The first example of generating home interiors fills me with indescribable hatred. Recently real estate agents have taken to running every dilapidated unsellable apartment through these AI filters, and you have to scroll through a dozen of these Ikea-chic images of what the apartment presumably could look like, before you are allowed to see the horrors they are trying to peddle at insane prices.
Why should that be illegal?
It’s multiplying the productivity of our economy, instead of someone having to waste time and money making the apartment actually look like that, you can just generate an image of it, that’s massive productivity boost with no harm done to the final product, unless the tenant cares about the slippage between a generated image of an apartment that looks nice and an apartment that’s actually nice.
And plus thats time the real estate agent could have spent prompting claude to cure cancer so its a double win
I'm not sure if you're being serious but it should be illegal because they're producing images that are often not physically possible. At least if an agent stages an apartment with real furniture they are doing something a tenant really could do. But these AI images tend to change the physical dimensions of a room, use images of furniture that don't make sense dimensionally, shift the "natural" light of the room in a way that the sun will never provide and sometimes even change the view through the windows of the room.
Just having a good photographer is amazing. When my friend was selling their place I was amazed and how good the house looked in the listing. How big it looked, when I know it was not big. This was before AI filters were available. So not a new issue but certainly made worse and cheaper to do.
And it's borderline fraud, I think I saw an apartment on Streeteasy where they were able to 'fit' an entire desk, drawers and a queen size bed, obviously these image models just scale these down to proportions that just don't exist in real life.
the actual bedroom could only fit queen size bed ;(
In a sane world, this would be a clear cut case of false advertisement, and the real estate agents would be held liable for fraud. Sadly, we don't live in a sane world.
I received early access to test this model. (through work — Google still does not like me personally lol)
It works as advertised here, and it does behave like a distilled Nano Banana 2 with respect to certain elements such as good text rendering, which Nano Banana 1 does much worse with. It is definitely not at the level of the base Nano Banana 2 of course particularly with highly-nuanced prompts. My main criticism is that you cannot programmatically force aspect ratios with NB2L but you can with NB2.
That said, the price of $0.034/image is higher than expected since price is generally correlated with generation time, and it takes half the time to generate than a Nano Banana 1 image which costs $0.039/image. Google's assertion that you can directly replace NB1 pipelines with NB2L is fair.
Yesterday, Google announced that the Gemini app will allow free image generations (https://blog.google/innovation-and-ai/products/gemini-app/pe...) but did not specify which model would be used: I suspect it's the main motivation for Nano Banana 2 Lite.
You can set aspect ratios with NB2 Lite programmatically through Vertex [1]. I updated the program I use to help create all the images for GenAI Showdown, set the model ID to `gemini-3.1-flash-lite-image`, and was able to use aspect ratios like 16:9, 4:3, and others.
... does it? Are you seeing something I'm not seeing? Number one is that this just doesn't appear to be true (non-lite versions beat it across the board it seems), number two is that this specifically is a low-cost bulk model and not a SOTA frontier model, of course the benchmarks are lower.
That is fair to point out. For those who don't know, ChatGPT Image 2 has an absurd ELO of 1387; compared to the #2 model at 1273, it's over 100 points higher (https://arena.ai/leaderboard/text-to-image). The tradeoff is latency, and ChatGPT Image 2 at High is...slow (~2 minutes at 1024x1024). In both cases it would have skewed the charts here to uselessness.
I want to do a writeup on ChatGPT Image 2 but at this point I don't think people care about nuanced image generation anymore...even though ChatGPT Image 2 crushes all my existing tests.
I definitely appreciated your post about Nano Banana Pro. It's also a genuinely useful time-capsule for how these systems evolve and where they fall short. I've preferred the output of ChatGPT Image 2. I think a post would be very helpful for folks to see what they're missing.
That arena leaderboard has some questionable results. Anyone who's used these models would know that ranking HiDream above Krea2 is a pretty hot take.
Many of these ELO comparative tests (ArtificialAnalysis is guilty as hell on this as well) also have other problems such as a considerable number of "amateur judges" tending to prioritize aesthetics over actual instruction-following given the prompt.
Also (less a critique of Arena.AI necessarily), but the MAI models are so incredibly locked down (e.g. censored) as to be functionally useless. I have a sneaking suspicion its fallout from Tay.
The speed is definitely impressive. I'm seeing under 5 seconds per image vs ~30 seconds for base NB2.
I built an app for my kids that generates illustrated stories for them with them as the characters. I wanted to prioritize likeness while still stylizing the illustrations. I tested a bunch of models but none seem to come close to maintaining likeness when stylized. I find the others generate generic looking characters.
I'm excited to incorporate this into the onboarding of my app since I want the users to experience the aha moment as soon as possible and waiting half a minute+ isn't ideal. I'll still be using the main NB2 for the actual illustrations as this lite version still has slight issues with nuance and consistency as others have pointed out.
I'm way behind in imagegen - only using it occasionally for roleplaying tokens, goofing around, and random personal assets. To me, this is nuts. It's able to create images in like 2 seconds... before with chatgpt it would take 30s-1m for the same quality image. I don't get the negative comments here
Expensive and Google doesn't even have enough resources to decently deploy a model like that. Creating 10 images in parallel gives me RESOURCE_EXHAUSTED error, which is a painfully common error when using Google AI products.
> Google doesn't even have enough resources to decently deploy a model like that
Probably not for free but tbf, Google did scale "AI Mode" globally to its billion+ users, with its Gemini 3 series. Pretty much broke my habit of searching the web with pplx & Chat.
Wow, that's a pretty massive decrease in latency, which should unlock some use cases, but the linked web page doesn't exactly make it straightforward to understand the differences between the models.
However, based off my personal experiences with general images models, Google in my opinion is the best for my workflows. Granted, I haven't tried far-east providers yet.
Nano Banana is head and shoulders above the rest, but still too steep for personal use, and half off doesn't really mean much for enterprise if the results are worse. Hopefully this drives the rest to catch up at least.
The first example of generating home interiors fills me with indescribable hatred. Recently real estate agents have taken to running every dilapidated unsellable apartment through these AI filters, and you have to scroll through a dozen of these Ikea-chic images of what the apartment presumably could look like, before you are allowed to see the horrors they are trying to peddle at insane prices.
I think that should be illegal and misrepresenting. Lots of gray area with AI usage.
Why should that be illegal? It’s multiplying the productivity of our economy, instead of someone having to waste time and money making the apartment actually look like that, you can just generate an image of it, that’s massive productivity boost with no harm done to the final product, unless the tenant cares about the slippage between a generated image of an apartment that looks nice and an apartment that’s actually nice.
And plus thats time the real estate agent could have spent prompting claude to cure cancer so its a double win
What, after all, is a bit of light fraud, if it saves an estate agent some time?
I'm not sure if you're being serious but it should be illegal because they're producing images that are often not physically possible. At least if an agent stages an apartment with real furniture they are doing something a tenant really could do. But these AI images tend to change the physical dimensions of a room, use images of furniture that don't make sense dimensionally, shift the "natural" light of the room in a way that the sun will never provide and sometimes even change the view through the windows of the room.
I think their last sentence is a pretty clear indicator that they were not being serious.
As with bitcoin fans before them, Poe's Law is in full effect with the AI boosters.
Just having a good photographer is amazing. When my friend was selling their place I was amazed and how good the house looked in the listing. How big it looked, when I know it was not big. This was before AI filters were available. So not a new issue but certainly made worse and cheaper to do.
And it's borderline fraud, I think I saw an apartment on Streeteasy where they were able to 'fit' an entire desk, drawers and a queen size bed, obviously these image models just scale these down to proportions that just don't exist in real life.
the actual bedroom could only fit queen size bed ;(
How is that borderline, that’s just actual fraud.
In a sane world, this would be a clear cut case of false advertisement, and the real estate agents would be held liable for fraud. Sadly, we don't live in a sane world.
I received early access to test this model. (through work — Google still does not like me personally lol)
It works as advertised here, and it does behave like a distilled Nano Banana 2 with respect to certain elements such as good text rendering, which Nano Banana 1 does much worse with. It is definitely not at the level of the base Nano Banana 2 of course particularly with highly-nuanced prompts. My main criticism is that you cannot programmatically force aspect ratios with NB2L but you can with NB2.
That said, the price of $0.034/image is higher than expected since price is generally correlated with generation time, and it takes half the time to generate than a Nano Banana 1 image which costs $0.039/image. Google's assertion that you can directly replace NB1 pipelines with NB2L is fair.
Yesterday, Google announced that the Gemini app will allow free image generations (https://blog.google/innovation-and-ai/products/gemini-app/pe...) but did not specify which model would be used: I suspect it's the main motivation for Nano Banana 2 Lite.
You can set aspect ratios with NB2 Lite programmatically through Vertex [1]. I updated the program I use to help create all the images for GenAI Showdown, set the model ID to `gemini-3.1-flash-lite-image`, and was able to use aspect ratios like 16:9, 4:3, and others.
[1] - https://cloud.google.com/developers/vertex-ai
What kind of work are you doing that requires automated image generation at scale?
but until this comes to edge gallery i won't care
> Google still does not like me personally lol
Please elaborate.
It's sort of amazing that Grok's image model beats Nano Banana on nearly every one of the metrics they chose to highlight.
... does it? Are you seeing something I'm not seeing? Number one is that this just doesn't appear to be true (non-lite versions beat it across the board it seems), number two is that this specifically is a low-cost bulk model and not a SOTA frontier model, of course the benchmarks are lower.
They didnt include chatgpt in the comparison chart. That tells a lot
That is fair to point out. For those who don't know, ChatGPT Image 2 has an absurd ELO of 1387; compared to the #2 model at 1273, it's over 100 points higher (https://arena.ai/leaderboard/text-to-image). The tradeoff is latency, and ChatGPT Image 2 at High is...slow (~2 minutes at 1024x1024). In both cases it would have skewed the charts here to uselessness.
I want to do a writeup on ChatGPT Image 2 but at this point I don't think people care about nuanced image generation anymore...even though ChatGPT Image 2 crushes all my existing tests.
I definitely appreciated your post about Nano Banana Pro. It's also a genuinely useful time-capsule for how these systems evolve and where they fall short. I've preferred the output of ChatGPT Image 2. I think a post would be very helpful for folks to see what they're missing.
That arena leaderboard has some questionable results. Anyone who's used these models would know that ranking HiDream above Krea2 is a pretty hot take.
Many of these ELO comparative tests (ArtificialAnalysis is guilty as hell on this as well) also have other problems such as a considerable number of "amateur judges" tending to prioritize aesthetics over actual instruction-following given the prompt.
Also (less a critique of Arena.AI necessarily), but the MAI models are so incredibly locked down (e.g. censored) as to be functionally useless. I have a sneaking suspicion its fallout from Tay.
https://en.wikipedia.org/wiki/Tay_(chatbot)
Whats the use cases where cheaper and faster img models are key differentiators?
I mean 3 cents compared to 6 cents doesnt seem like much in my mind-unless youre running a consumer saas
The speed is definitely impressive. I'm seeing under 5 seconds per image vs ~30 seconds for base NB2.
I built an app for my kids that generates illustrated stories for them with them as the characters. I wanted to prioritize likeness while still stylizing the illustrations. I tested a bunch of models but none seem to come close to maintaining likeness when stylized. I find the others generate generic looking characters.
I'm excited to incorporate this into the onboarding of my app since I want the users to experience the aha moment as soon as possible and waiting half a minute+ isn't ideal. I'll still be using the main NB2 for the actual illustrations as this lite version still has slight issues with nuance and consistency as others have pointed out.
I tried something like that but an error told me it couldn't do anything with children. Did that change?
I'm way behind in imagegen - only using it occasionally for roleplaying tokens, goofing around, and random personal assets. To me, this is nuts. It's able to create images in like 2 seconds... before with chatgpt it would take 30s-1m for the same quality image. I don't get the negative comments here
Different use cases.
People making images, where the image is the focal point, want to spend more per image.
Where images are parts of reports or throwaways or demos, cheap is the better approach.
Expensive and Google doesn't even have enough resources to decently deploy a model like that. Creating 10 images in parallel gives me RESOURCE_EXHAUSTED error, which is a painfully common error when using Google AI products.
> Google doesn't even have enough resources to decently deploy a model like that
Probably not for free but tbf, Google did scale "AI Mode" globally to its billion+ users, with its Gemini 3 series. Pretty much broke my habit of searching the web with pplx & Chat.
Wow, that's a pretty massive decrease in latency, which should unlock some use cases, but the linked web page doesn't exactly make it straightforward to understand the differences between the models.
However, based off my personal experiences with general images models, Google in my opinion is the best for my workflows. Granted, I haven't tried far-east providers yet.
What does everyone else think?
It seems to respond to edits much better than the current production image model, which often stubbornly locks on to prior iterations of the images.
NB2 is an impressive tool. Camera File -> Heavy Changes in NB2 -> Final Tweaks in Photoshop -> Production Image
gemini is so far behind. starting to wonder if their strategy is launching the low cost alternative to image/text models. last release was 3.5 flash
Imagine saying Gemini is so far behind when Llama has unreleased max models from last year that are currently beat by quantized Qwen2.5 models.
It's only half the full model price, $30/m output: https://cloud.google.com/gemini-enterprise-agent-platform/ge...
Nano Banana is head and shoulders above the rest, but still too steep for personal use, and half off doesn't really mean much for enterprise if the results are worse. Hopefully this drives the rest to catch up at least.