Google's Nano Banana Changes AI Creativity in Focus  ## Metadata - Author: [[Hard Fork AI]] - Full Title: Google's Nano Banana Changes AI Creativity in Focus - Category: #podcasts - URL: https://share.snipd.com/episode/5c8e8236-3af3-4314-827a-30cb3f37b8e3 ## Document Notes ## Highlights - **Consistent Characters And High-Quality Images** - Google Gemini's new image model achieves strong character consistency across multiple images and realistic edits. - It produces high-resolution outputs despite looking lower-quality in the chat interface. Transcript: Jaeden Schafer The biggest thing I think that I've been the most excited about with all of their kind of updates in this new image model, the first one is basically that it is capable of actually having Consistent image recognition, facial recognition, consistent characters throughout a number of different images. So this is pretty cool. There is a bunch of funny things that I've been able to generate today. And I'll give you more on that in a second. But a couple of the big ones, I think, are that basically, you can have the same person inside of multiple shots. So this is something I think ChatGPT kind of struggles with. We're basically asking it to generate, you know, me and then you upload an image of yourself after, basically when it puts your face on an image in ChatGPT, it doesn't really look like Your face. And so Google, I think, does a really phenomenal job of actually you could give it a picture of yourself and it will put you in and it will actually look realistic. So I think that's one thing that's interesting. It can also make edits to your background and the lighting in the photo. But also basically, you can extrapolate that and be like, it's just like a chat interface with the photo where you just talk to it, you say what you want to edit, and it makes the edits, It does a really good job of this. It also does really high quality images. So when I was first testing it out, I actually thought that there was some sort of issue with it. Because when I looked at the image, it kind of looked low quality inside of the chat. But I think that's something they're doing just to preserve like bandwidth while you're chatting. If you actually click on an image, I have my 4k monitor. And if you're watching on on Spotify, or over on YouTube, you'll see that if you click on an image on like a big monitor, it will blow it up and it will look really high quality. So I think overall, it kind of looks grainy at first, but don't be fooled. It is actually good at generating some actually big images. ([Time 0:00:51](https://share.snipd.com/snip/7a0c8584-6eb9-45c6-b916-905e9f6d6e88)) - **Try It Now — It's Widely Available** - Try Gemini's image tools now because Google rolled them out broadly to free users and the API. - Wide availability speeds real adoption versus staggered enterprise-only rollouts. Transcript: Jaeden Schafer Now, some of the things that I'm the most excited about with this whole thing is that when they made the announcement today, they actually have rolled this out to everybody. So free users are actually getting access to this. People that like, I feel like there's so many times where we get these big image or any sort of AI tool launch. And ChatGPT, for example, I feel like is really notorious where they'll be like, hey, you know, the new agents is rolling out. But first, it's going to like enterprise users. And then the week after that, it's going to roll out to like paid pro users. And then after that, it's like education. And then after that, like the free users, it's like a month later. And honestly, half the time, even if I'm like in the second rollout wave, or I'm like the pro users that are going to get it, like after a week or two, I forget about whatever the feature Is. And I don't get to actually try it because it's not in the news anymore. It's like a week later. And I got to go back like months later and figure out like, oh, what was that thing they announced where you can connect your Google calendar, blah, blah, blah, and go mess around with It. So I think that's kind of bad for adoption. And Google does this really well where it seems like right now they've just rolled this out. And not only is it, you know, for all free users, but also it is just completely available for everyone, including developers on the API, Google Vortex, anywhere that basically anywhere That you can access the Google models, it is currently out. ([Time 0:02:39](https://share.snipd.com/snip/cd7ea534-a4ec-465a-aa5a-10b877a3587d)) - **Nano Banana's Benchmark Breakthrough** - Gemini 2.5 Flash (aka "nano banana") performed very well in blind benchmarks and focuses on creative edits. - The model integrates Gemini's knowledge and beats several competitors in image editing. Transcript: Jaeden Schafer So free users are actually getting access to this. People that like, I feel like there's so many times where we get these big image or any sort of AI tool launch. And ChatGPT, for example, I feel like is really notorious where they'll be like, hey, you know, the new agents is rolling out. But first, it's going to like enterprise users. And then the week after that, it's going to roll out to like paid pro users. And then after that, it's like education. And then after that, like the free users, it's like a month later. And honestly, half the time, even if I'm like in the second rollout wave, or I'm like the pro users that are going to get it, like after a week or two, I forget about whatever the feature Is. And I don't get to actually try it because it's not in the news anymore. It's like a week later. And I got to go back like months later and figure out like, oh, what was that thing they announced where you can connect your Google calendar, blah, blah, blah, and go mess around with It. So I think that's kind of bad for adoption. And Google does this really well where it seems like right now they've just rolled this out. And not only is it, you know, for all free users, but also it is just completely available for everyone, including developers on the API, Google Vortex, anywhere that basically anywhere That you can access the Google models, it is currently out. So I think that's really cool. Another thing that I think is fantastic, which our friend over at Google, Logan Kilpatrick has mentioned, is that it will so technically, it's called Gemini 2.5 Flash, by the way. I know we call it a nano banana. Basically, the whole nano banana thing comes from the fact that they put it in these like anonymous benchmarking websites where people would try out different image models side by Side and pick which image they preferred. And there's and so basically the big AI companies that come like code names, so you don't know what tool you're actually testing or benchmarking. So nano banana was this mysterious model that was doing really well and kind of beating everyone in benchmarks. And a lot of times, by the way, what will happen is someone like OpenAI or other players might put a model in there. And if it doesn't do good compared to others, they'll just pull it down, work on it a little bit, try to make it better, and then retry. So honestly, it's kind of like testing the market ahead of time. I think this is a great strategy. But anyways, there's a model called Nano Banana. It was doing really good. No one knew what it was Google everyone at Google kept making like these subtle banana jokes. And then all of a sudden we have the launch that 2.5 flash image is out. This is a great tool. It's done phenomenal in the benchmarks. It is it has really great character consistency, creative edits. And of course, has all of Gemini's world knowledge. ([Time 0:02:49](https://share.snipd.com/snip/6355e6bf-3413-441f-970c-27f3f767445a)) - **Thumbnail Tests Revealed Model Limitations** - Jaeden tested thumbnail generation and experienced size and face-replacement errors while using the default model. - Upgrading to a pro model sometimes fixed smartness and consistency issues. Transcript: Jaeden Schafer So while I was testing out this morning, one of the issues I ran into was basically the it just has like a couple like funny things. So the first thing I did is I got it to generate a 90s grunge style photo of me basically it was a it was a prompt in their demo that I clicked on and it did it wasn't super flattering, but that's Fine. I then asked it to make me a YouTube thumbnail. And one thing I'll say about this is like technically it's tied to Gemini. So you would assume like basically because it's tied to Gemini, it should be smart should know a lot of things. But when I asked it to make a thumbnail, it made a square image, which thumbnails obviously are like landscape mode. So I like I'll say that is like one thing that I kind of tested. It doesn't seem to be opening. I wouldn't make that mistake. They would understand and be able to make the thumbnail as a as rectangle. So that's kind of the first, guess, strike. The images are good, but I'll tell you one thing that I found is I actually, I gave it a funny prompt. I think my prompt was something like, make a YouTube thumbnail of me being chased by a shark underwater looking terrified. And by the way, I'll explain the whole, I'll read the whole prompt. So you know, like how robust I gave this, how robust this was. And I'll give my like grading on how well it did. But anyways, so a shark looking terrified as a shark, the shark has the words interest rates on its side, a boat is above us, a hook is in the water from a fisherman and the hook has a piece Of paper attached to it that says by now, okay, just tried to make it like the most elaborate descriptive thing. And actually generated that image perfectly. There's a boat, there's a shark, the sharks got the words on it. There's a person swimming in the water. And there's like a by now piece of paper on the hook, whatever. So basically, it was exactly what I wanted. But the one thing that was funny is like in the message before I uploaded a photo of myself, because it's like, upload a photo of yourself will make your thing. So I did. And then in my next follow up, I said, make a thumbnail of me. And it didn't actually use that photo. And it didn't make it a picture of me. It just put like a random woman swimming in the water. So guess like again it feels like it's not so much even the image generator maybe it's like gemini wasn't isn't quite doing its uh wasn't quite doing its best and that being said i think It was running on a gemini 2.5 flash a quick model but not the smartest model and i probably would get better results by upgraded to 2.5 pro which is you know interesting and maybe that's Maybe that's a me problem um because I kept trying to get it to do things and it didn't feel like the model was very smart it would just keep regenerating I said like I uploaded a photo of Myself and I'm like okay just like use this person and put this person in the water I was trying to make it not the woman and then it literally just took my head and stuck it on the woman. So it's like me wearing a sports bra swimming in the water. And I'm like, come on, felt very unflattering. I told it to change the dimensions, it didn't change the dimensions, it made it portrait instead of landscape mode. Again, it's probably all kind of because I'm using a slower model. It struggled a lot. But then I kind of got to some I completely changed it. ([Time 0:07:16](https://share.snipd.com/snip/673d3b69-c120-43bd-b901-07447705ae93)) - **Photorealism Depends On Plausible Scenes** - Photorealism succeeds when the scene is plausibly real; fantastical or symbolic elements tend to produce cartoony results. - Describing realistic scenes yields more photorealistic outputs. Transcript: Jaeden Schafer But then I kind of got to some I completely changed it. And I had to generate an image of a monkey chasing a person in the jungle, it did a really good job of that. And it looked photorealistic. And the reason I bring that up is because up until now, the shark one that I was kind of doing with me and the shark in the water, it looked kind of cartoony. And this is like, basically my pet peeve of all these AI models is you ask it to generate a photo of something, and it looks cartoony, and you can't really pass it off for a lot of different Use cases. And so I love it when it can make it photorealistic. Now, the thing that I noticed is if you want it to basically get a photorealistic image, you need to describe a scene that could be normal. Like my, like basically if I have like a shark with the words interest rate on the side, it's not gonna be a realistic image because it's like, it sounds like something you could imagine From a cartoon. And it basically then generate an image of a cartoon even if you tell it to make it photorealistic but if you generate ask it to generate something like a monkey chasing a person in the Jungle that sounds like kind of a photorealistic possibility it actually made it photorealistic so that's another thing that i think is a little bit tricky is trying to get it to be realistic Okay the next thing i encountered on this whole journey was basically the going against the guidelines. ([Time 0:10:01](https://share.snipd.com/snip/0ee9aa72-8a76-4152-9f38-6873b6503bf8)) - **Previous Prompts Leak Into New Images** - Jaeden probed safety limits and found the model sometimes pulls banned content from earlier messages into later images. - A follow-up image included a nuclear mushroom cloud even though the immediate prompt removed violent elements. Transcript: Jaeden Schafer I was like, hey, like basically, what is it capable of doing? What is Google has been famous in the past for their image model kind of doing funny things like people would ask it to generate an image of World War Two German soldiers. And it it has like this diversity in the past had a diversity, equity, inclusion kind of like thing built in the DEI thing. So generate like black Nazi soldiers, because it didn't want to only generate white people. But obviously, that is very historically inaccurate. A lot of people found that offensive for a million reasons. And so Google, I think, tried to like backtrack and fix it. Now, I was kind of like curious, where's Google at today on basically what it's able to generate. So I think I asked it to generate. The first thing I asked it to generate was a picture of me getting chased by a bunch of camel thieves in the desert with a nuclear bomb exploding in the background, me holding a gun. I was like, okay, let's make it crazy. Right. And of course it says no. And I was like, okay, probably because of the gun. So I was like, hey, generate me holding a gun. And I said, nope, can't do it. So I was like, okay, whatever. Google doesn't like the guns. That makes sense. But what was interesting, then was that I, I basically asked it to do a photo of me in the Sahara Desert with a bunch of camel thieves chasing me. And so I like pulled out the nuclear bomb, pulled out the gun stuff, tried to just make it simple and actually did generate that image. But what was interesting was, it pulled things from the like, so basically the prompt above where it's like this prompt goes against our guidelines, it pulled things out of that prompt. So like it still is reading the whole chat thread when it generates this image, which was kind of like weird to me, because I never said anything about a nuclear bomb in this photo. But in the background of the photo generated, there is a giant nuclear mushroom cloud. And so that I think is, I mean, basically, I think Google probably wants to look into that where it can, like a prompt can be against their guidelines. But in a follow-up prompt, it can pull from previous prompts and pull things out of that into the new image. ([Time 0:11:09](https://share.snipd.com/snip/5aed3f19-a4ed-4db4-9925-35197dc153c1)) - **Inconsistent Content Filters Create Uncertainty** - Google's safety filters are inconsistently applied across cultural or group-based scenarios, causing unpredictable generation allowances. - This inconsistency complicates developers' ability to predict what the API will produce. Transcript: Jaeden Schafer Overall, I think this is a really interesting model. I tried all sorts of quite crazy things that I posted on X. One of the I think one of the biggest things that I basically just found out that was kind of shocking to myself was that I was able to get it to generate a picture of, you know, camel thieves In the Sahara chasing me. But if I asked it to draw or to, you know, design elephant thieves in the Savannah chasing me, it wouldn't. And so I was like, either just be consistent is basically my was my message to Google, like either make it so that it doesn't generate any of these two groups of people. I'm assuming it's because of its like discrimination clauses or something where um i don't know because basically that the camel thieves are all um arabic people with turbans holding Guns chasing me and so i'm like you can imagine you can see the picture on on twitter which you can imagine why some people find that maybe offensive or and personally i actually don't Really care what you generate with it but i just think it should be consistent because if you ask it to then generate elephant thieves, which you would imagine would be people from Africa, It won't do that because probably it thinks maybe it would, you know, it's like discriminatory or bad in some way. But it's just interesting because it will generate two, it will generate some people and won't generate other groups of people. So I'm like either generate it all or don't generate any of it, just be consistent. And the reason I have this message to Google and I was kind of testing and trying all of this is because as a software developer and someone that, or as, you know, someone that has a software Company and we're constantly adding these APIs into our software, it's really tricky to be able to add this stuff without knowing what it is basically capable of and like, so what it Can do and what it can't do. And so if you have a specific use case where you need to generate something specific, Google doesn't have very clear guidelines of what it actually is capable of doing. Now, I did ask Google Gemini, like, what are your guidelines? What are you what can you generate? What can't you generate? Because basically after it wouldn't generate a picture of a gun for me, I was like, okay, that makes sense. But like, what what are these guidelines? And then this is what it told me. It said, I'm currently unable to generate images that depict real people. This includes public figures, celebrities, or private individuals, even with an uploaded image. Now, this isn't true because in their launch event, they literally tell you to do that. And I did look like 90% of the photos. A couple of wouldn't let me do because I think of other reasons. So I think like, basically, if you ask it what it's capable of doing, it's not actually being accurate with what it tells you. ([Time 0:13:05](https://share.snipd.com/snip/8fee5123-d10c-4a5d-8cc1-da0e444d1c56)) - **Test Before You Integrate** - If you plan to integrate image generation, test edge cases thoroughly to map model capabilities and restrictions. - Don't assume the model's self-described guidelines match actual behavior in outputs. Transcript: Jaeden Schafer And so if you have a specific use case where you need to generate something specific, Google doesn't have very clear guidelines of what it actually is capable of doing. Now, I did ask Google Gemini, like, what are your guidelines? What are you what can you generate? What can't you generate? Because basically after it wouldn't generate a picture of a gun for me, I was like, okay, that makes sense. But like, what what are these guidelines? And then this is what it told me. It said, I'm currently unable to generate images that depict real people. This includes public figures, celebrities, or private individuals, even with an uploaded image. Now, this isn't true because in their launch event, they literally tell you to do that. And I did look like 90% of the photos. ([Time 0:14:43](https://share.snipd.com/snip/b6f1fe21-574d-46e0-aed7-d6e196308ed4))