Inside Nano Banana 🍌 and the Future of Vision-Language Models with Oliver Wang - #748

31:24Like what's most exciting for you in the broader space? Well, I think that all the teams that are working in this space are full of really brilliant people. And every time there's a new model that comes out, I'm always amazed to see what it can do. So I think I still have the wonder that like we all felt when Stable Diffusion 3 came out with every future release, maybe especially because I've worked on these models for so long. So I kind of know how hard it is to do all that stuff. But I think we're, I don't think we're anywhere near the ceiling for quality. Like I think just, you know, just improving the quality of these models will make them more and more useful.

32:02I mean, I think we get to the point now where, you know, if you know what to do, you can kind of cherry pick a really good image out of most of the top models. And it'll be, it'll be like a very high quality, you know, kind of exactly what you want to get. But if you're just exploring the models, like as soon as you leave this space, the comfortable space of things that work really well, I think you often can fall into like, oh yeah, this doesn't work or this doesn't look right or it's not what I was trying to create. So I think that we still have a number of years ahead of us for just making these models better.

32:33And every time they're made better, they'll become more useful. And that's like totally ignoring the whole world knowledge aspect, which I think is coming, as I mentioned, with all the information-seeking use cases. the other thing that I think is of course really interesting is thinking so we know the models all models now think in text traces and of course when models can generate images they can think in images and I think that we as a community don't really know yet what is the potential of being able to do this but I think we'll find out really soon like what happens when these models can think and create images as well as text in their thinking process Yeah, with kind of this focus on quality being the big challenge, like what do you think is required to get there?

33:19What do different teams need to be focused on? Or what would you expect to see different teams focusing on as kind of vectors for driving model quality? Yeah, it's an interesting question. So I think there's, you know, one thing is, of course, scale. So we're seeing what happens at scale. I think that the image community has been behind the language models for scaling up their models and actually figuring out how to do it. I mean, I think a lot of people have tried, but it's very hard to sort of successfully scale up an image model to be the same size as our language models and sort of still see gains.

33:59So I think there's sort of questions as like, how do we scale the models? And if we do scale them, what's the impact of doing that? So that's one thing that I think is kind of unknown at this time, but it's kind of interesting. Um, and then, uh, I don't know, maybe this is like a bit in the weeds, but, uh, it's, but evaluation of the models I think is significantly harder than, than, um, than text models. I think a lot of times we're trying to, like, there's no clear right and wrong answers. So a lot of what the model does, like, like, how do you know it's getting better? Um, a lot of it is really kind of personal preference.

34:35Um, and, and there's a wide variety of personal preference that different people have. So something that's getting better for one person may be getting worse for another person. And so it makes it really interesting, I would say, trying to figure out how to hill climb different aspects of your model and still be able to measure, like, is it getting better or not? So this is like something I think people and all the other people who work on images and media generation that really struggle with. But it's something where I'm excited to see what people come up with as sort of proxies or ways that people can predict quality.

35:08Can you talk a little bit about how the industry has taken on this evaluation challenge? And, you know, historically, like, are there any papers that come to mind that, you know, have become standard practice or seminal and kind of laying out an approach? Or how do you get from, you know, personal preference to some kind of function that you can optimize over? Yeah, I mean, I think from the beginning, there's always been aesthetic models that people have used. And these aesthetic models are typically trained with pairwise human preference scores. But I think that, and this dates all the way back to the beginning of imagination models.

35:51We've collected preference data and trained our models in them. But I think that there's, it's kind of interesting because it's not clear that the mean preference, kind of the average overall preference, human preference is something that you want to get to. Because it's, it's a little bit like, oh, you're in this space. You can imagine a watering down of, yeah. It's like you mix all the colors together and you get brown, you know, it's kind of like, you're kind of there. So it's, it's not like a, it's very hard to, yeah, it's very hard to say that that's, that's like the best model. So if you just optimize for very large scale preference, you end up with maybe with something that's generic and doesn't have like character or it doesn't generate interesting images um so i think like uh i mean if you look at like um midjourney for example they have spent a lot of effort on on having kind of stylistic control and being able to to generate images that that are sort of personalized to users stylistic preferences um and this is something that's that's uh i mean it's very useful for their for their users and generating these really nice images.

36:59But yeah, I think there's a lot of different ways that people can do it. I don't think we actually know what the best way is. Talk a little bit about what you see Nano Banana as being like really good at and what you think it's less good at. That came to mind because you mentioned Mid Journey really focusing on style transfer. I don't think I've tried style transfer with Nano, but I don't know how well it does, but presumably it does really good at some things and less good at other things. How do you think about the things that it's best at versus not best at? Yeah, I think that the things it's best at, I think, are

37:44preserving the content and people's identity in images that you edit. I think this is one of the reasons why it took off when we released the first version. I think, yeah, I mean, I think it's definitely not an artistic model, like a mid-journey model. So if you put a random prompt in, you don't expect to get back something that looks amazing. I think we try to make it really expressible and put a lot of effort into making sure that it's following the user instructions as closely as possible. And sometimes for creative tasks, those two things are a little bit contradictory because you might want to actually take the user away from the thing they're asking because it's more interesting or cool to look at.

38:25So I would say that because we've pushed for a really controllable and high amount of instruction following, that it's maybe less suitable for this creative exploration kind of tasks. Does that degree of controllability and instruction following imply that the bar, the prompting bar is a lot higher? Like you have to really, if you want to get amazing output, you have to, you know, give it amazing input essentially. I think we try to make it so you can give very simple input and get a reasonable output. I do, like we also have a team who are professional prompters and every time we give them the model, like I am stunned by the things that they get back.

39:09So like you can, if you know what you're doing, you can really direct this model into generating very impressive things. But I think for most users, they're not going to get to that point where that's the control. And I think that the goal there is to really have good instruction following with simple prompts and simple instructions as really a primary goal. For a while, it seemed like the direction the space was going was prompt rewriting. So the user gives a kind of average prompt, some secondary thing, like makes that prompt more elaborate and tries to produce a better result. Just based on what you're saying here, it doesn't seem like it's necessarily doing that.

39:51Can you talk about that technique in the broader industry? Yeah, I guess I'll say that I think prompt rewriting, one way to think about prompt rewriting is it's a way for a language model which has world knowledge to communicate that world knowledge to an image model that only takes text as important. Got it. So if you've got an integrated model, it's like less of a thing that you want. Yeah, I think that the thing that you want to think about is like how does the information get exchanged between world knowledge that you learn from text data and image generation that you create, learn from images.

40:35And with integrated models, you have a lot more choices there, how you can communicate information between the language models between image generation models. And I think, so, you know, I think we're probably, everyone who's working in this space probably has a different solution exactly to how that works. But like, that's kind of how I think about the role of prompt rewriters. And prompt rewriters were kind of the first step of doing that when you have like a fixed language model and you can kind of chain them together, right? You can call a language model, you can call your image model and you can pass information through the rewritten prompt.

41:09Are there any papers that come to mind and I feel like I asked this and I don't know that we got the papers. Like, are there any papers that come to mind that like you feel have been super important for the space over the past, whatever? It feels like a year is too long in this space. I'm not going to call out any specific papers. I think I've seen a lot of very good papers. I think that some of, well, some of the labs actually are still, some of the big labs are publishing papers still, although fewer and fewer. and I think those are always interesting to look at. You know, this is kind of a space where the forefront is being pushed by the big labs that are usually keeping secrets pretty close to themselves.

41:52So I think like, I like spaces when the academic community is leading and every paper is like, you know, amazing and going to change the field. And I think we'll get back to that point again at some point. but right now I mean I think that the paper is kind of like all together if you take all the papers together to kind of like enable the progress that we've seen but I don't think there's one specific paper that's really opened up an entire new area for example. Given the degree of secrecy in the space if you know for folks that are interested in the space from a technical perspective, but not at a big lab, like what's the best way, you know, from your perspective to try to gain an understanding of what's really happening behind the scenes and what's like pushing the frontier forward?

42:47Actually, I think that for sort of enthusiasts in the area, it's completely possible to have a home set up and do interesting work and interesting research. I think that the sort of required boilerplate to get systems like these up in training is pretty well supported with today's tools. The things that you can't do, you know, not as working at a smaller scale is training models that are like sort of across all axes. But I think that it's definitely possible to say, take some narrow data slice and like do experiments and kind of like, you know, make improvements in this kind of small constrained way.

43:27And, um, and those are things that'll probably set you up for like knowing how the larger systems work. So I, I mean, I think that like, it's definitely, it, you know, it's not inaccessible, I would say just because like, we do have these big well-resourced groups that are pushing the frontier of the quality. Like I don't think that means that it's an inaccessible technology for people to pick up. And, and actually I think it's quite, um, it's quite, it's quite accessible. There's actually not, you know, like I said, the base to get a baseline running is totally doable by like one excited person in their room like over a couple months and is that assuming that you're starting from some open pre-trained baseline or not necessarily i mean i think a lot of a lot of interesting work started in small scale ablation experiments you know on small toy data sets like for a long time we were using cfar and before that MNIST and ImageNet.

44:26And so like you still see, I mean, I still think that's the right place to start is to start on like smaller models, simpler data sets and work from there and make it work there. And along the way, you'll discover things that make it better. And those things you discover are probably really useful in the broader scheme of things. What do you think of as some of the most exciting, like open problems, like beyond something general like pushing model quality forward? Like, are there, you know, specific, you know, dimensions or capabilities that you feel like, you know, we've not been able to, you know, crack for these models yet?

45:06Yeah. I mean, I mentioned earlier the evaluation problem, and I think that kind of well thought out evaluations and kind of testing domains are still really interesting. I saw this great paper on vision language models about how they're fooled by, it's called the illusion illusion, and it's about how if you give the model something that's not actually an illusion, but is close to an illusion, it'll be fooled into thinking it's an illusion. What's an example of an illusion in this case? so an example is like an optical illusion for example like there's the classic one where you have the checkerboard and the shadow of a cylinder and and these two squares are the same color right so we've all seen that image a lot of times even though our brains interpret them as different colors because one is in the shadow um well it turns out if you uh if you give it to a vision model and you make them very different colors it'll still say oh these are the same color because it's seen the illusion before.

46:07So I think it's just a really clever... It's kind of the vision analogy of like the crossing the river thing where the model just can't seem to get out of that rut. Right, right. Like I think in the generation space, people have been playing with like trying to get full wine glasses for a while. And it's like the same concept, right? We've seen so many wine glasses that are not quite full that like it's very difficult to get a full wine glass. So I think like there's all kinds of interesting ways to probe the model and understand failures. And this is something that's very easy to do and really can be done distributed.

46:37And I do see a lot of interesting work coming out on kind of like evaluating where these models fail, where they don't. So that's something that I think that's always interesting. I also think that we're kind of on the early days of understanding post-training and reinforcement learning for multimodal models and generation models. So I think that's like an interesting area that there's a lot of cool papers and work coming out about. I, yeah, there's so much, you know, like once you start working on these problems, you realize that like it's not, none of this is solved. Like everything we're doing is really in the early stages.

47:12I mean, we've only been doing this for a few years. So like there's a lot more to learn. Along the lines of, you know, taking things that are happening in the tech space and bringing them the images you mentioned, RFT, another is this idea of like verifiable, data? Like, is there an analog for that in the image space? Does verifiable data mean verifying that it was generated by a model or what is actually No, more like the idea that, you know, we're going to find a data source that, you know, we can self-supervise because like, you know, we get it to generate code and we can evaluate the code.

47:55We get it to generate a math problem. We can evaluate the math problem. Like, I think it's related to this idea this broader idea of evaluation that we've been talking about um but i'm wondering you know for example like if we're talking about the domain of like give the model a bunch of like hype pythagorean theorem problems in an image like that could be a verification like there could be a verifiable loop there and i'm wondering if there are other things like that, if that's a thing that people are thinking about? Yeah, I like to say, which is something I think text people probably get annoyed by, that text is a subset of images, because you can imagine any image and just render the text in that image, right?

48:38So I think when we talk about kind of verifiable generation accuracy, I think anything that is in the text side, and of course, all the mathematics and geometry problems, I think this is like a space where we can verify images. And that's probably a pretty interesting thing to look at. But even more generally, I think you can kind of pick sort of sub axes of generation and that are verifiable by vision models. So, you know, a common one that people look at is like counting, like did the models get counting right? Because that's something you can evaluate pretty well with models out there. So I think - For perspective, for example, I'm looking at the dresser behind you and thinking that that could be an interesting way to train a model to like be able to manipulate and force perspective or being able to look at a scene from different angles.

49:25Yeah, that's right. So you could imagine if you generate multiple camera views of a scene, you could do some 3D construction and see if it actually makes sense. And yeah, and then these maybe can form reward signals. I think this is an interesting space. So like the problem I see there is like coming up with a way to get a wide enough range of these verifiable domains that you can like then generalize to being good at everything else. Um, and, and I think here we're talking mostly about accuracy, which is definitely like a, a big thing we want to improve, but you also, we also have to consider aesthetics, which is a much harder thing to measure.

50:00Um, so like, is this, is, is the lighting nice on this image? It's like, it's very much more difficult to tell, um, whether we've succeeded in that space. So I think probably we'll see that if you target a few verifiable domains and figure out how to maximize the quality on these spaces, it will be beneficial in other areas, but, I haven't seen much of this yet. And do you find like a direct tension between, you know, accuracy, whatever that means in a particular context, image context and aesthetics? Like, are they, you know, I can imagine a world where like you, you know, have sufficient data and evaluation that you train a model that's good at aesthetics and it would, you know, and then you start pushing on accuracy and it will just do things more accurate.

50:48do aesthetic things more accurately. But I can also imagine a world where like there's some fundamental tension. And, you know, I think we saw, we've seen lots of examples of this, like, you know, when in the like GPT-4 timeframe when we were trying to like get the model to, you know, do better at following instructions, like the output was less creative. What's your take on that? So I think as a community, we are just starting to get to this point because I think we've been working really hard on just making the images not be full of artifacts. And I feel like these kinds of, you know, six fingers and like solving that problem.

51:31Like, bro, I just got to five fingers. Right, so we just got to five fingers. And like, until we got to five fingers, kind of all the improvements are benefiting both of these axes because like you still need an image that's like structurally sound in order to be either correct or aesthetic, I mean, like, or like believable as an image. I mean, maybe, so I don't think we've yet really struggled with this, but I do think we're starting to get to the point where like different, you know, different teams can take different roads where they optimize for one or the other. And I think it's a, you know, in some sense, it's a bit of a data distribution problem.

52:12Like if you are able to narrow your data distribution into something that's just the good looking images, it can actually be easier to generate those because it's a narrower distribution. So it's easier to model. So I think like a lot of what we did at the beginning of these image models was figure out how to like restrict the generation to this very, very small subset of things that looked reasonable. And in the process, that was also the way to generate the best images. now as we get better at modeling and we can model this like wider distribution of things that are maybe look good and things also that don't look good but maybe have interesting information into them it opens up more more avenues for us to be able to do kind of both at the same time in the same model so i think we're we're basically broadening the distribution of things we can generate well enough that it's not just instantly thrown out you know like most of the like 99 % of things you generate before you just look at that and be like oh that's unusable image because it's full of problems.

53:08But that number is going down and the amount of things we can generate is becoming more broad. So I think we will have to choose in the future more between like, what's the purpose of this model? Is this model supposed to generate things that look good? Is it supposed to generate things that are accurate and believable? Or like, you know, what's the purpose? It seems really counterintuitive to me that generating things that look good is like that you would characterize that as a narrow domain. You know, maybe when I think about it from the perspective of like you know i don't know if we're still saying this but like there's a period of time when like you know that's a mid-journey image like all the mid-journey images kind of look the same they have this or ai generated images even more broadly they all kind of look have this look um i i feel like we've pushed that a little bit but like aesthetically aesthetic pleasingness doesn't feel like a narrow domain it feels like kind of a broad quality that you know is applied different to different to different things like break that down for me like how you think about that so what you said is definitely true that like the first models i mean i think it's a bit of a precision recall thing because if you have a model that can generate only aesthetic things it doesn't mean that you're generating all the aesthetic things right you haven't covered everything that is aesthetic but you've restricted yourself to only generating things that look good.

54:34So that's what that's the kind of like narrow view that is is a was a starting place and was easier to generate. Now that we can have models that generate broader distributions, we probably actually can get a better recall of things that different people would find aesthetically pleasing. Like maybe someone really likes crayon scribbles, like, you know, like, this is something that we can do now that we have broader distribution. So I think So I'm not saying that we covered everything that everybody has ever liked, but I think that at the beginning, we focused on things that people did like. And what's interesting is that this AI look is something that came out of that first round of trying to narrow the distribution to things that we can generate well and still look good.

55:18And then people got saturated by it. So now it's a problem if your model generates only these things in this AI look. So the tastes are actually changing a bit as well, because like when that first came out, I think everyone was just so amazed that you could generate those things that that it was still really cool and fun to like play into space of like, what is this? This like an AR looking space. But now we have to, you know, the people expect more. So we need to be able to generate more broad images. And how do you think about the challenge of as generated images become, you know, a greater portion of the images on the Internet that like you're like smoking your own exhaust, so to speak?

56:03Oh, yeah. You know, it's clearly a problem on the tech side as well. But like, is it, you know, you know, are there unique aspects to it that you think about on the image side? Yeah, I mean, I think it's definitely a problem. um i think uh you know everyone who's uh you don't want to add a bunch of ai bad ai generated images and into your training data because you'll learn to generate bad ai generated images it's like uh i think the same thing exists in the tech space as well um but yes this concept of i think the concept of of ai slop is interesting too because like there's also good ai images and it's a little less clear is like, do you actually want to train on good images or not?

56:45I mean, there's been some distillation process by which people have said, oh, that's a really good image. And like, we do know that if you can distill synthetic data to the point where it's, there's like high quality, it can actually be beneficial to train on. So it's not like you never, it's not like all AI generated images are necessarily detrimental, but like the ones that are bad are. So I think there's a, there's an interesting, there's a couple of interesting problems. There's like, How do we handle the fact that there's more AI-generated images out there and AI-generated text that we don't necessarily want to train on because we might learn factual inaccuracies or we might learn unrealistic lighting with the fact that some of this data is highly curated and is actually, people think it's really good and the quality might be much higher.

57:27And how do we tell the difference between those two? These are all interesting problems. And also this idea that we were just referring to, like if you have only produced good images in kind of a narrow, you know, band or domain, you know, if you're then training on just that band or domain, that doesn't necessarily help you get a broader ability to, you know, generate, you know, more diverse aesthetically pleasing images. Yeah, yeah, that's right. Yeah. Yeah, I think that that's something that has been a topic of discussion in AI for a while is like how much can you eat your own tail? You know, like, is it possible to train like models on fully synthetic data and use self-supervision.

58:14And there's a lot of people who have been looking at different parts of this. And I think like how much human supervision versus synthetic supervision is needed. Yeah, like I don't think we know the answers to these questions. And everything that exists in the tech space has an analogy also in images for this. You just said like everything that has an analogy, everything in the tech space has an analogy. Like one of the things that I'm thinking about in the tech space is like a revisiting of thinking about, you know, model scale versus data curation. Do you spend a lot of time thinking about that?

58:47Yes, yes. I mean, there's always discussions. I think, like I said earlier, we're just scratching the surface and it's possible that someone could come up with a much more efficient recipe that does way better is more compute efficient like i think there's still a lot of gains that are going to be made by the community in this in this area um i wouldn't predict which direction we're going like like is it all going to be like just finding the exact right data points to train a small model on or is it going to be like you know just scaling everything up 10 times bigger and hoping it all works out like yeah it's um i don't i'm not gonna predict which one is is the winner if you think about the tech side of things like from a broad perspective you know it's all still kind of early and um you know we're early in kind of this ai curve generally in terms of like it you know it's a broader impact.

59:51But, you know, in the text domain, like you get the sense that like, you know, a lot of the core techniques are established and like we're, you know, pushing really hard at, you know, a, you know, I don't even know if optimal or approaching optimal is the right way to characterize it, but like we're operating at a scale that it's, you know, it's hard to get like big, to get big leaps beyond that. And in fact, like a lot of the innovation we're seeing are folks that are like doing things to like reduce the scale and try to do more with less. Like, and I'm wondering, when I'm trying to get as like on the image side of things, is it kind of the same?

1:00:37Like, or are we like much further behind and Dan, like there's just so much headroom and, you know, all we need to do is convince Sundar to write the check and then we could like 10x this thing. Yeah, I mean, I think this is interesting. I don't know how far we'd get by scaling sort of naively. I mean, I think if I look at the text side, I feel like we saw a huge boost in accuracy come when all the main labs started doing thinking, right, like test time scaling. And that ended up being really, important. So these ideas, I think, still can also land in images, like I think test time scaling images, something we haven't really figured out quite yet.

1:01:21And probably all the labs are trying different ways to solve this problem. And so I would say that the recipes are not very well known. I hope we get to the point where everyone starts publishing again, and we can see what we've all been doing. Because maybe we'll be able to put our heads together and figure out like oh these are the things that actually mattered um you know barring so barring that i think that um the discoveries will happen sort of independently but people will figure these types of things out um and it's a lot easier to do something when you've seen it be done so when other labs make progress and you're like oh this is possible it's like okay now i kind of know what to try because i can see that something can work so i do think like there's there's this like lifting up of the the the models you know just the fact that we have all these different labs working on them um and i i do expect actually large more large improvements coming um i think you know we also saw with genie 3 that that kind of interactivity is um is really on the horizon and these like world models that you can you can have like you know you can walk around you can interact with components of them like this stuff's all really early and just launching and very exciting.

1:02:32So I think like the space is really open with possibility. Well, Oliver, thanks so much for jumping on and sharing a bit about kind of the way you think about the space and Nano Banana image generation. Congratulations. Very cool product. And it's been amazing to see like the uptake and interest out there. And I certainly had a lot of fun and gotten a lot of value out of using it. Excited to see what comes next. Thanks so much for having me. And yeah, definitely go out there and play with Nano Banana and tag me on X. If you find failures or things that you really liked, you know, I'll check these.

1:03:10So like, it's, we always appreciate feedback. And yeah, thanks so much for having me on. It's been great to chat. Awesome. Thanks so much.

1:03:27Thank you.