Nano-Banana can produce some astonishing results. I maintain a comparison website for state-of-the-art image models with a very high focus on adherence across a wide variety of text-to-image prompts.
I recently finished putting together an Editing Comparison Showdown counterpart where the focus is still adherence but testing the ability to make localized edits of existing images using pure text prompts. It's currently comparing 6 multimodal models including Nano-Banana, Kontext Max, Qwen 20b, etc.
Gemini Flash 2.5 leads with a score of 7 out of 12, but Kontext comes in at 5 out of 12 which is especially surprising considering you can run the Dev model of it locally.
Don't know if it's the same for others, but my issue with Nano Banana has been the opposite. Ask it to make x significant change, and it spits out what I would've sworn is the same image. Sometimes randomly and inexplicably it spits our the expected result.
Anyone else experiencing this or have solutions for avoiding this?
Just yesterday, asking it to make some design changes to my study. It did a great job with all the complex stuff, but asking it to move a shelf higher, it repeatedly gave me back the same image. With LLMs generally I find as soon as you encounter resistance it's best to start a new chat, however in this case that didn't wok either. Not a single thing I could do to convince it that the shelf didn't look right half way up a wall.
Yeah I've definitely seen this. You can actually see evidence of this problem in some of the trickier prompts (the straightened Tower of Pisa and the giraffe for example).
Most models (gpt-image-1, Kontext, etc) typically fail by doing the wrong thing.
From my testing this seems to be a Nano-Banana issue. I've found you can occasionally work around it by adding far more explicit directives to the prompt but there's no guarantee.
I've had this same issue happen repeatedly. It's not a big deal because it is just for small personal stuff, but I often need to tell it that it is doing the same thing and that I had asked for changes.
Great comparison! Bookmarked to follow. Keep an eye on Grok, they're improving at a very rapid rate and I suspect they'll be near the top in not too distant future.
Will do! I just added Seedream v4.0 a few hours ago as well. It's all I can do just to keep up and not get trampled under the relentless march of progress.
By the way, some of the results look a little weird to me, like the one for the 'Long Neck' prompt. The giraffe of Seedream just lowered its head but its neck didn't shorten as expected. I'd like to learn about the evaluation process, especially whether it is automatic or manual.
Hi Isharmla, the giraffe one was a tough call. IMHO, even when correcting for perspective, I do feel like it managed to follow the directive of the prompt and shorten the neck.
To answer your question, all of the evaluations are performed manually. On the trickier results I'll occasionally conscript some friends to get a group evaluation.
The bottom section of the site has an FAQ that gives more detail, I'll include it here:
It's hard to define a discrete rubric for grading at an inherently qualitative level. To keep things simple, this test is purely PASS/FAIL - unsuccessful means that the model NEVER managed to generate an image adhering to the prompt.
In many cases, we often attempt a generous interpretation of the prompt - if it gets close enough, we might consider it a pass.
To paraphrase former Supreme Court Justice Potter Stewart, "I may not be able to define a passing image, but I know it when I see it."
That's a mistake. Gpt-image-1 is a lot stricter in the supported output resolutions so it's using a cropped image. I'll fix the test later this week. Thanks for the heads up!
Add gpt-image-1. It's not strictly an editing model since it changes the global pixels, but I've found it to be more instructive than Nano Banana for extremely complicated prompts and image references.
It's actually already in there - the full list of edit models is Nano-Banana, Kontext Dev, Kontext Max, Qwen Edit 20b, gpt-image-1, and Omnigen2.
I agree with your assessment - even though it does tend to make changes at a global level you can least attempt to minimize its alterations through careful prompting.
> - Case 16 labels the tricuspid in the wrong place and I have no idea what a "mittic" is
> - Case 27 shows the usual "models can't do text" though I'm not holding that against it too much
16 makes it seem like it can "do text" — almost, if we don't care what it says. But it looks very crisp until you notice the "Pul??nary Artereys".
I'd say the bigger problem with 27 is that asking to add a watermark also took the scroll out of the woman's hands.
(While I'm looking, 28 has a lot of things wrong with it on closer inspection. I said 26 originally because I randomly woke up in the middle of the night for this and apparently I don't know which way I'm scrolling.)
EDIT: Yeah, on closer inspection, 28 is definitely a bit screwy. I wasn't clicking on the images themselves to view the enlarged ones, and from the preview I didn't see anything that immediately jumped out at me. I have no idea what that line at the bottom is meant to represent!
Also you're right, I didn't notice the scroll had gone, though on another inspection, it's also removed the original prompter's watermark
Yeah, I appreciate this kind of benchmarking too. That other Gen AI Showdown in the comments also does a good job with this - mentions that it was best of 8 attempts and so on.
Yep, Google actually recommends using Imagen4 / Imagen4 Ultra for straight image generation. In spite of that, Flash 2.5 still scored shockingly high on my text-to-image comparisons though image fidelity is obviously not as good as the dedicated text to image models.
Came within striking distance of OpenAI gpt-image-1 at only one point less.
This is the first time I really don't understand how people are getting good results. On https://aistudio.google.com with Nano Banana selected (gemini-2.5-flash-image-preview) I get - garbage - results. I'll upload a character reference photo and a scene and ask Gemini to place the character in the scene. What it then does is to simply cut and paste the character into the scene, even if they are completely different in style, colours, etc.
I get far better results using ChatGPT for example. Of course, the character seldom looks anything like the reference, but it looks better than what I could do in paint in two minutes.
When Nano Banana works well, it really works -- but 90% of the time the results will be weird or of poor quality, with what looks like cut-and-paste or paint-over, and it also refuses a lot of reasonable requests on "safety" grounds. (In my experience, almost anything with real people.)
I'm mostly annoyed, rather than impressed, with it.
Ok this answers my question to the nature of the page. As in: Are these examples that show results you get when using certain inputs and prompts. Or are these impressive lucky on offs.
I was a bit surprised to see quality. Last time I played around with image generation is a few months back and I’m more in the frustration camp. Not to say that I believe some people with more time and dedication at their hand can tickle better results.
From having used Nano Banana over the past few days, I think that they're extremely cherry-picked, and that each one is probably the result of multiple (probably a dozen+) attempts.
In my experience, Nano Banana would actively copy and paste if it thinks it's fine to do so. You need to explicitly prompt that the character should be seamlessly integrated into the scene or similar. In the other words, the model is superb when properly prompted especially compared to other models, but prompting itself can be annoying from time to time.
Play around with your prompt, try ask Gemini 2.5 pro to improve your prompt before sending it to Gemini 2.5 Flash, retry and learn what works and what doesn't.
I understand the results are non deterministic but I get absolute garbage too.
Uploaded pics of my (32 years old) wife and we wanted to ask it to give her a fringe/bangs to see how would she look like it either refused "because of safety" and when it complied results were horrible, it was a different person.
After many days and tries we got it to make one but there was no way to tweak the fringe, the model kept returning the same pic every time (with plenty of "content blocked" in between).
Seedream 4.0 is not always better than Gemini Flash 2.5 (nano-banana), but when it is better, there is a gulf in performance (and when it's not, it's very close.)
It's also cheaper than Gemini, and has way fewer spurious content warnings, so overall I'm done with Gemini
Through that testing, there is one prompt engineering trend that was consistent but controversial: both a) LLM-style prompt engineering with with Markdown-formated lists and b) old-school AI image style quality syntatic sugar such as award-winning and DSLR camera are both extremely effective with Gemini 2.5 Flash Image, due to its text encoder and larger training dataset which can now more accurately discriminate which specific image traits are present in an award-winning image and what traits aren't. I've tried generations both with and without those tricks and the tricks definitely have an impact. Google's developer documentation encourages the latter.
Unfortunately NSFW in parts. It might be insensitive to circulate the top URL in most US tech workplaces. For those venues, maybe you want to pick out isolated examples instead.
(Example: Half of Case 1 is an anime/manga maid-uniform woman lifting up front of skirt, and leaning back, to expose the crotch of underwear. That's the most questionable one I noticed. It's one of the first things a visitor to the top URL sees.)
As a non-US citizen - even though I've been the only Brit in remote teams of Americans - I find this really hard to make sense of.
At least in the UK, if I saw this loaded on someone else's screen at work, I might raise an eyebrow initially, but there wouldn't be any consequences that don't first consider context. As soon as the context is provided ("it's comparing AI models, look! Cool, right?!") everyone would get on with their jobs.
What would be the consequence of you viewing this at work?
How would the situation be handled?
Is the problem a HR thing - like, would people get sacked for this? Or is it like a personal conduct/temptation, that colleagues who see it might not be able to restrain themselves or something?
I'm really surprised that it can generate the underwear example. Last time I tried Nano Banaba (with safety filter 'off', whatever it means), it refused to generate a 'cursed samurai helmet on an old wooden table with a bleeding dead body underneath, in cartoon style.'
I’m Italian, and I really struggle to rationalize this attitude.
I honestly don’t understand. Maybe it’s because I’m surrounded by 2,500 years of art in which nudity is an essential and predominant element, by people (even in the workplace) who have a relaxed and genuinely democratic view of the subject — but this comment feels totally alien to me. I suppose it’s my own limitation, but I would NEVER have focused attention on this aspect.
I don’t know, maybe I’m the one who’s wrong…
I'm more bothered by the fact that this reference image is clearly a well-made piece of digital art by some artist.
We all know the questionable nature of AI/LLM models, but people in the field usually at least try to avoid directly using other people's copyrighted material in documentation.
I'm not even talking about legality here. It just feels morally wrong to so blatantly use someone else's artwork like this.
I agree that proper permission should be used for these examples, but I’m quite sure the image in question is AI generated. The quality is incredible these days as to what can be generated, and even to a trained eye it’s getting more difficult by the day to tell if its AI or not.
My favorite (or should I say, anti-favorite?) is calling real artists' art AI, which I'm starting to see more and more of, and I've already seen a couple of artists rage-quit social media because of the anti-AI crowd's abuse.
Yeah that's bad too, but what the parent comment did was the opposite: calling an AI-generated image "clearly a well-made piece of digital art by some artist."
It boils down to the same thing - it's getting harder to distinguish AI generated art from non-AI art, and since the models are constantly getting better it's only going to get worse.
Personally, I'm underwhelmed by this model. I feel like these examples are cherry-picked. Here are some fails I've had:
- Given a face shot in direct sunlight with severe shadows, it would not remove the shadows
- Given an old black and white photo, it would not render the image in vibrant color as if taken with a modern DSLR camera. It will colorize the photo, but only with washed out, tinted colors
- When trying to reproduce the 3 x 3 grid of hair styles, it repeatedly created a 2x3 grid. Finally, it made a 3x3 grid, but one of the nine models was black instead of caucasian.
- It is unable to integrate real images into fabricated imagery. For example, when given an image of a tutu and asked to create an image of a dolphin flying over clouds wearing the tutu, the result looks like a crude photoshop snip and copy/paste job.
I thought the the 3rd example of the AR building highlighting was cool. I used the same prompt and seems to work when you ask it for the most prominent building in a skyline, but fails really hard if you ask it for another building.
I uploaded an image I found of Midtown Manhattan and tried various times to get it to highlight the Chrysler Building, it claimed it wasn't in the image (it was). I asked it to do 432 Park Ave, and it literally inserted a random building in the middle of the image that was not 432 Park, and gave me some garbled text for the description. I then tried Chicago as pictured from museum campus and asked it to highlight 2 Prudential, and it inserted the Hancock Center, which was not visible in the image I uploaded, and while the text was not garbled, was incorrect.
The "Photos of Yourself in Different Eras" one said "Don't change the character's face" but the face was totally changed. "Case 21: OOTD Outfit" used the wrong camera. "Virtual Makeup Try-On" messed up the make up. "Lighting Control" messed up the lighting, the joker minifig is literally just SH0133 (https://www.bricklink.com/catalogItemInv.asp?M=sh0133), "Design a Chess Set" says you don't need an input image, but the prompt said to base it off of a picture that wasn't included and the output is pretty questionable (WTF is with those pawns!), etc.
I mean, it's still pretty neat, and could be useful for people without access to photoshop or to get someone started on a project to finish up by hand.
This is amazing. Not that long ago, even getting a model to reliably output the same character multiple times was a real challenge. Now we’re seeing this level of composition and consistency. The pace of progress in generative models is wild.
Huge thanks to the author (and the many contributors) as well for gathering so many examples; it’s incredibly useful to see them to better understand the possibilities of the tool.
I've come to realize that I liked believing that there was something special about the human mental ability to use our mind's eye and visual imagination to picture something, such as how we would look with a different hairstyle. It's uncomfortable seeing that skill reproduced by machinery at the same level as my own imagination, or even better. It makes me feel like my ability to use my imagination is no more remarkable than my ability to hold a coat off the ground like a coat hook would.
As someone who can’t visualize things like this in my head, and can only think about them intellectually, your own imagination is still special. When I heard people can do that, it sounded like a super power.
AI is like Batman, useless without his money and utility belt. Your own abilities are more like Superman, part of who you are and always with you, ready for use.
But you can find joy at things you envision, or laugh, or be horrified. The mental ability is surely impressive, but having a reason to do it and feeling something at the result is special.
"To see a world in a grain of sand
And a heaven in a wild flower..."
We - humans - have reasons to be. We get to look at a sunset and think about the scattering of light and different frequencies and how it causes the different colors. But we can also just enjoy the beauty of it.
For me, every moment is magical when I take the time to let it be so. Heck, for there to even be a me responding to a you and all of the things that had to happen for Hacker News to be here. It's pretty incredible. To me anyway.
I always thought I had a vivid imagination. But then the aphantasia was mentioned in Hello Internet once, I looked it up, see comments like these and honestly…
I’ve no idea how to even check. According to various tests I believe I have aphantasia. But mostly I’ve got not even a slightest idea on how not having it is supposed to work. I guess this is one of those mysteries when a missing sense cannot be described in any manner.
A simple test for aphantasia that I gave my kids when they asked about it is to picture an apple with three blue dots on it. Once you have it, describe where the dots are on the apple.
Without aphantasia, it should be easy to "see" where the dots are since your mind has placed them on the apple somewhere already. Maybe they're in a line, or arranged in a triangle, across the middle or at the top.
When reading "picture an apple with three blue dots on it", I have an abstract concept of an apple and three dots. There's really no geometry there, without follow on questions, or some priming in the question.
In my conscious experience I pretty much imagine {apple, dot, dot, dot}. I don't "see" blue, the dots are tagged with dot.color == blue.
When you ask about the arrangement of the dots, I'll THEN think about it, and then says "arranged in a triangle." But that's because you've probed with your question. Before you probed, there's no concept in my mind of any geometric arrangement.
If I hadn't been prompted to think / naturally thought about the color of the apple, and you asked me "what color is the apple." Only then would I say "green" or "red."
If you asked me to describe my office (for example) my brain can't really imagine it "holistically." I can think of the desk and then enumerate it's properties: white legs, wooden top, rug on ground. But, essentially, I'm running a geometric iterator over the scene, starting from some anchor object, jumping to nearby objects, and then enumerating their properties.
I have glimpses of what it's like to "see" in my minds eye. At night, in bed, just before sleep, if I concentrate really hard, I can sometimes see fleeting images. I liken it to looking at one of those eye puzzles where you have to relax your eyes to "see it." I almost have to focus on "seeing" without looking into the blackness of my closed eyes.
Exactly my experience too. These fleeting images are rare, but bloody hell it feels like cheating at life if most people can summon up visualisations like that at will.
Watching someone clearly just transfer what's in their mind to a drawing is just jaw-dropping to me.
Like they'll start at an arm and move along filling the rest of the body correctly the first time. No sketching, no finding the lines, just a human printer.
I can see I my head with ~80% the level as seeing with my eyes. It's a little tunnel visiony and fine details can be blurry, but I can definitely see it. A honeycrisp apple on a red woven placemat on a wooden counter top. The blue dots are the size of peas, they are stickers in a triangle.
It not just images either, it's short videos.
What's interesting though is that the "video" can be missing details that I will "hallucinate" back in that will be incorrect. So I cannot always fully trust these. Like cutting the apple in half lead to a ~1/8th slice missing from one of the halves. It's weird.
I'm a 5 on the VVIQ. I can see the 3D apple, put it in my hand, rotate it, watch the light glint on the dimples in the skin, imagine tossing it to a close friend and watch them catch it, etc.
It's equally astonishing to me that others are different.
I absolutely do. For example, when I'm playing D&D, or listening to a podcast of other people playing D&D, I can "see" a fully realistic view of what is happening in my head. With the apple test, I can see a nice red apple, with the little vertical orange streaks, three blue dots arranged in a triangle, and I can rotate the apple in my head and have the dots move as you would expect from a real apple. I have a very vivid imagination
There are people who actually "see" a full-ass movie in their head when they read.
These are also the people who get REALLY angry when some live-action casting choice isn't exactly like in the book. I just go "meh", because I kinda remember the main character had red hair and a scar and that's it. :D
After reading your first sentence, I immediately saw an apple with three dots in a triangle pointing downwards on the side. Interestingly, the 3 dots in my image were flat, as if merely superimposed on an image of an apple, rather than actually being on an apple.
How do people with aphantasia answer the question?
I guess it's a spectrum with varying abilities. If you ask me, I can see a red apple - or a photo of a red apple precisely. It's not in 3D though, I cannot imagine it from other angles so I cannot image the dots around it. But if I were to sit in a quiet and dark room without any distractions, and tried concentrating super hard (with my eyes closed), then I would be able to see it as other can. Perhaps even manipulate it in my mind.
Then maybe, at least in my case, it is my inability to focus my imagination when my senses are already being bombarded with external stimuli. But I cannot speak for anyone else.
For me the hard question to answer is whether I have aphantasia because people describing “actually seeing” things like with their eyes is an absolutely wild concept.
To answer the question I imagine an apple with three dots in a triangle, closely together. There is no color because there is no real image, it’s just an idea. As other have said if prompted the idea gets more detailed.
That said, when I tried to learn building mind palaces it has worked. I can “walk through” places I know just fine, even recall visual details like holes in a letterbox. But again, there is no image.
I found out recently that I have aphantasia, based on everything I've read. When you tell me to visualize, I imagine. I don't see it. An apple, I can imagine that. I can describe it in incomprehensibly sparse details. But when you ask details I have to fill them in.
I hadn't really placed those three dots in a specific place on the apple. But when you ask where they are, I'll decide to put them in a line on the apple. If you ask what color they are, I'll have to decide.
I'm pretty sure I don't have aphantasia. I don't see the apple either; it doesn't occupy any portion of my visual field and it doesn't feel similar to looking at an image of an apple. There's more of a ghostly, dreamlike image of an apple "somewhere else" whose details I only perceive when I think about them, and fade when I pay less attention. But the sensation of this apparition is a visual one; the apple will have an orientation, size, shape, and colour in the mental image, which are defined even if they're ghostly, inconsistent, and change as I reconsider what the apple should look like.
How fair is it to ask people to self report whether details existed in their original image before or after a second question? Does the second question not immediately refine the imagined image? Or is that the point, that there’s now a memory of two different apple states?
Edit: This iDevice really wants to capitalise Apple.
There's no apple, much less any dots. Of course, I'm happy to draw you an apple on a piece of paper, and draw some dots on that, then tell you where those are.
oh just close your eyes and imagine an apple for a few moments, then open your eyes, look at the wikipedia article about aphantasia and pick the one that best fits the level of detail you imagined.
So my mind briefly jumps to an apple and I guess I am very briefly seeing that the dots happen to be on top of the apple, but that image is fleeting.
I have had some people claim to me that they can literally see what they are imagining as if it is in front of them for prolonged periods of time, in a similar way to how it would show up via AR goggles.
I guess this is a spectrum and it's tough to dealineate the abilities. But I just looked it up and what I am describing is hyperphantasia.
For me the triggering event was reading about aphantasia, and then thinking about how I have never, ever, seen a movie about a book I've read and said, "that [actor|place|thing] looks nothing like I imagined it" Then I tried the apple thing to confirm. I have some sense of looking at things, but not much.
It's a great aspect to evaluate (fiction books/movies), thanks for mentioning it. I think it's much easier to use as an evaluation tool than techniques like the apple example. One of the tests, for example, is to recall a book that you have never seen a movie adaptation of and try to remember the characters and scenes. For me, in these cases (when I try to recall), the characters appear faceless, while places are more detailed, but they usually remind me of some real places I have encountered before in my life.
It's interesting that if non-aphantasia people are so common, I wonder why so few paintings have scenery based solely on imagination. I even remember asking a person who paints (not in the context of this condition) how hard it was for him to paint something not directly before his eyes, but from imagination, and why he didn't do it more often. I recall that he definitely did this (painting from imagination) rarely or not at all, and the question really puzzled him
Ask people to visualize a thing. Pick something like a house, dog, tree, etc. Then ask about details. Where is the dog?
I have aphantasia and my dog isn't anywhere. It's just a dog, you didn't ask me to visualize anything else.
When you ask about details, like color, tail length, eyes then I have to make them up on the spot. I can do that very quickly but I don't "see" the good boy.
The proof in the pudding will be if machines will be able to develop new art styles. For example, there is a progression in comic/manga/anime art styles over the decades. If humans would stop (they probably won't) that kind of progression, would machines be able to continue it? In principle yes (we are biological machines of sorts), but likely not with the current AI architecture.
I think it's a mistake to look at developing new art styles as simply continuing a linear progression. More often than not art styles are unique to the artist - you couldn't, for instance, put Eichiro Oda, Tsutomu Nihei and Rumiko Takahashi on the same number line. And trends tend to develop in reaction to existing trends, usually started by a single artist, as often as they do as an evolution of a norm.
Arguably, if creating an art style is simply a matter of novel mechanics and uniqueness, LLMs could already do that simply by adding artists to the prompts ("X" in the style of "A" and "B") and plenty of people did (and do) argue that this is no different than what human artists do (I would disagree.) I personally want to argue that intentionally matters more than raw technique, but Hacker News would require a strict proof for the definition of intentionality that they would argue humans don't possess, but somehow LLMs do, and that of course I can't provide.
I guess I have no argument besides "it means more to me that a person does it than a machine." It matters to me that a human artist cares. A machine doesn't care. And yes, in a strictly materialist sense we are nothing but black boxes of neurons receiving stimuli and there is no fundamental difference between a green field and a cold steel rail, it's all just math and meat, but I still don't care if a machine makes X in the style of (Jack Kirby AND Frank Miller.)
> More often than not art styles are unique to the artist
I'd disagree. Art styles are a category of many similar works in relation to others or a way of bringing about similar works. They usually build off of or are influenced by prior work and previous methods, even in cases where there is a effort to avoid or subvert them. Even with novel techniques or new mediums. "Great Artists Steal" and all that.
Some people become known for certain mediums or the inclusion of specific elements, but few of them were the first or only artists to use them. "Art in the style of X" just comes down to familiarity/marketing. Art develops the way food does with fads, standards, cycles, and with technology and circumstance enabling new things. I think evolution is a pretty good analogy although it's driven by a certain amount of creativity, personal preference, and intent in addition to randomness and natural selection.
Computers could output random noise and in the process eventually end up creating an art style, but it'd take a human to recognize anything valuable and artists to incorporate it into other works. Right now what passes for AI is just remixing existing art created by humans which makes it more likely to blindly stumble into creating some output we like, but inspiration can come from anywhere. I wouldn't be surprised if the "AI Slop" art style wasn't already inspiring human artists. Maybe there are already painters out there doing portraits of people with the wrong number of fingers. As AI is increasingly consuming it's own slop things could get weird enough to inspire new styles, or alternately homogenized into nothing but blandness.
To be fair, we're the beneficiaries of nature generating the data we trained on ourselves. Our ability came from being exposed to training in school, and in the world, and from examples from all of human history. Ie. if you locked a child in a dark room for their entire lives, and gave them no education or social interaction, they wouldn't have a very impressive imagination or artistic ability either.
> Nano banana saves literally millions of manual human pixel pushing hours.
At the low, low cost of burning incredible amounts of energy!
This is also he same logic as “lost sale” software piracy calculations. 90% of those claimed hours would not have been spent if the tool did not exist. Most of the generated images are idle throwaways that no human would bother with creating.
Vision has evolved frequently and quickly in the animal kingdom.
Conscious intelligence has not.
As another argument, we've had mathematical descriptions of optics, drawing algorithms, fixed function pipeline, ray tracing, and so much more rich math for drawing and animating.
Smart, thinking machines? We haven't the faintest idea.
> Vision has evolved frequently and quickly in the animal kingdom. Conscious intelligence has not.
Three times, something like intelligence has evolved - in mammals, octopuses, and corvids. Completely different neural architectures in those unrelated speces.
Seriously? One could always cut-and-paste (not the computer term) a hairstyle over a photo of a person.
You are now marvelling at someone taking the collective output of humans around the world, then training a model on it with massive, massive compute… and then having a single human compete with that model.
Without the human output on the Internet, none of this would be possible. ImageNet was positively small compared to this.
But yeah, what you call “imagination” is basically perturbations and exploration across a model that you have in your head, which imposes constraints (eg gravity etc) that you learned. Obviously we can remix things now that they’re on the Internet.
Having said that, after all that compute, the models had trouble rendering clocks that show an arbitrary time, or a glass of wine filled to the brim.
>Having said that, after all that compute, the models had trouble rendering clocks that show an arbitrary time, or a glass of wine filled to the brim.
I know you're probably talking about analog clocks, but people when dreaming have trouble representing stable digits on clocks. It's one of the methods to tell if you are dreaming.
Does a pretty good job (most of the time) of sticking to the black and white coloring book style while still bringing in enough detail to recognize the original photo in the output.
Impressive examples but for GenAI it always comes down to the fact that you have to cherry pick the best result after so many fail attempts. Right now, it feels like they're pushing the narrative that ExpectedOutput = LLM(Prompt, Input) when it's actually ExpectedOutput = LLM(Prompt, Input) * Takes where Takes can vary from 1 to 100 or more
ML researchers have been used Top-5 accuracy for a quite long time, especially when it comes to computer vision.
Of course it's a ridiculous index in most use cases (like in self-driving car. Your 4th guess is that you need to brake? Cool...). But somehow people in ML normalized it.
That's why I always record the number of rolls it takes to get to an acceptable result on my GenAI Comparison site for each model - it's a broad metric indicating how much you have to fight to steer the model in the right direction.
Man, I hate this. It all looks so good, and it's all so incorrect. Take the heart diagram, for example. Lots of words that sort of sound cardiac but aren't ("ventricar," "mittic"), and some labels that ARE cardiac, but are in the wrong place. The scenes generated from topo maps look convincing, but they don't actually follow the topography correctly. I'm not looking forward to when search and rescue people start using this and plan routes that go off cliffs. Most people I know are too gullible to understand that this is a bullshit generator. This stuff is lethal and I'm very worried it will accelerate the rate at which the populace is getting stupider.
> Most people I know are too gullible to understand that this is a bullshit generator.
I'm more worried about the cases that aren't trying to be info diagrams. There's all this "safety" discourse around not letting people generate NSFW, and around image copyrights etc. but nobody talks about the potential to use things like #11 for fraud. "Disinformation" always gets approached from a political angle instead of one of personal gain.
One thing that couldn't be done is transparent background. The model just generates the pattern in the background. Not real alpha channel transparency. You can even see artifacts in the pattern.
The training data is presumably full of examples of people using the pattern to indicate transparency (and explaining that they do so — like the input for 50!), and much less of people actually creating such images (if the training data even preserves the alpha channel in the first place).
I think a bigger problem is the "artifacts" you describe (worse than that sounds to me).
Yeah, mangled checkerboard patterns are common when prompted to "remove" the background. It can be worked around by generating multiple images with only the background color varying (e.g. black and white) and reconstructing the alpha channel from their difference, as the model generally prefers to just copy and paste when no other prompts override that preference.
“Just do more manual work and waste even more energy so you can take yet another manual step and finally get what you wanted.” A real time-saver, that.
So it seems like image generation/deepfake proliferation is pretty inevitable. I imagine we can't trust any image anymore (for e.g. identification verification purposes) unless it is done in person or otherwise notarized somehow. Is there a way (NFT-ish?) to "tag"/sign an image to say it was taken by an actual camera?
I theory you could install some kind of TPM-like device to every hardware that signs the data with key generated by manufacturer. Should be designed in such a way that it is very easy to break it when trying to tamper with it
I'm furnishing a new apartment and Nano Banana has been super useful for placing furniture I want to purchase in rooms to make a judgment if things will work for us or not. Take a picture of the room, feed Nano Banana with that picture and the product picture and ask it to place it in the right location. It can even imagine things at night or even add lamps with lights on. Super useful!
I think it might be the same as with programmers. It might look like AI Agents can do all the programming, but when you actually try to use it do do things it quickly turns out to be not so much reliable.
Maybe they're better off switching careers? At some point, your customers aren't going to pay you very much to do something that they've become able to do themselves.
There used to be a job people would do, where they'd go around in the morning and wake people up so they could get to work on time. They were called a "knocker-up". When the alarm clock was invented, these people lose their jobs to other knockers-up with alarm clocks, they lost their jobs to alarm clocks.
A lot of technological progress is about moving in the other direction: taking things you can do yourself and having others do it instead.
You can paint your own walls or fix your own plumbing, but people pay others instead. You can cook your food, but you order take-out. It's not hard to sew your own clothes, but...
So no, I don't think it's as simple as that. A lot of people will not want the mental burden of learning a new tool and will have no problem paying someone else to do it. The main thing is that the price structure will change. You won't be able to charge $1,000 for a project that takes you a couple of days. Instead, you will need to charge $20 for stuff you can crank out in 20 minutes with gen AI.
I agree with this. And it's not just about saving time/effort--an artist with an AI tool will always create better images than an amateur, just as an artist with a camera will always produce a better picture than me.
That said, I'm pretty sure the market for professional photographers shrank after the digital camera revolution.
Does anyone else cringe when they see so many examples of sexualised young women? Literally, Case 1/B has a women lifting up her skirt to reveal her underwear. For an otherwise very impressive model, you are spoiling the PR with this kind of immature content. Sheesh. I guess that confirms it: I am a old grumpy man! I count 26 examples with young women, and 9 examples with men. The only thing missing was "Lena": https://en.wikipedia.org/wiki/Lenna
I had to scroll down way too long for someone to point this out. Its messed up how casually racialised all these image gen examples are towards young asian women.
wait until you learn what prehistoric sculptors spent their time carving
I read your comment before checking the site and then I saw case one was a child followed by a sexy maid and I thought "oh no dear god" before I realized they weren't combining them into a single image.
the archeological evidence is rather consistent and clear. I'm aware of critiques trying to change the interpretation of what the female figures are for, but nobody denies that they are naked female figures. And the critiques don't seem to have found much purchase among archeologists.
> the archeological evidence is rather consistent and clear.
What are you referring to?
> but nobody denies that they are naked female figures.
No, but the suggestion above that they were the prehistoric equivalent to cartoons of school girls lifting their skirts hasn't been the dominant theory about thirty years.
> And the critiques don't seem to have found much purchase among archeologists.
This is simply incorrect. They became part of the general archeological discourse as far back as the 1900s and are now a normal part of any such discussion. Multiple theories now coexist and to frame those critical of the original Venus ideas as being somehow more fringe than the fertility/pornography theories is just misleading.
I didn't try it but I've seen really good results, is some innovation going on under the hood that we don't know? Is the technology the same of similar models? I can't find technical info on the internet
While I think most of the examples are incredible...
...the technical graphics (especially text) is generally wrong. Case 16 is an annotated heart and the anatomy is nonsensical. Case 28 with the tallest buildings has the decent images, but has the wrong names, locations, and years.
Yes, it's Gemini Flash model, meaning it's fast and relatively small and cheap, optimized for performance rather than quality. I would not expect mind-blowing capabilities in fine details from this class of models, but still, even in this regard this model sometimes just surprisingly good.
After looking at Cases 4, 9, 23, 33, and 61, I think it might be suited to take in several wide-angle pictures or photospheres or such from inside a residence, and output a corresponding floor plan schematic.
If anyone has examples, guides, or anything to save me from pouring unnecessary funds into those API credits just to figure out how to feed it for this kind of task, I'd really appreciate sharing.
I can't provide a definitive answer for this - but I will say that the Google's SDK docs state that a single edit request is limited to a maximum of THREE images so depending on how many you have - you might have to sort of use the "Kontext Kludge", aka stitching together many of input images into a single JPEG.
I'm pretty sure these are cherry-picked out of many generation attempts, I tried a few basic things and it flat out refused to do many of them like turning a cartoon illustration into a real-world photographic portrait, it kept wanting to create a pixar style image, then when I used an ai generated portrait as an example, it refused with an error saying it wouldn't modify real world people...
I then tried to generate some multi-angle product shots from a single photo of an object, and it just refused to do the whole left, right, front, back thing, and kept doing things like a left, a front, another left, and weird half back/half side view combination.
I use the API directly but unless I'm having a "Berenstein Bears moment" I could have sworn those safety settings existed under the Advanced Options in AI Studio a few weeks ago.
Computer graphics playing in my head and I like it! I don't support Technicolor parfaits and those snobby little petit fours that sit there uneaten, and my position on that is common knowledge to everyone in Oceania.
The #1 most frustrating part of image models to me has always been their inability to keep the relevant details. Ask to change a hairstyle and you'd get a subtly different person
Seedream 4 is yet another phenomenal Chinese model. On my comparison site it's currently ranked #1 scoring 9 out of 12. Nano-Banana trails at 7 out of 12.
This is a Google's Gemini flash 2.5 model with native image output capability. It's fast, relatively cheap and SOTA-quality, and available via API.
I think getting this kind of quality in open source models will need some time, probably first from Chinese models and then from BlackForestLabs or Google's open source (Gemma) team.
Outside of Google Deepmind open sourcing the code and weights of AlphaFold, I don't think they've released any of their GenAI stuff (Imagen, Gemini, Flash 2.5, etc).
The best multimodal models that you can run locally right now are probably Qwen-Edit 20b, and Kontext.Dev.
so many little details off when the instructions are clear and/or the details are there. Brad Pitt jeans? The result are not the same style and missing clear details which should be expected to just translate over.
Another one where the prompt ended with output in a 16:9 ratio. The image isn't in that ratio.
The results are visually something but then still need so much review. Can't trust the model. Can't trust people lazily using it. Someone mentioned something about 'net negative'.
Yes, almost all of the examples are off in one way or another. The viewpoints don't actually match the arrow directions, for example. And if you actually use the model, you will see that even these examples must be cherry-picked.
The way you formulated your message just made me realize that we got somehow duped into accepting the term "model" (as in "scientific model") as a valid word for this AI stuff. A scientific model has a theoretical foundation and specific configuration parameters.
The way current AI is set up, you can't even reliably adjust the position of the sun.
You need to wait until someone does the exact picture you want, annotates it in their android photo library, and google uses it to train their AI models. But then they will be able to provide you the perfect result for your query, totally done with AI! ;)
It is HIGHLY unlikely that site is using actual Gemini Flash 2.5 - ICANN shows the domain was registered on Aug 14th 2025, practically a week before Google even announced the availability of Nano Banana.
While these are incredibly good, it's sad to think about the unfathomable amount of abuse, spam, disinformation, manipulation and who know what other negatives these advancement are gonna cause. It was one thing when you could spot an AI image, but now and moving forward it's be basically increasingly futile to even try.
Almost all "human" interaction online will be subject to doubt soon enough.
Hard to be cheerful when technology will be a net negative overall even if it benefits some.
By your logic email is clearly a net negative, given how much junk it generates - spam, phishing, hate mails, etc. Most of my emails at this point are spams.
If we're talking objectively, yeah by definition if it's a net negative, it's a net negative. But we can both agree in absolute terms the negatives of email are manageable.
Hopefully you understand the sentiment of my original message, without getting into the semantics. AI advancement, like email when it arrived, are gonna turbocharge the negatives. Difference is in the magnitude of the problem. We're dealing with whole different scale we have never seen before.
Re: Most of my emails at this point are spams. - 99% of my emails are not spam. Yet AI spam is everywhere else I look online.
Their argument is false equivalence. You can’t just say “if you’re saying X is negative, you must believe that Y is negative because some of the negatives could be conceptually similar.” A good faith cost benefit analysis would rank both the cost and risks of an extremely accurate, cheap, on-demand commercial image generation service and an entirely open asynchronous worldwide text communication protocol, in different universes.
I recently finished putting together an Editing Comparison Showdown counterpart where the focus is still adherence but testing the ability to make localized edits of existing images using pure text prompts. It's currently comparing 6 multimodal models including Nano-Banana, Kontext Max, Qwen 20b, etc.
https://genai-showdown.specr.net/image-editing
Gemini Flash 2.5 leads with a score of 7 out of 12, but Kontext comes in at 5 out of 12 which is especially surprising considering you can run the Dev model of it locally.
reply