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.

California recently added new laws regarding AI in real estate. I think lighting corrections and cropping is allowed but other digitally altered images should include a link to the original.

https://lewisbrisbois.com/insights/clientalerts/new-californ...

https://dre.ca.gov/Licensees/Advisories/Advisory_2026_03_17_...

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?

Quite a few times I've seen permanent light fixtures that don't exist, vents that don't exist, room sizes that are obviously implied to be much larger than reality (e.g. they show a full-size bed, but there's only like 4 feet of space in that location), etc.

I don't particularly mind fake furniture, but if it's very much not to scale I think it's pushing "probably fraud". And when permanent fixtures are fabricated, "blatant fraud, penalize immediately, revoke license on repeats". Using an automated tool does not absolve you of consequences, particularly one nigh-universally well-known to fabricate things.

Oh, and little touches like an ugly fence being replaced by a sweeping view of a beach or a mountain range.

I think the real issue here is lack of progress in AR technology. Tenants may be disappointed by the difference between marketing material and reality, but that can be easily solved by AR glasses that the tenants can wear 24/7 to make the apartment look just like in marketing material, perhaps for a small monthly subscription fee. It's cheaper than renovation, that's for sure.

Mark Zuckerberg should take note.

‘Catfishing is fine’

Honestly, given the dangerously unmitigated power of Claude Mythos, we should really look into arresting the people who have failed to ask Claude to cure cancer already.

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.

/s?

Wouldn't that fall under existing false advertising laws, if you're putting fake/altered images in the listing?

It should, I would assume. But for some reason, it seems nobody is enforcing consumer protections like they used to.

Pretty soon they'll take the stickers off mowers warning people to not put their hand under it while it's running.

The world isn't in a good place...

> it seems nobody is enforcing consumer protections like they used to

I'll preface this by saying that I don't agree with how things are working currently. But, the way the existing protections get enforced is by an individual or group of individuals filing lawsuits.

Our regulatory agencies have either been completely gutted by DOGE, or just no longer have an appetite to do any kind of proactive enforcement, so its now up to the victims who have been wronged to bring the violation to the attention of the courts/regulatory bodies responsible for enforcement.

Part of me feels like this is intentional and by design, because bringing suit is an expensive and time consuming process and naturally locks out the people harmed the most by these violations. Its the same method by which slumlords/bad landlords get away with so many blatant violations of the various landlord/tenant laws; their tenants can't afford the lawsuits and the lawyers to protect themselves.

The legal system needs to be made much more accessible, but I'm not sure how that happens or what that looks like.

I mean, what’s the problem with taking such stickers off? I’d love if it we had fewer retarded warnings. And I fail to see how they have anything to do with enforcing consumer protections.

Trump Mobile did the same with the false advertising of their phone, and that’s apparently okay.

> that’s apparently okay

Everything is "okay" with Trump. This social experiment got out of hand long ago. I mean come on, if we (as a society) allowed and okayed pedophilic tendencies, why do we keep looking for the lowest point on the bar? I'm pretty sure, Trump can consume human flesh on live TV and we'd just shrug it off and forget that even happened two weeks later.

It seems the solution could be quite simple.

Basically you buy/rent whatever was advertised, and if reality doesn't match, welp thats a defect the seller/landlord must fix at their own expense.

Room shows a vent but in reality there isn't one? Well, the seller has to install it or cover the costs of what an installation would cost.

Accepting 100% that it should be in some way deemed unacceptable (socially or legally) to fake what an apartment actually looks like, I did find using an image model really helpful in making design choices for my bathroom remodel. Mostly about whether to tile certain things where we couldn't quite visualize ourselves what the effect on the entire space would be.

It should be the same as rule #1 of machine translation: never machine translate something for your recipient unless they ask for it. They may not need it. If they do need it, they almost certainly know where to find machine translation. A bad translation with no original is worse than nothing, because you often end up having to mentally backtranslate a broken version of your native language in order to understand what they were trying to say.

Likewise, if I want to see AI renders of what the apartment may one day maybe look like - I can ask for it. Or make the render myself with my tool of choice. But I'll need to know what it actually looks like to do that.

Sadly Google in particular don't obey rule #1 even for machine translation, so it's going to be an uphill battle to get companies to understand.

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.

This

Where I live (NYC) putting altered images like that has been the norm for more than a decade.

It’s just used to be more expensive to hire someone to do it for you.

The altered images always e free stirs the same bright walls and grey magazine style furniture.

AI is just making it cheaper, but this was bound to happen.

(Images altered this way do have a small watermark stating so)

Also NYC. A classic was mounting a bright light outside a window so it appears as “sun-drenched” as the description claimed.

(Unrelated, my favorite one was getting to the apartment and learning the “bedroom” was a flex wall in the kitchen)

AI has very uniquely made creating these faked/impossible layout images one of the cheapest & easiest things you can do at the moment, even if it didn't introduce the concept. Simultaneously, AI has had very little cost reduction impact on much else. This change in relative balance is how AI has created the new version of the problem and it's not apparent how this imbalance was always bound to occur without AI.

I just started seeing these pop up a few weeks ago after some very obvious AI edits appeared in my searches. It’s entirely possibly realtors have been doing this for years now, just in less obvious ways. This crosses the line for me as they’re clearly making spaces look far bigger and brighter than they actually are. Straight up fraudulent and deceptive behavior.

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.

AirBnB folks are doing the same.

This!!

2 months ago while looking for apartments, the majority of the pics shown were generated by AI. The pictures generated by AI often looked much more brighter, cleaner and larger and when I visited them in person, they were the opposite. I wasted so much time visiting due to this.

I understand the intention but the pictures are so wrong most of the time and hide so much imperfection that it should be illegal for false advertisement.

Instead of fighting the use of AI for home interior picture, it might be more useful to have an AI that can correct the fabricated images. If the listing includes room sizes, an AI should be able to give you more realistic images. Maybe a browser plugin that makes all content honest?

Why would that be more useful than fighting this fraud? Surely the original images would be more realistic than one passed through AI two times.

I think it makes houses/apartment less likely to sell. When you see the idealized version, and then the reality, the impact is much bigger then just showing reality.

Unless people prove me wrong, and they really fall for that...

Its like we used to be flooded with fisheye lens pictures of homes, that made the rooms way bigger then reality. I noticed that this trend (on the immo that i follow for years) has heavily reduced. Because nothing beats a sale, as people seeing something looking spacious on pictures and then in person seen its way more small/cramped/compact.

I love that new trend of 3D home viewing... It saves you so much time, and saves time for the immo people, filters out a lot of people with less interest.

When you do get the privilege of seeing the unit in person, yeah. This is obviously the case for most home sales.

But there are plenty of rental markets where you can be forced to rent without seeing that exact unit first. Common in big complexes, where you might get shown a "similar unit" or in markets where rental vacancy is so low that if you don't apply & sign within hours, you aren't getting that apartment because there's 20+ potential tenants for every rare vacancy. The current renal home I live in I rented without seeing it in person first because it was the only vacancy at the time, and in that market you must be first to sign the lease or you lose.

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.

Honestly a great start up would be a review system for house listings.

Users can rate how accurate the description was, the real life flaws and even upload their own photos.

Side note: last time I looked for a house I really wasted 95% of my time because every house had one unique major flaw that would have made me not even bother going to see it.

Announcers get very touchy with listings data, so even compiling listings from multiple sources is hard without getting cease-and-desisted. Then, realtors will certainly flood competing announcements and post fake reviews. It's an aggressive market.

That's actually illegal in most states.

There needs to be lawsuits over stuff like that. I don't get why people accept blatant false advertising just cause the tech used to do it is new. They may as well be uploading pictures of a real, nicer apartment with a similar layout. What's the difference?

The difference is you'll get caught a luddite and hear you're opposing progress if you try to get in the way of AI doing any and everything.

I don't think that matters in a court of law.

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.

I read it as "Google is indifferent to me" rather than "Google specifically dislikes me", as a point of clarification that they didn't get access by being personally selected.

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?

Haven't experienced any errors related to children. Will get the random copyright error though.

That's nice but it's on google's broken AI studio thing. I can't use half the stuff on there because it requires a Google One account. Which I'm ineligible for because I'm on a workspace account. Can I switch? No, because Google One doesn't support own domains.

So I need to run (AND PAY!) two accounts to have both a nice email address and Banana? Starting to think the correct number of paid google accounts here is zero.

Yeah I'm kinda in the same position. I pay for both One and Workspace for myself and not sure which to use for some of this stuff. I mostly default to my personal account since it has more context, but then bringing in stuff from the Workspace Drive/etc. is an extra few steps.

And some stuff like Project Genie is just flat out unavailable on Workspace which seems weird to me.

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.

While I have no experience with it personally (no interest in image gen) my aunt was raving about current chatgpt image model for "restoring" / working with old photos - sharpening, changing some small details like ill-fitting background. It takes her a bunch of prompting but eventually she gets things just right. In comparison, current gemini output (supposedly) tends to be subtly off, details aren’t quite right, proportions are subtly changed etc.

This is purely about generating images with people in them, I don’t think she’s doing any logic puzzles with gotchas and specific alignments of differently colored blocks and whatnot

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)

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.

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.

How do you get the ~real time prototype things shown in the "hands on" section of this page?

gemini.g lets you add a canvas or use image gen, but it isn't clear to me how you stick in the "space lift" prompt and out comes what is demo'd

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

ChatGPT detail is a lot better though. You can do stuff like complex 6-panel comics that Nano Banana can't match.

Also a lot of the negative comments are from people who hate the very idea of AI art and want it to fail.

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.

I loved Nano Banana Pro. Are there local alternatives yet? I heard about Qwen Image, Klein and recently Krea, any suggestions?

Krea-2 is fantastic. If you can get around the restrictive license, output speed, and JSON prompting, Ideogram 4 probably comes the closest to SOTA models. See my profile for GenAI Showdown, where it's benched against other local and proprietary models.

It actually scored above Gemini 2.5 (aka the original NB) which is pretty impressive.

Public AI benchmarks are fucking stupid and can be completely ignored. Too gamed.

However, your advice is right. KREA2 and Ideogram4. The latter being less commercially usable without what I would surely assume is a pain in the ass.

To a degree.

There are plenty of well-known public tests that have been around since SD 1.5 that I'd have to say if companies are trying to optimize for then they’re failing pretty badly (wine glass filled to the brim, the inverted piano, the nine-pointed star, etc.)

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?

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.

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.

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.

People inside deep mind are using Claude code

NB2 is an impressive tool. Camera File -> Heavy Changes in NB2 -> Final Tweaks in Photoshop -> Production Image

Can it do that with high fidelity to the source; e.g., denoise and restore?

I've been testing for the last 2 hours and compared to GPT Image 2. The pricing for 1K resolution it costs $0.0342, takes 8 seconds (not 3 sec). Compared to the GPT Image 2 api ($0.0414 for the same res, responding in 35-45 seconds) I am pleased with its quality and pricing. The texts come garbled, it adds them to every image generated without the prompt, but negative prompts do work well to remove those.

It tends to generate depressing-looking lighting, but if instructed, it overcomes it. Up close, faces also look pretty decent. I can still tell it is AI-generated, but to an untrained eye, it is nearly impossible to distinguish, especially for regular blog posts and ads.

At least from this page it looks like the images generated don't have a watermark on them. Hopefully there isn't an invisible one either.

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

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.

Where is the pelican on bicycle test results?

Because for imagegen models it'd be a meaningless benchmark. It's designed to test codegen models for their UI/UX capabilities.

So the image model's benchmark is to generate an image with the corresponding SVG sources.