Ep 75: Nano Banana’s Oliver Wang and Nicole Brichtova - Behind the Breakthrough as Gemini Tops the Charts

This week on Unsupervised Learning, Jacob sits down with Nicole Brichtova and Oliver Wang, the Google researchers behind "Nano Banana" - the breakthrough AI image model that achieved unprecedented character consistency and took over social media. The conversation covers how their model fits into creative workflows, why we're still in the early innings of image AI development despite impressive current capabilities, and how image and video generation are converging toward unified models. They also share honest perspectives on current limitations, safety approaches, and why the expectation of going from prompt to production-ready content is fundamentally overhyped.

• 0:00Speaker 00

  Jim and I recently reached the top of the App Store, finally overcoming ChatGPT for the first time since ChatGPT's launch. And what drove it? Well, if you've been on Twitter or the internet, you've definitely seen it. It's Nano Banana. Their incredible new image model that has huge breakthroughs in character consistency, image quality. And today on unsupervised learning, I got to talk to two folks who are really driving that at Google, Nicole and Oliver. We hit on a bunch of different things, including their favorite use cases for the models and how folks are actually using Nano Banana today. We talked about some of the challenges in building our products around really good models and how they thought about solving the blank canvas problem, as well as interacting with other image editing products. And we hit on the future of image models, what's next, frontiers, and how it'll interact with video models too. This was just a fascinating conversation with what I'd say is the main character of the AI ecosystem right now.

• 0:44Speaker 00

  And so without further ado, please enjoy this conversation with Nicole and Oliver. Nicole and Oliver, thanks so much for coming on the show. Really appreciate it. I've been looking forward to this. I feel like you have, you have pretty much taken over the entirety of my Twitter feed with Nano Banana, as well as like any spare moment of free time I have. You know, I figured a lot of things we'll dive into today, but maybe just to start, you obviously were sitting with this incredible model and product and before it was released into the world, and I guess it was maybe anonymously released into the world first, but there was folks, you know, you were some of the first to play around with it. I'm curious, like, some of the use cases you thought would be most prevalent or got you most excited, and then what actually has happened now, having released this out into the world?

• 1:22Speaker 01

  Oliver has seen a lot of pictures of my face in various iterations. The most excited, exciting thing for me was character consistency and seeing yourself in new scenarios. So I literally, there are slide decks full of my face in, on like, wanted posters and as an archaeologist and all my childhood dream professions, basically. We've now created an eval set with, that has, you know, my face in it and like other, other people on the team that we just kind of eyeball when we develop a new model.

• 1:53Speaker 00

  That's like the ultimate honor in AI world to have your own, like your own eval set.

• 1:57Speaker 01

  I'm very, I'm very excited. So I was, I was really excited about the character consistency capabilities because it just felt like it's, it's giving people a new way to imagine themselves in a way that just wasn't very easy to do before. And that is one of the things that people ended up, you know, being really excited about. We're, we're seeing people turning themselves into figurines, which is like a very popular use case. The one that surprised me, but it shouldn't have probably, is people colorizing old photos has been a really emotional use case that comes up for people of like, I can now see what I actually looked like as a baby, or I can see what my parents actually looked like from these black and white photographs. So that's been really fun.

• 2:37Speaker 00

  I mean, I'm sure alongside seeing all the ways people use these things, one of the joys, I'm sure, of having a very popular product is, and I've seen this on Twitter, you must get like a million feature requests, right? And everyone wants, you know, these models to do this thing or that thing. What are like some of the most common things people want? And how do you think about, I guess, like the next milestones or frontiers for these types of products and models?

• 2:55Speaker 01

  The things we get most on Twitter are higher resolution. So a lot of pro feature requests. Currently are 1K resolution, so higher than that. We get a lot of requests for transparency, because it's also a really popular pro use case. And those are probably the two biggest ones that I've seen, plus better text rendering.

• 3:17Speaker 00

  You know, obviously, I think character consistency for so long was like this big thing that needed to be solved, and feels like you guys have done a great job on that. What are like the next frontiers of image model improvement in your minds?

• 3:27Speaker 02

  Yeah, so I think like, one of the things that was most exciting to me about this model is that you can start to give it harder questions. So instead of having to kind of like define every aspect of the image you want to see, you can really ask kind of like for help, like you would from a language model. So for example, people are using this for like, I would like to redecorate my room, but I don't know how. Like, give me some ideas. And the model is able to come up with like, oh, you know, this might be a plausible suggestion, or like, given the color scheme, like these things would go together well. So the thing to me that's really interesting is like, is really like getting the world knowledge from the language model, so that we can make these images truly helpful for people, and maybe show things that like they hadn't thought of,

• 4:09Speaker 02

  or like, you know, they don't know what the answer is going to be, or maybe there's an information-seeking request, like, I would like to know how this thing works, and like, being able to show an image of like, oh, this is how it works. Like, I think that's a really important use case for these models going forward.

• 4:24Speaker 00

  Where are we in solving that right now?

• 4:26Speaker 02

  Well, you know, of course, sort of that specific example, aesthetics is always a bit tricky. Because it requires pretty deep personalization to be able to give useful information. And I think personalization is an area that's that's still being improved also on the tech side as well. So I think we're quite a ways away from really being able to like to tell exactly what the user is asking. But I think like with a little bit of clarification and being able to sort of have a conversation with the model, which is another thing I'm really excited about, that's models, you can talk to it like in a thread. Then you can kind of refine down to get the to get the the images you're trying to get in the end.

• 4:58Speaker 00

  Do you think personalization will happen at like just the prompt layer? Like ultimately you clarify enough that you can then feed the model enough context to personalize? Or is it like, will folks have different aesthetic models?

• 5:08Speaker 02

  I think it'll be more at the prompt layer. I think we'll have like, you know, given things the users told you about themselves. For example, we'll be able to make much more informed decisions. So at least I hope it's that way, you know, of course, everybody having their own model and serving that sounds like kind of a nightmare, but maybe that's the direction we're moving.

• 5:27Speaker 01

  But but I do think that you will have very different aesthetics, right? And I I do think that there's probably some level of personalization that has to happen at that level. Because even, you know, you you see it probably when you go to the shopping tab on Google right now, right? Like you're you're looking for a sweater and then you get a bunch of recommendations and then you actually wanted to hone into your own aesthetic and like be able to pull in from your closet to see what other things work with it. So I I hope that a lot of that can happen in the context window of that model, right? Because we should be able to feed the model to images of the things that you have in your closet and then try to find something that actually goes with that. And I'm really excited about that, so I hope we can do it there. Maybe you will need some level of aesthetic control kind of beyond that that like is with the model.

• 6:10Speaker 01

  But I suspect that might happen more in the pro workflows again.

• 6:13Speaker 00

  Let's start by, obviously, in the LM world, it feels like a lot of, or even even in the image world, a lot of the decisions that are made in data that's used for pre-training like actually really does impact the, you know, end capabilities or aesthetics of the models. And so I wonder whether we'll we'll have a world where there's one model that's just prompted incredibly well for like the entire variety of image use cases? Or actually, you know, you'll have your cool futuristic one and your, you know, I don't know. You're smarter than me in what the categories are, but it'll be interesting to see how it all all plays out.

• 6:40Speaker 01

  I think we continue to be oppressed with just how. Broad the range of use cases are that you can support with the off the shelf model. But I think one thing to your point is like you get a lot of mileage on it on some of these like consumer facing use cases where you know you're trying to just sketch out what something would look like in your room etc. But once you get again into the more advanced capabilities, like you do need to bring in other tools to actually make it a final product and have it be useful in a workflow for you know, marketing or design and those kinds of capabilities.

• 7:11Speaker 00

  Well, I'm sure folks are really curious like what made these models so much better?

• 7:15Speaker 01

  A lot of special sauce.

• 7:18Speaker 02

  I think it's you know, it's one of these boring cases where it's not one specific thing. It's like figuring out all the details, right? And and really tuning the recipe and and having a team that's been working on the problem for quite a while. I think we were even kind of a little bit taken aback by the degree of success that the model had. Like, I think, you know, we knew we had a cool model and we were really excited to get it out there. But you know, the we saw when we released on LM Arena, we had, you know, not just the Elo scores were high and that's great. I think, you know, it's a good sign the model's useful. But the real the real metric for me was that like we actually had so many users going to LM Arena in order to use the model.

• 7:56Speaker 00

  Yeah.

• 7:57Speaker 02

  That we had to like keep increasing the like the amount of queries per second that we can support. And we definitely didn't expect that. And I think that was really the first sign that like, oh, this this is something that's really special. And and there's a lot of people that have use for this model.

• 8:10Speaker 00

  I think that's one of the most fun parts of this this whole ecosystem where it's like you obviously have some sense of the model, building it yourselves, but it's only in releasing it in the wild that you really understand like the the extent of the power it has. And clearly, this is this has struck a chord. Obviously, some of the reasoning capabilities of the models are really driven by improvements in LLMs themselves, right? And I think maybe just contextualize a little bit, like, how much do image models benefit from these improvements in LLMs? And do you kind of expect that that to continue as as LM progress continues?

• 8:36Speaker 02

  Yeah, I mean, definitely, they benefit, you know, almost 100% from the rural knowledge of language models. I mean, the thing about Gemini 2.5 flash image, which is the name of the model, is that it is a Gemini.

• 8:48Speaker 00

  A little more fun of a name, but yes, that too.

• 8:51Speaker 01

  That's a bit easier to say.

• 8:54Speaker 02

  I do wonder how much of our success has to do with just the fact that people like saying the name Nano Banana. But yeah, you know, it is a Gemini model. So like we're, you know, you can, you can talk to this model like you do Gemini and it understands all the things that Gemini understands. And that's, and that's actually, I think, was a really important kind of step function and utility for these models is to, is to integrate them with the language models. Yeah.

• 9:17Speaker 01

  Yeah. And if you, you probably remember this, but it used to be the case that, you know, two, three years ago, you had to be very, very specific in what you asked the models to do. It was like a cat sitting on a table with this exact background and these are the colors. And now you don't have to do that anymore. And a lot of it is because the language models have just gotten so much better.

• 9:34Speaker 00

  Yeah. It's not like magic prompt transformation that's happening on the back end. I feel like that was the hack back in the day. You'd like have a sentence and they would turn it into like a 10-sentence prompt that made sure it was specific enough to get things right. But now it feels like the models are sophisticated enough to, to get it, which is exciting. One thing I'm curious about from the product perspective is, you know, I feel like you have so many different kinds of people that use these products, right? You have the folks that I guess are flocking to L M Arena the second the product's out. They know what they want to do with it. They are, are experts in playing with it. And you have a lot of folks that are just probably Gemini users and, you know, have this blank canvas problem. They're like, what do we, what exactly do we do with the problem, do with the product? I'm curious, how did you think about building product for like those two different types of users?

• 10:14Speaker 01

  There's so much more we can do. And to, to your point, we're, we're doing a lot, right? I think the L M Arena crowd and even just developers, they're, they're really sophisticated. They know how to use these tools. They came up with use cases that are new to us. We've seen people do, you know, turn objects into holograms within a photo. And that's not something that we trained for, expected to be good at, but apparently the model's very good at it. For consumers, having, making it really, really easy is super important. And so even now, when you go into the Gemini app, people will notice there's banana emojis everywhere. And we did that because we realized that people were having a really hard time finding the banana.

• 10:52Speaker 01

  When they hear about it and then they go into the app to try it, but there was no obvious place to actually do it. And we've done a lot of work to, you know, pre-seed some of these use cases with creators that we partner with. And to kind of just like put examples out there that link directly then back to the Gemini app. And so then the prompt pre-populates. I think there's a lot more that we can do on kind of just the zero state problem of like giving you visual guidance. There's probably a lot more that we could do with making gestures a thing when you're trying to edit images so that it's not all just prompt based. And sometimes when you do want something very specific, you still need a fairly long prompt. And that's just not a natural mode of operating for most consumers.

• 11:36Speaker 01

  And so I try to give it the parent test. Like if my parents can use it, then it's probably good enough. And I don't think we're quite there. So I think we have a long way to go, but a lot of it does come down to just showing, not telling, giving people the examples that they can easily reproduce, like making it really easy to share things. And so it's to Oliver's point, a lot of things that it's not kind of one magical answer.

• 12:00Speaker 02

  The other thing we've noticed is that this is that sort of social sharing is a really important part of like the blank slate problem. Like people see things that other people are doing. And because the model is sort of personalizable by default, you know, you can try it on yourself or your friends or your pets. Like there's this like very easy way for people to see something and be like, I'll try, I'll try that for myself and see how it works. That's like a really big way that we, that like this model is being spread around.

• 12:31Speaker 00

  Today, the way you interact with these models is all, you know, via text. What other kind of like design interfaces get you excited about ways people could interact with them, you know, longer term?

• 12:40Speaker 01

  I think we're just scratching the surface on what's possible here. Ultimately, I envision all of these modalities kind of blending together and then having some sort of an interface that like picks the right one for whatever it is that you're actually trying to do. And I think even now, kind of moving more towards a place where the LLMs don't just output text, but they can output images and visual explainers when it's actually relevant to the user query. So I think there's a lot of potential in voice, especially. It's a very natural way for people to interact. I don't think anyone has really cracked like how that could actually show up in a user interface. We're still very much, you know, typing the thing that you want to put in. And so some combination of that plus again gestures.

• 13:24Speaker 01

  So if you wanna erase, you know, an object from an image, you should be able to just erase it like you would on a scratch pad. Again, like, how do you do that? And how do you seamlessly transition between those modalities based on what the task is? Is something I'm really excited about. And I think there's a lot of headroom in figuring out, like, what does that actually look like?

• 13:44Speaker 00

  Yeah, what feel like the limitations to that voice UI today? I mean, it makes it, you know, I could totally imagine talking to these images.

• 13:50Speaker 01

  I, I, I think some of it might also just be kind of prioritizing it because, you know, we're, we're still pushing on a lot of model capabilities. Voice is obviously getting really good also in the last couple of years. So I think we could, I think we're gonna start seeing some people take this on and think about and probably, you know, will do some of this work too. I'm trying to think about what this could look like. I think part of the problem is actually the, how do you detect the intent? And then how do you switch different modes based on the user intent? And what they're actually trying to accomplish? Because it's not obvious. And you could also end up with a surface that again is like a blank slate. And then how do you actually show the user what's possible? Because it's one of the other things that's really challenging. I think we see users come into chatbots and they just expect the chatbot to be able to do anything, right?

• 14:34Speaker 01

  Because you can just talk to it like you would talk to a human. And it's actually very hard to explain the limitations. It's very hard to show people what you can do when the tool can now do so much. And so I think it's part of it is just figuring out how do you scope the problem? How do you show people what's possible in a UI? That can ultimately help them accomplish almost anything.

• 14:53Speaker 00

  Totally. And it feels like if you teach them at any moment in time what the chatbot can and can't do, that changes three months later at a maximum. And so you're always having to reteach those, those very same capabilities, which I think is a really interesting product challenge. A lot of folks have, you know, in consumer and enterprise products. You alluded to it earlier around, around evals. I guess you have your own eval data set, Nicole, of yourself. But I'm curious, like, image models generally, like, how do, what, what do model evals look like for this? You know, besides just putting things out in L M arena? And kind of any learnings since you got started on, on tracking? What makes these models better?

• 15:26Speaker 02

  Yeah, um, so, well, one of the nice things about the, the language models and the vision language models getting better is that there is start to be like a feedback loop where we can use the intelligence from the language model to help with evaluating its own generations. And this kind of like is a virtuous circle where we can continue to improve both of these dimensions. So that's pretty exciting. But I think ultimately, people are the arbiters of the images that they're trying to create for themselves. So that's why I think cases like Elmerina, where users are entering their own prompts, like this is the best way to evaluate a model.

• 16:03Speaker 01

  So taste also matters a lot. Oliver won't say this because he's being modest, but like, he's one of the people on the team who just has a really good eye for what these things look like and, you know, whether or not things are working, where are the flaws in the images. So we do have a couple of people on the team like that who will do a lot of this initial eyeballing, as we call it, very technical term, on what the outputs from a model that, you know, training actually look like. And then that still plays a pretty significant role. And then, you know, to your question, we do get a lot of feedback from people, including on X, on what's working, what's not working. And then we do try to adapt our evals to capture the things that are both working, so we don't regress on them. But then also the things that are not working, that the community wants us to push on. So keep sending us feedback.

• 16:48Speaker 00

  It feels in some ways like a much harder problem than, you know, some of these LLMs that are working on, you know, like a legal use case where there's some, there's some like relatively right answer. You get the ones where the models deviate and like there's kind of this pure eval data set that really works. You know, because images are so subjective, it almost feels hard to know like what, I'm sure they're, you know, did the character stay consistent in generation to generation? There's probably some things that carry over. But the subjectivity of it must make it really challenging, right, to figure out what to hill climb on. Well, I do have to ask, like, what is the story behind the name?

• 17:18Speaker 01

  Uh, there's a PM on our team. Her name is Nana. She was up at 2:30 in the morning working on this release. And she came up with the name then. And it stuck because it was fun. And it even now stuck as a semi-official name because, to your point, Gemini 2.5 flash image is a bit harder to say.

• 17:39Speaker 00

  Yeah, I mean, clearly very successful when you have the CEO of Google tweeting out bananas like the name has, uh, the name has, has, has taken, uh, has taken charge.

• 17:48Speaker 01

  The brand, the branding takeaway, I guess, is your, your name should have an appropriate emoji to go with it because it makes it stickier.

• 17:56Speaker 00

  Yeah, I feel like Hugging Face was the OG of figuring that out in the AI world with their emoji. But it seems like we're not too far away from stock tickers just being emojis for companies. I guess switching gears, another interesting part of, we were alluding to earlier, this idea that you have really sophisticated users that use this stuff and then people that come to the blank screen figuring out what to do. What have you seen your most sophisticated users do?

• 18:20Speaker 02

  I have a favorite sophisticated use case and I used to work most of my career in video. So I really like video tools and creation tools. And so we found that having this, so Nano Banana ends up being a really useful way to make AI-generated videos in combination with video models like V O three. Because you can idea quicker, you can sort of plan shots out. And interestingly enough, that's also how people plan movies, right? They start with the shot boards where you have the story and the views. And so people have started to do that and build more coherent, longer-form video.

• 18:56Speaker 01

  I've been really impressed with people using this in real, say, architecture workflows, where you can go from a blueprint to something that looks like a 3D model, but it's not actually a 3D model, to something that then starts to look like a design where you can iterate on it. And it shortens a lot of the part of the pipeline that is very tedious. And then it allows people to actually work on the parts of the pipeline that are creative and fun and that they actually enjoy doing. And so I think I was a little bit surprised that it worked as well as it did out of the box.

• 19:28Speaker 00

  It's like all these VOD coding image use cases of getting the first basics up and running for across these different spaces.

• 19:36Speaker 01

  And the other thing is you could vibe code website UI now. And it always felt a little bit abrupt to me that we went from a prompt to a website being coded. And it feels like there's a missing step where you can actually iterate on the design and change things and do that really quickly. And so now we're at a point where you can actually start to do those things. And then you can code it once you're happy with the way that it looks.

• 19:57Speaker 00

  Yeah. Do you think that's the future workflow? I mean, it makes a tremendous amount of sense, like why execute, why spend the tokens on all the code? If it's going to come out and you're like, no, that aesthetic is completely off from what I wanted. Or that actually looks nothing like what I wanted.

• 20:09Speaker 01

  It also just feels more fun to Oliver's point is how people have actually been using these, you know, the previous technologies in their existing workflows. It feels very natural to adapt, but I think just, you know, the LLMs are progressing so quickly that you can go straight from a prompt to a website and it's super impressive. I think it's really fun for people to iterate kind of in that in-between space to like make it, make sure that it, you know, adheres to your aesthetic.

• 20:34Speaker 00

  I mean, you obviously offer the model as well as an API. So obviously there's going to be all sorts of interfaces and use cases on top of the model. How do you think about the ones that make sense to live in kind of a general, you know, chat tool like Gemini and then the ones you want to power, you know, through other product experiences?

• 20:50Speaker 01

  I think they're just very different. So we do see people going to Gemini to do a lot of this quick iteration, right? So if you're, we've had folks on the team already, they're trying to redesign their garden. And so they will go to the Gemini app to try to imagine what it could look like. And then they may actually go and, you know, work with a landscape architect to kind of mold it to their vision, collaborate and like take that idea further, right? So it's really kind of that first step in your ideation process. It's very rarely like the final product of the thing that you're actually trying to do, right? Versus for a lot of these pros, for developers, you are building much more sophisticated tools that chain multiple models at this point, right?

• 21:33Speaker 01

  And so I think it is a very, it's a much more sophisticated, much more complex kind of multi-tool journey. And the chatbots are really great at kind of getting you started, giving you inspiration, supporting a lot of these use cases that are just really fun and shareable. So you can share things with family and friends. And I kind of see that probably sticking around for a bit because the more sophisticated users who have like more demanding workflows will always want to go back to something that is just like a more visual forward UI than, you know, a chatbot.

• 22:03Speaker 00

  And where does like the editing workflow end up fitting into this? Obviously, it seems like first generation and getting started makes tons of sense to doing these tools. I think you guys power like the APIs available in Adobe and some of these other tools. Oliver, I know you used to work there, like where the classic editing gets done. Do you think those editing workflows end up looking incredibly different? Obviously, you can do some light editing in, you know, using just the core models today. Or do you think to bring something from 95% done to 100% done, it's still going to look similar to kind of the classic editing workflows of the past?

• 22:33Speaker 02

  I think it depends a lot on the user. So some users will have like a very, very clear idea of what they want, like to the pixel. And for this type of use case, like I think you really have to integrate this with the existing tools and all the existing, like, you know, things you can do in Adobe products, for example. And some users are really looking for inspiration and have a bit less like restrictive requirements. And for these users, then just being able to get, you know, quick ideas out in a chatbot is totally valid. So, you know, both of them are important applications of the model.

• 23:06Speaker 01

  And to the pixel level control, one thing that I learned two days ago is that, you know, when you're, when you're creating ads, right, for different products, different brands, things including where a model is looking in the ad end up mattering a lot because where your gaze is at kind of influences what message it communicates. And that's just a level of control that is very, very hard to get out of a chatbot. And I think for those kinds of users, for those kinds of use cases, you will continue to need specialized tools and just like a very precise level of control.

• 23:39Speaker 02

  I think it comes down to things that can be specified by language versus not specified. Because like, you know, language is a great way to communicate high level ideas. But like, if you're trying to move something left by three pixels, it's like, it's not the most elegant way to do that. So I do think there is a place for both.

• 23:53Speaker 00

  I guess if you watch like a real artist or creator like duel their workflows end to end, like they probably would be hard for them to dictate exactly what they're doing. There's like some sort of just like feel as you're doing that. There's lots of folks within Google that are excited about using this in different ways. Like, what are some of the ways you get excited about, you know, an image model proliferating throughout, you know, the many different things that Google does?

• 24:12Speaker 01

  I I I think there's a lot of, I mean, on the creative side, you can imagine that photos is a really exciting place to be doing any kind of photo editing, right? Because your your library is right there. And so like turning a family photo into a birthday card, which is a use case that I have, you know, a couple of times a year. You get kind of being able to do it right there, I think is really exciting. And then to Oliver's point at the beginning, I think a lot of this kind of factuality, whatever we want to call it, set of use cases is really exciting on a lot of different Google product surfaces, where if you want the model to explain photosynthesis to you and you're a five-year-old and do it in a way that, you know, it probably doesn't exist anywhere on the internet visually, we should be able to do that.

• 24:56Speaker 01

  And so I think that will just open up again like a whole set of use cases and opportunities for people to learn in a way that is customized to them and is very visual, because a lot of people are visual learners.

• 25:07Speaker 02

  Another one I think is really cool is Workspace as well, like, you know, PowerPoint slides and Google Slides, like being able to get to the point where people can have like really compelling presentations that don't all look the same, like same set of Google points.

• 25:20Speaker 00

  As someone who started their career in consulting, that would be phenomenal.

• 25:24Speaker 01

  Me too.

• 25:24Speaker 00

  If we get that, yeah. But lord knows, too much time in formatting all this stuff.

• 25:30Speaker 01

  Well, and you used to start from just storyboarding what the slides look like on a whiteboard, right? With like the exact headline and then, oh, I need this this chart to represent this data set on the left-hand side of the slide. Like being able to feed that to an LLM and like take a lot of that legwork off of you. I'm very excited about this part.

• 25:50Speaker 02

  To snapshot the whiteboard.

• 25:52Speaker 00

  Yeah.

• 25:52Speaker 01

  Snapshot the whiteboard and then you just get it. That sounds great.

• 25:55Speaker 00

  I guess maybe taking a step back for just like, you know, the broader kind of image model landscape. I feel like it's just been a torn of progress since maybe Stable Diffusion and then Mid Journey. Maybe, Oliver, just contextualize, like, as you think about the last two, three years in image generation models, like what have been the major milestones? Like, how would you kind of characterize the path we've been on? And the extent to which things have changed?

• 26:15Speaker 02

  Yeah, my God, it's been rocket ship, definitely. I mean, I I worked in this space, you know, back when GANs were the de facto method for generating images. And we were amazed by what GANs could do. But they could really only generate in very narrow distributions, basically like, okay, we had images of people that looked pretty good, but you could just generate these like front-facing images of people. And I think that We started to see models come out that could really generalize and be like fully text-driven, control over anything, but they started out really small and really noisy. But I think a lot of us at that point were like, oh, wow, this is going to change everything. And so we all started to like, you know, pour our energy into this area, but none of us could predicted the rate which it improved in.

• 26:57Speaker 02

  I think I attribute this to there being quite a few labs of absolutely great people working on these problems. And then this kind of like friendly competition between all the labs, where we see another group come out with this amazing model. You know, for a long time, you know, Mid Journey just had, just like far ahead of everyone. It just looked amazing. And this was really motivating. People were like, man, how did they do that? How does it, how does it look so good? And then, of course, Stable Diffusion coming out as an open source model really showed a little bit of like, what is the size of the developer community? And like, how many people out there actually want to build on these models and make things? So that was like clearly another launching point. But, but since then, it's been, it's just been really fun working in this area. Although, maybe, maybe a little bit, a little depressing sometimes,

• 27:41Speaker 02

  because I think people's, not, not only have the models gotten so much better, but people's expectations have gotten so much better, too. So you now see people like complaining about these like small little problems, where you're like, oh, man, don't you know how hard we worked to make this model? And like, just a year ago, we were looking at these images that don't even look realistic at all. And like, everyone was amazed. So, you know, people, people who have a remarkable ability to get jaded about new technology.

• 28:03Speaker 00

  It does feel like, yeah, I mean, with LLMs, right? So we have this incredibly powerful and technology that if you told us in 2017, we had, we absolutely dumbfounded by. And still, we always complain about all the places where they, where they have shortcomings. It's a funny part of, of human nature. In retrospect, like, what allowed Mid Journey to get out to such a strong lead in the space? I feel like for a while, they were like the brand, you know, and clearly the place everyone was going.

• 28:26Speaker 02

  Yeah, I mean, Mid Journey figured out a little bit before everybody else, how to do the post-training and specifically how to do post-training of these models in order to give you sort of stylistic and artistic imagery. And that's really been their bread and butter, is focusing on like, how do we, how do we give control over style? And how do we make sure that like, no matter what you generate, it's going to look amazing? And at the beginning, that was really important, because the, you know, when you can sort of narrow the domain of images you generate to just the good-looking ones, you can actually do a better job generating those images as well. So, so starting with that sort of small. Section of really, really good looking images was like a, was a great stage at that point. And then I think the, you know, all the models, Midjourney included, but all the, all the other models like, you know, Flux and GPT

• 29:11Speaker 02

  like have expanded since then and can now generate much broader categories of images while still kind of retaining the quality.

• 29:18Speaker 00

  What, what allowed that ability to generate like a broader set of images and not have to like hand pick just the, just the perfect ones to put in?

• 29:24Speaker 02

  Well, a bunch of things. We all figured out the details and, and in particular, you know, what the data should look like. And at the same time, you just have this kind of natural scaling curve where all the models get bigger, everyone's compute gets bigger. So things that like weren't possible to do back then start working now because you can do them at a bigger scale.

• 29:42Speaker 00

  Yeah. And then you kind of alluded to it earlier. We've made so much progress in image models and I can't tell whether there's like only 10% more to go or there's actually like, we're going to look back three years from now and be like, that was funny that we thought we had good image models. How do you think about that? And like, what, you know, I mean, the generations are pretty good. Like I'm trying to even conceptualize like what the next, you know, 10x improvement even would be.

• 30:02Speaker 02

  Yeah, I, I'm completely in the latter camp there. I think we have, we have much, much farther to go on. Just, just in terms of image quality, like let's not even think about all the other applications, but like just the image quality, I think has, has a long room to improve. And, and really I think that's going to be in the expressibility of the models. So like, we know we can generate some things basically perfectly, like no one could tell that it's a generated image and it's indistinguishable from, from, from reality. But like, as soon as you start going outside the realm of like very common things that people try to generate, then the quality degrades really quickly. And you can look at like the prompts that require more imagination, more like composing multiple concepts together. Like these break down really fast. So I think like the, the models in the future, while the, the best case images today

• 30:46Speaker 02

  may look as good as the best case images in a few years, the worst case images today will be significantly worse. So we'll just make the models more useful and sort of broader in applications. And, you know, we found the broader we make them, the more use cases people discover and the more useful they become.

• 31:02Speaker 00

  I'm curious, as you think about like the evolution of this broader, you know, image model space, you know, if I were to compare to the LLM world, you know, there's really, I don't know. I, I could say it feels like it's you guys, OpenAI. And then like a host of other folks maybe doing, you know, slightly behind different things. Do you think the image model landscape plays out similarly?

• 31:19Speaker 02

  Yeah, I think it's a good question. So until now, I think images have been an area where it's been possible for smaller teams to develop really state of the art image models. I mean, we've seen this. There's, there's a few small labs that have just amazing models.

• 31:34Speaker 02

  I hope this could maintain because I like the, the fact that there's small labs doing it, but like, um, you know, I said before that the, that the world knowledge of these models and making them more useful, I think is, is something that really benefits at scale. And especially with the scale of the language model. So, um, you know, I suspect we might see the same groups that can do the large language model trainings are the ones that have the, the image models that have all the world knowledge as well.

• 32:00Speaker 02

  We're seeing similar trends with the, the big Chinese labs also coming out with like amazing models and the same as language models. So I think we'll see that as like a major player in the, the image space as well.

• 32:10Speaker 00

  Do you think it's that big a disadvantage for an image model to use like the best open source model versus some of like L M versus the, the, the cutting edge closed source ones?

• 32:18Speaker 02

  This is a great question. I feel like this is going to depend a little bit on the future of open source models, which is something that, you know, changes quite frequently. I think maybe a year ago, uh, it would have seemed like that was a very safe solution and now, um, maybe it, the, it's less so, but I don't know the, I don't know what the future of, of open source models is. It's definitely a possibility and, and that could sustain a lot of sort of small, um, smaller labs training, good image models, definitely.

• 32:43Speaker 00

  Well, Oliver, I'd like to just shamelessly ask you a question because you were, you mentioned you worked in video for a while and I'm always trying to just make sure I grew up like the relationship between image models and video models. Obviously, uh, your colleagues have had tremendous advances, uh, on the video side, uh, with, with their recent video model releases. Like, are these separate efforts? Like, do they build off each other? Like, how are the image and video worlds interacting these days?

• 33:04Speaker 02

  They're very closely related. I mean, like, I think in the future. We have, we're moving, everyone is moving towards Omni models, right? Which are sort of models that do everything. And for many reasons, these models have advantages and may end up winning in the long run. I don't know. But I would say that a lot of the techniques that we learn for image generation also sort of went into the video generation models and vice versa. And so like, this is kind of one of the reasons why video generation took off as well is because we, as a community, we kind of like learned how to do this, solve this problem. So I think that, I would say, very close friends and we share a lot. And maybe we'll live together someday.

• 33:49Speaker 00

  Yeah, I guess the techniques, you mean, the techniques to do a lot of this stuff end up looking similar across, across the models.

• 33:56Speaker 02

  Yeah, exactly.

• 33:57Speaker 01

  And even in the workflows, right? A lot of people do end up using these models in very complementary ways. So if you're a movie maker and you're making a movie, a lot of the initial iteration, first off, it happens in the LLM land. And then iterating in the frame or image space makes a lot of sense because it's faster and cheaper to do that. And then you actually move on to video. So I think even from just like workflows and usability perspective, there's a lot of complementarity that happens between these models. And a lot of the use cases and problems that we have to solve are similar. So consistency, characters, objects, scenes, right? Exists in image land, exists in video, gets a little more complex in video because you have to do it over multiple frames.

• 34:35Speaker 00

  Yeah, what do you feel like the next problems to solve on the video side?

• 34:39Speaker 02

  Yeah, I think that, that having the same kind of control over the models that we get with the latest generation of image models is something that obviously will be really impactful in the video space. So that's, that's one thing to look out for. I think, you know, video teams are also working on just improving the sort of resolution and the consistency over time. And of course, like being able to generate the same character across many scenes is, whenever you look at like. What people want, that's always sort of top on the list. So I would say moving towards longer form content with more coherency. Definitely a future direction.

• 35:16Speaker 00

  And it sounds like you can figure out some of that stuff in the image context. And a lot of that ends up being transferable to the to the video side, which is cool. It's been a fascinating conversation. We always like to end our interview with a standard set of a quickfire questions where I stuff some overly broad questions into the into the end of the interview. And so maybe to start, would love your thoughts on one thing that's overhyped and one thing that's underhyped in the broader AI world today.

• 35:40Speaker 01

  I think overhyped is this idea that you can just go from like one short prompt to something that's like a production ready solution that you can ship anywhere. I do think that there's a lot of iteration that needs to happen. And if you even if you look at a lot of the stuff that people post on socials, there's actually still a lot of work that goes into, you know, producing the final product and sharing it. So I think that's a little bit overhyped. Underhyped, I think it's the it's what are the UIs of the future? We already talked about this, but like, how does this actually all come together? And how do you make it easier for people to use these models, show them what's possible? And then make it useful for specific workflows?

• 36:19Speaker 00

  Have you seen any products that you're like, oh, that's a really interesting new take on a on a UI for the AI age?

• 36:24Speaker 01

  Not yet. I'm still I'm still waiting. I'm still waiting.

• 36:28Speaker 02

  I like node based interfaces, but I don't think I'm like representative of everybody.

• 36:32Speaker 01

  It's a very specialized audience for this.

• 36:35Speaker 00

  Yeah, well, hey, well, everyone gets their their own UI, right? In the future, anyway. So maybe that that it is, it will be an era of specialization. Do you think image model progress is going to be more or lesser the same next year as it's been this year?

• 36:48Speaker 01

  Oh, more.

• 36:49Speaker 02

  More. Yeah, there's more smart people working on the problem and putting more resources into it. I think we're going to see in all axes that the change will accelerate. Yeah.

• 37:02Speaker 00

  Well, you've obviously taken over the internet with everyone's focus on Nano Banana. Are there other things happening in the AI image world that like you're paying attention to that you don't think enough people are thinking about or paying attention to?

• 37:13Speaker 01

  I do think it's this factuality dimension. Even when we look at the things that people are doing with Nano Banana, you know, people have been making infographics or taking a picture of Niagara Falls and asking the models to annotate it. And it sort of does a good job as a demo. It looks, you know, it looks fine, but when you, when you hone in, the text is a little bit garbled. It's not factual. It repeats information. So I think that is a frontier is something that not a ton of people are paying attention to, but I do expect to get better.

• 37:46Speaker 02

  There's a real analogy with the text language models as well because, you know, when, when GPT one and two came out, people found like, oh, these are, these are kind of cool. I can have them write haikus and kind of do creative tasks where there's like a broader range of acceptable answers. But now people don't use language models for this as much at all. Now they use them for like, you know, for information seeking and also for like kind of conversation and like, you know, personal connections and like having someone to talk to. So like, I think that we could probably find analogies for all of these in the image space as well. Where we went from like creative tools to maybe like information seeking tools. And then also maybe like people will end up talking to a video model, you know, when they need to talk to someone. And this is something that we could see in the future.

• 38:29Speaker 01

  And the models, I think, should also get a lot more proactive because, because right now you always have to ask, right? I want, I want, I want an image, but if, if the query warrants an image, right? And we're used to this from search actually, right? Like you're used to going to search, asking for a thing. And sometimes it comes with text, sometimes it comes with text and an image. Sometimes image is the right answer. And so I also expect these models to just get a lot more proactive and smarter about how they talk to you and how they leverage these different modalities based on what you're asking about.

• 38:57Speaker 00

  I love that you'll be able to seamlessly go back and forth. And then already your earlier point about, you know, it seems like so much of the improvement here is really around reliability. And, you know, there's in the early days of LLMs, there were glimpses of, oh, wow, that's, that's really cool. But it was not nearly consistent enough to use for, for more kind of work use cases. And it seems like a similar pattern to be followed with, with image models as well. The most important question of this interview, definitely. What is your single favorite piece of content that you've seen made using NANABANAN?

• 39:25Speaker 02

  Okay, so my favorite piece of content is going to be like playing with the model with my kids. Where we like put them in in like, you know, funny locations or and like um make their stuffies come to life and these kinds of things because they're they're they're so personal and it's like something I can do with them and they love it and so like this kind of thing I think is the the most yeah definitely the most valuable for me.

• 39:49Speaker 00

  That's amazing.

• 39:49Speaker 01

  I don't think I can top that, yeah.

• 39:51Speaker 00

  Yeah, people love like the LM, you know, bedtime stories for their kids and now we get to move into image and soon you'll have movies that you can put your your kids and and do and and whatnot and that'll be that it's a really exciting uh exciting future. Um, well, I want to make sure to leave the mic to you folks. Uh, besides like clicking on, you know, various bananas uh around the internet and Gemini. Where can folks go to learn more about you, uh, the product, uh, anything you want to point folks to? The uh, the floor is yours.

• 40:17Speaker 01

  Please do go click on the bananas. Um, Gemini app, AI studio, you know where to find us. Um, and then we're both on X, so do you um, do you find us, um, give us feedback?

• 40:29Speaker 02

  Yeah, if you find something that doesn't work or or sucks, definitely like tag me in it. It's very helpful.

• 40:35Speaker 00

  Amazing, well, we'll we'll hope people contribute uh to to that and we'll we'll link to you as well. But thank you both so much. This is a, this is a ton of fun.

• 40:42Speaker 01

  Thank you.