I really really want to see how these images are starting to form into videos. The stills are clearly getting better and better, but what about when you need the stills to organically conform to a keyed script?
Did gemini-2.5-flash-image get an upgrade as well? I just got the following, which is fascinating, and not something I've seen before:
> I'm sorry, but I cannot fulfill your request as it contains conflicting instructions. You asked me to include the self-carved markings on the character's right wrist and to show him clutching his electromancy focus, but you also explicitly stated, "Do NOT include any props, weapons, or objects in the character's hands - hands should be empty." This contradiction prevents me from generating the image as requested.
My prompts are automated (e.g. I'm not writing them) and definitely have contained conflicting instructions in the past.
A quick google search on that error doesn't reveal anything either
Interesting they get to rev this with the release of a new flash model. I'm speculating part of the distil pipeline includes the image gen stuff; that seems like internal tooling that will pay dividends over time, if true. New frontier model -> automatic new image model. Even if it's just incremental updates, it's good for both the product cadence and compounding improvements.
Google updated it early in AI Studio so I've been experimenting:
- Base pricing for a 1024x1024 image is almost 1.6x what normal Nano Banana is ($0.067 vs. $0.039), however you can now get a 512x512 image for cheaper, or a 4k image for cheaper than four 1k images: https://ai.google.dev/gemini-api/docs/pricing#gemini-3.1-fla...
- Thinking is now configurable between `Minimal` and `High` (was not the case with Nano Banana Pro)
- Safety of the model appears to be increased so typical copyright infringing/NSFW content is difficult to generate (it refused to let me generate cartoon characters having taken psychedelics)
- Generation speed is really slow (2-3min per image) but that may be due to load.
- Prompt adherence to my trickier prompts for Nano Banana Pro (https://minimaxir.com/2025/12/nano-banana-pro/) is much worse, unsurprisingly. For example I asked it to make a 5x2 grid with 10 given inputs and it keeps making 4x3 grids with duplicate inputs.
However, I am skeptical with their marquee feature: image search. Anyone who has used Nano Banana Pro for awhile knows that it will strongly overfit on any input images by copy/pasting the subject without changes which is bad for creativity, and I suspect this implementation appears the same.
Additionally I have a test prompt which exploits the January 2025 knowledge cutoff:
Generate a photo of the KPop Demon Hunters performing a concert at Golden Gate Park in their concert outfits.
That still fails even with Grounding with Google Search and Image Search enabled, and more charitable variants of the prompt.
tl;dr the example images (https://deepmind.google/models/gemini-image/flash/) seem similar to Nano Banana Pro which is indeed a big quality improvement but even relative to base Nano Banana it's unclear if it justifies a "2" subtitle especially given the increased cost.
I think this tech is cool, from an engineering perspective. I’m trying to figure out if there’s any justification for using it in a business world outside of: “We don’t want to pay an artist.”
You can argue things like code generation are an extension of the engineer wielding it. Image generation just seems like a net negative overall if it’s used at scale.
Edit: By scale, I mean large corporations putting content in front of millions. I understand the appeal for smaller businesses where they probably weren’t going to pay an artist anyway.
These image gen models are getting so advanced and life like that increasingly the general public are being duped into believing AI images are actually real (ex Facebook food images or fake OF models). Don't get me wrong I will enjoy the benefits of using this model for expressing myself better than ever before, but can't help feeling there's something also very insidious about these models too.
What a great thing this didn't exist in the past. We likely wouldn't have had any of the amazing artworks that we have now. Imagine an AI generated Mona Lisa, Nightwatch or Sistine Chapel ceiling because prompting would have been so much cheaper than paying Leonardo, Rembrandt or Michelangelo...
Now extrapolate to all other artforms. Sculpture seems safe, for now, but only barely so.
I only needed help of this banana boy twice, it managed to disappoint me each time. The most recent one, I was trying different beard and mustache styles on myself, on a photo I imported from my own Google photo gallery, and it consistently rejected me, claiming I'm a public figure. Nobody ever told me that I look like any famous person, so that's googles own bananination. ChatGPT nicely handled the job.
I'm building my personal home right now. The AI image models have been a game-changer in designing the look of the house. My architect did an OK job, but the details that Nano Banana added really bring the house up a notch. I just do hundreds of renders from the basic 3D models and I find looks that I like and iterate from there. We are implementing the renders from Nano Banana over our Interior Designers designs. We would not have hired the Interior Designers again after using Nano Banana to do our interiors.
I think part of the issue with architects and designers today is that they use CAD too much. It's easy to design boxes and basic roof lines in CAD. It's harder to put in curves and more craftsman features. Nano Banana's renders have more organic design features IMO.
Our house is looking great and we're very happy how it's going so far with a lot of the thanks to Nano Banana.
Any chance you'd be willing to share an album? I've considered doing this for my own home and I'd be psyched to study practical examples. Honestly this would make one helluva blog post (imo).
Out of curiosity: what is your input to the model? A CAD file or a drawing?
I find it does a good job at isometric views from floor plans. However, I needed Gemini 3.1 Pro to be able to have a chance at rendering 3D human point of view images from floor plans.
Is any of this intrinsically a strength of Nano Banana, while not of other models/generative tools? Have you tried doing the same with say Klein, ZIT, etc.?
The cost of doing more complex designs is analogous to the cost of doing more complex builds.
If you can afford the extra cost for someone to figure out how to build the blue sky designs that nano banana spits out, maybe you can afford something more thoughtful and interesting than a shitty mashup of other peoples mcmansions.
So far it does feel more iterative than an entirely new leap in terms of capabilities, but I haven't run it through the more multimodal aspects such as editing existing images.
That being said, it actually managed the King Louie jump rope test which surprised me.
I saw an item for sale on Ali Express's video and I thought "Wow, they hired some really attractive actors to pitch their little gadget." 30 seconds in, I realized they used GenAI. Not because it looked AI, but because the production values looked too high and professional for the item. I would get in on this if you sell anything online.
My naive question, can image generation make something novel eg. "show me a DNA structure that cures cancer" can it do that, or it has to have seen something before to generate it.
Just think we conceptually know what a brushless motor design looks like and it's just pixels. I guess even if it did produce the image we wouldn't know what it means.
What they've chosen as examples to illustrate the strength of the new model surprises me.
The "cubism" example seems like it would be a closer fit to something like stained glass or something. I don't think the thing really understands what cubism was all about. Cubist painters were trying to free themselves from the confines of a single integral plane of perspective by allowing themselves to show various parts of the image from different viewpoints, different times, different styles, etc.
The division of the image into geometric shapes is just a by-product of that quest, whereas the examples here have made it the sum total of the whole piece.
This feels to me like an example of how LLMs still don't "understand" what the art means, and are just aping its facade.
I'm sure this has been written about but here's what happens long term - images are commoditized and lose their emotional appeal.
Probably about half of us here remember photos before the cell phone era. They were rare, and special, and you'd have a few photos per YEAR to look back on. The feel of photos back then, was at least 100x stronger than now. They were a special item, could be given as a gift. But once they became freely available that same amount of emotion is now split across many thousands of photos. (not saying this is good or bad, just increased supply reducing value of each item)
With image/art generation the same thing will happen and I can already feel it happening. Things that used to be beautiful or fantastic looking now just feel flat and AI-ish. If claymation scenes can be generated in 1s, and I see a million claymation diagrams a year, then claymation will lose its charm. If I see a million fake Tom Cruise videos, then it oversaturates my desire for desire for all Tom Cruise movies.
> They were rare
> and special
> and you'd have a few photos per YEAR to look back on.
None of these things are true for me as a millennial in the 35-45 age group. And my family was poor to boot, and we were still drowning in photos and photo albums.
There's already a book about this: The Work of Art in the Age of Mechanical Reproduction (1935)[0]
> The conditions for an analogous insight are more favorable in the present. And if changes in the medium of contemporary perception can be comprehended as decay of the aura, it is possible to show its social causes.
100 comments
[ 4.6 ms ] story [ 85.7 ms ] thread> I'm sorry, but I cannot fulfill your request as it contains conflicting instructions. You asked me to include the self-carved markings on the character's right wrist and to show him clutching his electromancy focus, but you also explicitly stated, "Do NOT include any props, weapons, or objects in the character's hands - hands should be empty." This contradiction prevents me from generating the image as requested.
My prompts are automated (e.g. I'm not writing them) and definitely have contained conflicting instructions in the past.
A quick google search on that error doesn't reveal anything either
- Base pricing for a 1024x1024 image is almost 1.6x what normal Nano Banana is ($0.067 vs. $0.039), however you can now get a 512x512 image for cheaper, or a 4k image for cheaper than four 1k images: https://ai.google.dev/gemini-api/docs/pricing#gemini-3.1-fla...
- Thinking is now configurable between `Minimal` and `High` (was not the case with Nano Banana Pro)
- Safety of the model appears to be increased so typical copyright infringing/NSFW content is difficult to generate (it refused to let me generate cartoon characters having taken psychedelics)
- Generation speed is really slow (2-3min per image) but that may be due to load.
- Prompt adherence to my trickier prompts for Nano Banana Pro (https://minimaxir.com/2025/12/nano-banana-pro/) is much worse, unsurprisingly. For example I asked it to make a 5x2 grid with 10 given inputs and it keeps making 4x3 grids with duplicate inputs.
However, I am skeptical with their marquee feature: image search. Anyone who has used Nano Banana Pro for awhile knows that it will strongly overfit on any input images by copy/pasting the subject without changes which is bad for creativity, and I suspect this implementation appears the same.
Additionally I have a test prompt which exploits the January 2025 knowledge cutoff:
That still fails even with Grounding with Google Search and Image Search enabled, and more charitable variants of the prompt.tl;dr the example images (https://deepmind.google/models/gemini-image/flash/) seem similar to Nano Banana Pro which is indeed a big quality improvement but even relative to base Nano Banana it's unclear if it justifies a "2" subtitle especially given the increased cost.
You can argue things like code generation are an extension of the engineer wielding it. Image generation just seems like a net negative overall if it’s used at scale.
Edit: By scale, I mean large corporations putting content in front of millions. I understand the appeal for smaller businesses where they probably weren’t going to pay an artist anyway.
I'm old-fashioned so I still Photoshop it all together, but that's my use case here.
Now extrapolate to all other artforms. Sculpture seems safe, for now, but only barely so.
Let's give him 2015 tech instead. Imagine if he used Illustrator to create the Mona Lisa. Is that much better?
I think part of the issue with architects and designers today is that they use CAD too much. It's easy to design boxes and basic roof lines in CAD. It's harder to put in curves and more craftsman features. Nano Banana's renders have more organic design features IMO.
Our house is looking great and we're very happy how it's going so far with a lot of the thanks to Nano Banana.
I find it does a good job at isometric views from floor plans. However, I needed Gemini 3.1 Pro to be able to have a chance at rendering 3D human point of view images from floor plans.
If you can afford the extra cost for someone to figure out how to build the blue sky designs that nano banana spits out, maybe you can afford something more thoughtful and interesting than a shitty mashup of other peoples mcmansions.
Clearly i am triggered..
Here's some of my captions that tend to trip up even state-of-the-art models.
https://mordenstar.com/other/nb-pro-2-tests
So far it does feel more iterative than an entirely new leap in terms of capabilities, but I haven't run it through the more multimodal aspects such as editing existing images.
That being said, it actually managed the King Louie jump rope test which surprised me.
Just think we conceptually know what a brushless motor design looks like and it's just pixels. I guess even if it did produce the image we wouldn't know what it means.
The "cubism" example seems like it would be a closer fit to something like stained glass or something. I don't think the thing really understands what cubism was all about. Cubist painters were trying to free themselves from the confines of a single integral plane of perspective by allowing themselves to show various parts of the image from different viewpoints, different times, different styles, etc.
The division of the image into geometric shapes is just a by-product of that quest, whereas the examples here have made it the sum total of the whole piece.
This feels to me like an example of how LLMs still don't "understand" what the art means, and are just aping its facade.
Probably about half of us here remember photos before the cell phone era. They were rare, and special, and you'd have a few photos per YEAR to look back on. The feel of photos back then, was at least 100x stronger than now. They were a special item, could be given as a gift. But once they became freely available that same amount of emotion is now split across many thousands of photos. (not saying this is good or bad, just increased supply reducing value of each item)
With image/art generation the same thing will happen and I can already feel it happening. Things that used to be beautiful or fantastic looking now just feel flat and AI-ish. If claymation scenes can be generated in 1s, and I see a million claymation diagrams a year, then claymation will lose its charm. If I see a million fake Tom Cruise videos, then it oversaturates my desire for desire for all Tom Cruise movies.
What a time to be alive.
I think this is still true if you shoot film today.
None of these things are true for me as a millennial in the 35-45 age group. And my family was poor to boot, and we were still drowning in photos and photo albums.
> The conditions for an analogous insight are more favorable in the present. And if changes in the medium of contemporary perception can be comprehended as decay of the aura, it is possible to show its social causes.
[0] https://web.mit.edu/allanmc/www/benjamin.pdf
In economic terms it's diminishing marginal utility.