How Google’s Nano Banana Achieved Breakthrough Character Consistency

0:00There's something about like visual media that really excites people. that it's like the fun thing, but it's not just fun. It's exciting. It's intuitive. The visual space is so much of how we as humans experience life that I think I've loved how much it's moved people. I think we're really now making it possible to like tell stories that you never could. And in a way where like the camera allowed anyone to capture reality when it became very accessible, you're kind of capturing people's imagination. Like you're giving them the tools to be able to like get the stuff that's in their brain, out on paper, visually, in a way that they just couldn't before because they didn't have the tools or they didn't have the knowledge of the tools.

0:41Like, that's been really awesome.

0:59Today, we're talking with Nicole Brichtova and Hansa Srinivasan. the team behind Google's nano-banana image model, which started as a 2AM codename and has become a cultural phenomenon since. They walk us through the technical leaps that made single-image character consistency possible, how high-quality data, long multimodal context windows, and disciplined human evals enabled reliable character consistency from a single photo, and why craft and infrastructure matter as much as scale. We discussed the trade-offs between pushing the frontier versus broad accessibility and where this technology is headed.

1:35Multimodal creation, personalized learning, and specialized UIs that marry fine-grained control with hands-off automation. Finally, we'll touch on what's still missing for true AGI and white space where startups should be building now. Enjoy the show. Nicole and Hansa, thank you so much for joining us today. We're so excited to be here to chat a little bit more about Nano Banana, which has taken the world by storm. We thought we'd start off with a fun question. What have been some of your own personal creations using Nano Banana or some of the most creative things you've seen from the community?

2:10Yeah, so I think for me, one of the most exciting things I've been seeing is like the... it didn't occur to me, but this is very obvious in hindsight, is the use with video models to get actually consistent cross scene character and scene preservation. How fluid is that workflow today? How hard is it to do that? So what I've been seeing is people are really mixing the tools and using different video models from different sources. And so I think it's probably not very fluid. I know there's some products out there that are trying to integrate with multiple models to make this more fluid but i think the the difference in the the videos i've been seeing from before and after the nano banana launch has been pretty pretty remarkable and it's like much much smoother and much more like what you'd want in the video creating process with scene cuts that feel natural yeah so that's been cool um and i don't know why it didn't totally occur to me that people would immediately do that but yeah one of my favorite ways that i didn't expect is how people have hacked around the model to use it for learning new things or digesting information.

3:19I met somebody last week who has been using it to create sketchnotes of these like varied topics and it's surprising because text rendering is not something that it's not where we want it to be but this person has hacked around like these massive problems that like get the model to output something that's coherent and he's used it to try to understand the work that his father's doing He was like a chemist at a university and it's like a super technical topic. And so he's been feeding his lectures to Gemini with Nano Banana and then getting these sketchnotes that are like very coherent and like visually digestible.

3:56And for the first time, I think in like decades, they've been able to have a conversation with each other about his dad's work. And that was really fun and something that I didn't see coming. I think people are really working around, you know, like this model is amazing, but obviously it's not perfect. we have a lot of things we want to improve and I think I've been astounded by the ways people have found to to work with the model in ways we didn't anticipate and give inputs to the models in ways we didn't anticipate to bring out the best performance and unlock these things that are kind of mind-blowing did you guys in the building of it was there a moment like an aha moment where you kind of felt wow this thing's gonna be pretty good we just talked about it yeah I think Nicole had the aha moment.

4:40I had one where, so we always have an internal demo where we play with the models as we're developing them. And I had one where I just took an image of myself. And then I said like, Hey, put me on the red carpet and like full glass, just total vanity prompt. Right. And then it came out and it looked like me. And then I compared it to like all the models that we had before. Um, and no other model actually looked like me. And I was like, so excited. Wow. Um, And then people looked and they were like, OK, yeah, we get it. Like you're on the red carpet. And then I think it took a couple of weeks of other people being able to take their own photos and play with it and just kind of realize how magical that is when you get it to work.

5:21And that's kind of the main thing that people have been actually doing with the model, right? Turning yourself into a 3D figurine where it's like you want a computer, you want a toy box and then you as the figurine. So like you three times like that way to be able to kind of like express yourself and see yourself. in new ways and almost kind of like enhance your own identity has just been really fun. And that for me was like, oh, man, this is awesome. What was it about what Nano Banana did with you on the red carpet that was miles better than what everyone else has? It looked like me. And it's very difficult for you to be able to judge character consistency on people's faces you don't know.

5:58Yeah. And so if I saw, you know, a version of you that's like an AI version of you, I might be OK with it. But you would say like, oh, no, you know, like parts of my face are not quite right. And you can really only do it on yourself, which is why we now have evals on many team members where it's like their own faces and they're looking at the model's output with their own faces on it. Because it's really the only way that you can judge whether or not someone looks like you. Yourself and like faces you're familiar with. I think like when we started doing it on ourselves and it's like I see Nicole a lot.

6:29So like Nicole versus like random person we might eval on. Right. It's just a very big difference in terms of judging the model capabilities. And yeah, I think it's one of those things that it's like so fun, that preservation of the identity is so fundamental to these models actually being useful and exciting, but is surprisingly tricky. And that's why we see a lot of other models not quite hitting it. Well, I was going to ask you, I would imagine that character consistency is not just an emergent property of scale. And so maybe two questions. One, I'm sure there's stuff you can't tell us, but what can you tell us about how you achieved it?

7:08And then two, was that an explicit goal heading into the development of this model? Yeah. So I would say, I mean, yeah, I think there's definitely things that are tricky to say here, but I would say there's like sort of different genres of ways to do image generation. And so that definitely plays a part in how good it is. And I think it was definitely a goal from the beginning. It was definitely a goal because we knew it was a gap with the models that we released in the past. And generally consistency for us was a goal because every time you're editing images, right, like you want to preserve some parts of it and then you want to change something.

7:47And prior models just weren't very good at that. And that makes it not very useful in professional workflows, but it also doesn't make it useful for things like character consistency. And we've heard this for years from even advertisers who are trying to advertise their products and putting them in lifestyle shots. It has to look like your product 100%. Otherwise, you can't put it in an ad. So we knew there was demand for it. We knew the models had a gap. And we felt like we had the right recipe, both in terms of the model architecture and the data, to finally make it happen. I think what surprised us was just how good it was when we actually finally built the model.

8:21Yeah, right. Because I think we felt like we had the recipe exactly as Nicole said, but there's still always until you're seeing the model, you finish training, you're actually using it. You don't know how close you're going to get to that goal. And I think we were all surprised by that. yeah and I think the other thing is if we think about like what people expect out of editing when that you edit on your phone apps or like photoshop you expect a high degree of preservation of things you're not touching yeah and depending on how the models are made and how how the design decisions behind them that's very tricky to do but it's something people really like it's it's one of those things were like, it's shockingly technically difficult, even though it's something I think a lay person who's using the models would be expect to be like the basic thing about editing is like, you don't mess with the things you don't want to be messed with.

9:13Yeah. Back to that moment where you saw yourself on the red carpet and wow, that's actually me. And it took some of your colleagues a couple of weeks to have the same experience because they tried it with their own photos. The question is beyond, Hey, that's actually me that, you know, the qualitative test. Is there some sort of an eval that you can put against that to make it quantitative that, you know, we have achieved the thing that we set out to achieve here? Yeah. So I actually think, I think face consistency exactly for the reason Nicole said is, is quite hard. It's quite hard for other people to do.

9:46Yeah. Um, I will say in general, I think what we found with image generation in particular, that's unlocked a lot for us is like human evals are important. And so I think they're foundational. We have a team that works on helping us build sort of good tooling and good practices for evals and having humans actually eval these things that are very subtle. Like if you think about image generation, like faces, aesthetic quality, these are things that are very hard to quantify. And so I think human evals have been a big game changer for us. I think it's definitely, I think it's a combination of there's human evals, there is a very technical term, eyeballing, of the model results by different people.

10:30And there's also just community testing. And when we do community testing, we start internally. And we have artists at Google and at Google Define to play with these models. Our execs will play with these models. And that really helps, I think, kind of build that qualitative narrative around like, why is this model actually awesome? Because if you just look at the quantitative benchmarks, you could say like, oh, it's 10 % better than this model that we had before. And that doesn't quite grok that emotional aspect of like, oh, I can now see myself in new ways, or I can now finally edit this family photo that I cut up when I was five years old.

11:07And I probably shouldn't have. People have done that. When then I'm able to restore it. I think you really need that qualitative user feedback in order to be able to tell that emotional story. I think this is probably true of many of the Gen.A.I. and AI capabilities, but I think it's especially true of visual media where it's very subjective versus if you think about something like math reasoning, logic reasoning, where you can really ground it in an answer. And so it's more easy to have these very objective, automated, quantitative evals. To get to that level of character consistency from just one 2D image of someone is really, really hard.

11:50Can you walk us through maybe a little bit what are the technical breakthroughs that helped you drive to that level of character consistency that we actually haven't seen anywhere else? I mean, I think a key thing is like having good data that teaches the models to generalize. Right. And the fact that this is a base, it's a Gemini model. Yeah. It's a multimodal foundational model that's had seen a lot of data and has good generalization capabilities. And I think that like that's kind of the secret sauce. Right. It's like you really need models that generalize well to be able to take advantage of that for this.

12:27Right. Yeah, and I think the other nice part about doing this in a model like Gemini is that you also get this really long context window. So, like, yes, you can provide one image of yourself, but you can also provide multiple. And then on the output side, you can also iterate across multiple turns and actually have a conversation with the model, which wasn't possible before, right? One, two years ago, we were fine-tuning on 10 images of you, and it took 20 minutes to actually get something that looked like you. And that's why it never took off in the mainstream, right? Because it's just too hard.

13:00And you don't have that many images of yourself. It's too much work. And so I think it's both kind of the general, like Gemini gets better. You benefit from that multimodal context window. And you benefit from the long output and ability to maintain context over a long conversation. And then you also benefit from actually paying attention to the data, focusing on the problem. a lot of the things we get better at come down to there's a person on the team who's like obsessed with making them work like we have people on the team who are obsessed with text rendering and so our text rendering just keeps getting better because that person just like is obsessed with the problem yeah it's like it's not just about throwing high quantities of data in right i think that's one thing that's really important is it's there's this like attention to detail um and quality of, you know, all the things you're doing with the model.

13:50There's a lot of, there's a lot of small design decisions and decision points at every point. And I think that like detail orientedness of high quality is data and selections are really important. Yeah. It's the craft part of, I think, the AI, which we don't talk about a lot, but I think it's super important. Yeah. How big was the team that worked on it? To ship it, it took a village. Yeah. Especially because we ship across, many products so i think like there's like sort of the core sort of modeling team and then there's you know our close collaborators across like all the surfaces yeah when you put them all together you easily get into like dozens and hundreds but um the team who works on the model is much smaller and then the people who actually make all the magic happen and we had a lot of infrastructure teams like optimizing every part of the stack to be able to serve the demand that we were seeing which was really awesome um but really like to ship it we were joking that it takes like a small country.

14:44When you build something like this, do you build it with particular personas or particular use cases in mind? Or do you build it more with a capability first mindset? And then once the capabilities emerge, you can map it to personas? It's a little bit of both, I would say. Before we start training any new model, we kind of have an idea of what we want the capabilities to be. and some design decisions like how fast is it at inference time right they also impact which persona you're going after yeah so this model because it's kind of a conversational editor we wanted it to be really snappy because you can't really have a conversation with a model if it takes like a minute or two to generate that's really nice about image models versus video models like you just have to wait that long and so to us from the beginning it felt like a very consumer-centric model.

15:37But obviously, we also have developer products and enterprise products, and all of these capabilities end up being useful to them. But really, we've seen a ton of excitement on the consumer side in a way that I think we haven't before with our image models, because it was very snappy, and it kind of made these like pro level capabilities just really easily accessible through a text prompt. And so that's kind of how we started it out. But then obviously, it ends up being useful in other domains as well. Yeah. And I think one of the differences in philosophy, so previously we'd worked on the Imagine line of models, which were straight image generation.

16:15And I think one of the big philosophical goal changes in these Gemini image generation models is generalization, is a more foundational capability. So I think there is also a lot of... There's things where we want this model to be able to be good at this. like representing people and letting them edit their images and have it look like themselves. But I think there's also a lot of things like that are emergent from the goal of just having a baseline capable model that like reasons about visual information. Like I think one thing that's surprised me, I guess, as a callback to your earlier conversation is people can put in math problems, like a drawing of a math problem and like ask it to like render the solution, right?

17:01So you can put in a geometry problem and say, what is this angle? And that's an emergent thing of a foundationally capable model that has both reasoning, mathematical understanding, and visual understanding. So I think it's both. Can you maybe share, just out of curiosity, what's a good way to understand maybe the family mapping and the relationship between Gemini powering Nano Banana, VO, you know, all these other adjacent products and models that are all driven and benefit from the generalization and the scale of Gemini itself, how you co-develop and then where you want to take it from here?

17:43Our goal has always been to build the single most powerful model that can do all these things, right? You can take in any modality and you can transform it into any modality. And that's the North Star. We're obviously not quite there yet. And so on the way there, we had a lot of sort of specialized models that just got you great results in a specific domain. So Imagine was an example of that for image generation. VO is an example of that for video generation and editing. And so I think we're both kind of developing these models to push the frontier of that modality. And you get really useful outputs out of that, right?

18:19A lot of filmmakers are using VO in their creative process. but you're also learning a lot that you can then bring back into Gemini and then make it good at that modality. Image is always a little bit, I think, ahead of the curve because you just have one frame, right? It's cheaper both to train and at inference time. So I think kind of a lot of the developments you see in image, I expect you to see in video like six, 12 months down the line. And so that's always kind of been the goal. And so we have separate teams kind of developing these And then I think with image, we're now moving closer to Gemini and to that vision of that single most powerful model.

18:59And you will see that, I think, with some of the other modalities. And along the way, we'll release these experiences that are just like really powerful and like really exciting in that modality. So like VO3 was really awesome because it brought audio into video generation, right, in a way that we haven't seen before. Genie 3 was really awesome because it lets you in real time kind of navigate a world. And so in order to push that frontier, it's very hard to do all of that at the same time right now in one model. And so to some extent, these specialized models are kind of a testing ground. But I would expect that over time, Gemini should be able to do all these things.

19:34Oh, that's so interesting. Okay. We got to ask you about the name. I suspect that the name was a bit of a... It's an amazing product. I suspect that the name gave it a little bit of a boost. because it's so easy to remember and so distinct. So was it a happy accident or is there some creative genius who knew that this is going to be just the right name? It was a happy accident. So I think as many people know, the model went out on an ILL Marina where many models do. And part of that is you give a code name. If anyone hasn't used ILL Marina, you get to put in your prompt. You'll get back two responses from two models.

20:14they have code names until they're publicly released um and i think it was like we had to someone we were going out at like 2 a.m and uh nicole nicole's our wonderful pm there's another pm you have nina and someone messaged her being like what do we name it and she was really tired and exhausted and she was like this was the name of stroke of genius that came to her at 2 a.m this is you it was not me it was somebody on my team okay who named who named the model um i can't take credit for this another one of our pms um but what was really awesome was like a it was really fun i think that really helps yeah it's easy to pronounce it has an emoji which is critical for branding and she didn't overthink it in this era but she didn't overthink it and what was awesome is everybody just went with it once it went live and and i think it just like felt very googly and very organic and ended up looking like the stroke of marketing genius but no it was it was a happy accident and it just sort of worked out and people loved it and so we leaned into it and now there's you know bananas everywhere when you go into the gemini app um which we did because people were complaining that they were having a really hard time finding the model when they came into the app yeah and so we just made it easier yeah and yeah exactly i think there's like publicly people were like, nano banana, nano banana.

21:34How do I use nano banana? I had someone at Google I work with be like, how do I use nano banana? And I was like, it's Gemini. It's right there. Just ask for an image. Yeah. But I think that's the thing is like, I think Google's always had this really fun brand, right? Like it's like, it's not like, it's been a consumer oriented company at its inception and like i think it was really nice to to play on that that image people have of like google as a fun fun place fun company um and have this fun name it's also just like a really nice path to fun being kind of a gateway to utility right i i i think nano banana and just the model in general and what you can do with it like put yourself on the red carpet do all the childhood dream professions you had it's like a really fun entry point but what's been awesome to see is that like once people are in the app and they are using gemini they start to use it for other things yeah that then become useful in their day-to-day life like you use it to study and solve math problems or you use it to learn about something else and so i think it's maybe a little bit undervalued sometimes to like have a little fun um not just with the naming but also just like with the products that we built because it kind of gets people in gets them excited and then it helps them discover other things that you know the models are awesome at yeah i think other users like my my um like my parents and their friends are using it i think it's because it like had this reputation it was really easy it was really fun it felt unintimidating to try when you try it and you're like actually this this is very easy to work this works very easily it's very easy to interact with there's no like technology there's no like you know technology i think can sometimes be intimidating to people especially ai right now yeah um and i think the chat bot naturalness has broken a lot of that barriers but maybe more so with younger people yeah um and i think this like fun like yeah my mom like made was like making these images and having a great time and and and then realized she can use it to like remove people from the background of her images like these very practical things right it started very silly turned very practical then people can use it to realize like actually they can give you them diagrams or help them understand stuff.

23:48So I think there's also like a big accessibility component. Yeah. Where do you want to take from here? Maybe both from a model side and from a product side? On the product side, I think there's kind of a couple areas. Like on the consumer side, I still think we have a long way to go to just like make these things easier to use, right? You will notice that a lot of the nano banana prompts are like 100 words long and people actually go in and copy paste them into the Gemini app and like go through the work to make it work because the payoff is worth it. But I think we have to get past this prompt engineering phase for consumers and just like make things really easy for them to use.

24:25I think on the professional side, we need to get into like much more precise control, kind of robustness, like reproducibility to make it useful in actual professional workflows, right? So like, yes, we, you know, we're very good at editing consistency and not changing pixels, but we're not 100 % there. And when you're a professional, you need to be 100 % there, right? Like you really need kind of these precise, maybe even like gesture-based controls, like over every single pixel in the frame. So we definitely need to go in that direction. And then I think there's like a general direction that I'm really excited about, which is just about visualizing information.

25:01So the example I had about sketchnotes at the beginning and somebody kind of hacking their way around using Nano Banana for that use case, you could just imagine being able to do that for anything, right? And a lot of people are visual learners. I think we haven't really exhausted the potential of LLMs to be able to help you digest and visualize information in whatever way is most natural for you to consume. So sometimes it's a diagram, sometimes it's an image, and sometimes maybe it's a short video that you want to learn about some concept that you're learning in a biology class or something like that.

25:35So I think that's like a completely new domain that I'm really excited about, just these models getting better and getting past the point where 95 % of the outputs that you get out of these models are just text, which is useful, but it's not how we consume information in the real world right now. So it's really interesting. So on the product side, then, are you alluding to the fact that you might want to vertically integrate and build a little bit more product around it? And also, are you alluding to the fact that maybe the way you interact with some of these models isn't just through pure language and prompting over time, but more UI?

26:07Yeah, yeah. I definitely think the chatbots, I think, are an easy entry point for people because you don't have to learn a new UI. You just talk to it and then you say whatever you want to do, right? I think it starts to become a little bit limiting for the visual modalities. And I think there's a ton ahead room to think about what is the new visual creation canvas for the future? And how do you build that in a way that doesn't become overwhelming? Because as these models can do more and more things, it's very hard to explain to the user in something that's very open-ended what the constraints are and how do you work around that?

26:45How do you actually use it in a productive way? So I'm really excited about people kind of building products in those directions. And for us, you know, we have a team called Labs at Google that's led by Josh Woodward. And they do a lot of this kind of like frontier thinking experimentation. They work with us really closely where they take our frontier models and they think about like, what's the future of entertainment? What's the future of creation? What's the future of productivity? And so they've built products like Notebook LM and Flow on the video side. And I'm excited that maybe flow could kind of become this place where you could do, you know, some of this creation and think about what that looks like in the future.

27:21I think in the short term, it's very clear that, you know, this model has things that it's not perfect at. And so in the short term, it's obviously, it should work the way you expect it to every time, not just a lot of the time. And really make it so seamless and fix all these like small things where it's just like a little bit inconsistent in its performance. I think long term, it's, I think, Nicole covered that, which is, to me, it's in order to have that reality of really rich multimodal generation. So like right now, if you ask Gemini to explain something, it'll usually just explain in text unless you ask it for images.

28:03But if you think about like the platforms that have really taken off in the last like 10, 20 years for learning, right? Like we think of like Khan Academy started on YouTube. We think about like Wikipedia has a lot of images. Like it's very image focused. If you look up any math thing, you like diagrams. And so like that should become more like an actual part of the flow and a part of the way you use these models. And to enable that from a modeling point of view, it goes back to like we were talking about this multimodal understanding. and seamless generalization between modalities. Maybe the other interesting area, as we think about kind of these models being more proactive at pulling in, whether it's code or images or video, when it's appropriate for the user intent.