
https://genai-showdown.specr.net/?models=i3,i4,nb
Yeah... For some reason none of these are use cases in my day to day life. That said, I also don't open Photoshop very often. And maybe that's what this is meant to replace.
Not for everyone everyday, but a good tool to have in the toolbox. I recently was very easily able to mock up what a certain Christmas decoration would look like on the house. By next year, I'm sure that feature will be part of the product page.
I'm creating a team T-shirt from a bunch of kids drawings. The model has synthesize a bunch of disparate drawings into a cohesive concept, incorporate the team's name in the appropriate color and font, and make it simple enough for a T-shirt.
porn is probably the a biggest one?
but concept art, try-it-on for clothes or paint, stock art, etc
Nonconsenual pornography is the killer app.
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 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"
Yeah, the weird yellow tint, the kerning/fonts etc still immediately gives it away.
But I wouldn't mind being easily able to make infographics like these, I'd just like to supply the textual and factual content myself.
I would do the same. But the reason for that is because I’m terrible at drawing and digital art, so I would need some help with the graphics in an infographics anyways. I don’t really need help with writing text or typesetting the text. I feel like if I were better at creating art I would not want AI involved at all.
Oh what a day. What a lovely day.
https://www.youtube.com/watch?v=5mZ0_jor2_k
Honestly I think this is exactly how we're all feeling right now. Racing towards an unknown horizon in a nitrous powered dragster surrounded by fire tornadoes.
Not gonna lie - this is pretty cool.
But ... it comes from Google. My goal is to eventually degoogle completely. I am not going to add any more dependency - I am way too annoyed at having to use the search engine (getting constantly worse though), google chrome (long story ...) and youtube.
I'll eventually find solutions to these.
Time to expand my creation catalog. Lets see what we can get of out this pro version. It seems this week is for big AI announcements from Google
really missed an opportunity to name it micro banana (or milli banana). Personally I can't wait for mega banana next year.
I’ve been struggling with infographics. That’s my main use case but every tool seems to bungle the text.
> Starting to roll out in the Gemini API and Google AI Studio
> Rolling out globally in the Gemini app
wanna be any more vague? is it out or not? where? when?
Currently, it’s rolling out in the Gemini app. When you use the “Create image” option, you’ll see a tooltip saying “Generating image with Nano Banana Pro.”
And in AI Studio, you need to connect a paid API key to use it:
https://aistudio.google.com/prompts/new_chat?model=gemini-3-...
> Nano Banana Pro is only available for paid-tier users. Link a paid API key to access higher rate limits, advanced features, and more.
Phased rollouts are fairly common in the industry.
Already available in the Gemini web app for me. I have the normal Pro subscription.
I don't see in the ai studio
I see it but when I use it says "Failed to count tokens, model not found: models/gemini-3-pro-image-preview. Please try again with a different model."
does it handle transparency yet?
Anyone else think "Nano Banana" is an awful name? For some reason it really annoys me. It looks incredibly fancy, though.
If only there was a straightforward way to pay google to use this, with a not entirely insane UX...
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.
There's 2 takes here: First take is the AI is replacing jobs by making existing workforce more efficient.
The 2nd take is AI is costing companies so much money, that they need to cut workforce to pay for their AI investments.
I'm inclined to think the latter is represents what's happening more than the former.
On the contrary, it's encouraging to know that maliciously greedy companies like Adobe are getting screwed for being so malicious and greedy :thumbsup:
I had second thoughts about this comment, but if I stopped typing in the middle of it, I would've had to pay a cancellation fee.
Adobe, for all their faults, can hardly be said to be more malicious or greedy than Google.
Adobe, at least, makes money by selling software. Google makes money by capturing eyeballs; only incidentally does anything they do benefit the user.
Adobe makes money by renting software, not selling it. There are many creatives that would disagree with your ranking of who is more malicious or greedy.
What is up with these product names!? Antigravity? Nano Banana?
Not just are they making slop machines, they seem to be run by them.
I am too old for this shit.
Cool, but it's still unusable for me. Somehow all my prompts are violating the rules, huh?
In 25 years we'll reminisce on the times when we could find a human artist who wouldn't impose Google's or OpenAI's rules on their output.
the open-source models will catch up, 100%
Open models don't seem to be catching up the LLM-based image gen at this point.
ChatGPT's imagegen has been released for half a year but there isn't anything remotely similar to it in the open weight realm.
Give it another 50 years. Or maybe 10. Or 5? But there's no way it won't catch up.
Are you asking it to recreate people?
No, and no nudity, no reference images. Example: 'athlete wearing a health tracker under a fitted training top'
Can you give us an example?
'athlete wearing a health tracker under a fitted training top'
Failed to generate content: permission denied. Please try again.
It's not the censorship safeguard. Permission denied means you need a paid API key to use it. It's confusing, I know.
If you triggered the safeguard it'll give you the typical "sorry, I can't..." LLM response.
Have some examples?
Can Google Gemini 3 check Google Flights for live ticket prices yet?
(The Gemini 3 post has a million comments too many to ask this now)
https://gemini.google.com/share/19fed9993f06
Ah thanks, might have to make a throwaway account just for that.
Gemini 2 still goes "While I cannot check Google Flights directly, I can provide you with information based on current search results…" blah blah
Nano Banana Pro sounds like classic Google branding: quirky name, serious tech underneath. I’m curious whether the “Pro” here is about actual professional‑grade features or just marketing polish. Either way, it’s another reminder that naming can shape expectations as much as specs.