Canvaが挑むAIデザイン最大の悩み：Stefano Corazza氏が語るCraftとBrushの設計思想

As good as modern image generation models have become, designing with them can still feel like playing games. Put the right magic words in the prompt and you might win at the slot machine. Try to iterate on the result and you're playing whack-a-mole. That's fine for memes, but when design is your job, the game gets old. Canva's AI research team wants to make that process more predictable, with greater control over layouts, easier revisions, and models fast enough to keep up with the creative process. To get there, they've trained their own family of models using Canva's design data and integrated those models right into their tools. Here's Stefano Carraza, head of AI research at Canva, on what his team is trying to change. You have to build around iteration. AI will get 70%, 80% there. It will never get to 100%. And so you have to factor in that for the last mile, you need a human in the loop and be able to iterate in a way that is fast and also fun. I'm Sam Charrington, and this is the TwiML AI Podcast. For over a decade, I've been exploring the ideas and innovations shaping the future of AI

through conversations like this one that help you understand what's real, what's next, and what matters. Let's jump in. This has been a pretty busy week for you and your team. Well, first of all, the team was announced. Was this like the debut of AI research at Canva? Yeah, that was the debut last week, right? I think 12 hours before we actually announced Canva Craft, the model. We also announced Canva Research. We had like a blog post, a website with a lot of information about who we are and what we do. So it was really fun. And then this month has been the busiest month probably of my life, where like every week we are announcing a new model. And so exactly 40 minutes ago, we had this announcement about Canva brush. So very exciting. This visual language model for design, creation, and editing.

Excellent. And so have there been more models than just those two over the course of the month? Yeah, so we have a total of 100 IE models in production roughly right now. And some are smaller and may not be a huge, large language model or anything like that. But we started injecting and fusing AI in all parts of the product and to really help this design workflow in every possible way. But now we are taking bigger steps. We craft a financial model for image generation. And then with Brush, we are completely revolutionizing the way people can create content and also we are revolutionizing how fast and affordable that is. Because those two parameters, like, pioneers using AI are okay waiting a minute for the magic moment where something magical gets generated. But, like, if you want to really be using the day-to-day, like, that's not an option.

You know, in game development, there's a rule. Anything that takes more than 10 seconds is completely out of the question. and we wanted to get to like instantaneous AI. Yeah, I mean, particularly with the way that you need to iterate when you're creating, like those delays really stack up. Yeah, absolutely. And people, I think there's a little bit of Silicon Valley BIOS that I've seen myself being part of that BIOS that you think like the new tech comes and it's going to be single shot, perfect, and then people will just adopt it. But the reality is never the case, especially if the task has some creativity in it. It's not just a purely technical task. And so you have to build around iteration. AI will get 70%, 80% there, will never get to 100%. And so you have to factor in that for the last mile, you need a human in the loop and be able to iterate in a way that is fast and also fun. So that's the key part, I think, of what we do.

until the moment when we will have telepathic helmets that you can wear and you can transfer the full amount of your intent and your vision. But that is a little bit far out, I think. We're going to dig into the models and what you've created there, talk a little bit about how. But I'm trying to decide how interesting it is that you even did this. Like, on the one hand, you know, there are lots of, you know, their Frontier Labs have models and they, a big part of their businesses are allowing companies like Canva to use those models as part of broader processes. There are open weights models that you can build on. And so there's a lot, you know, that you could have, you, Canva, could have chosen to build on. On the other hand, it's like, you know, maybe it's obvious that a Canva with all of the data and user interactions that you have access to would build a model.

Do you think it's interesting or do you think it's obvious that you did it? No, I think it's very interesting because I think it also applies to other companies maybe of similar size. When basically there's two factors. So first of all, you have unique data that nobody else has. and then the second part is deep vertical integration of AI models in your stack is going to provide incredible performance and cost reduction. Those two factors are basically the driving input and I'm sure there's a lot of companies that are like in the thousands of employees or even smaller that will do that and where it makes sense. And so for us it was kind of the discovery of the fact that you can fine-tune or train a model that is two orders of magnitude fewer parameters if you have the data, and you can beat the bigger model, and then because it's smaller, you're going to be like another magnitude faster and cheaper. So that is the belief that we had,

and right now we are in the 40X cheaper than the most expensive frontier models. that can do the same design task, but they are using, you know, others or more parameters to do it. Instead, we have like smaller models that we can fine-tune, we can, you know, do reverse learning with unique data that we have, and we can basically squeeze a lot more, you know, same level of quality, sometimes better, but most importantly, like incredible speed and cost reduction. So that's one part. The second part is that the product itself, with AI is evolving dramatically, right? And the way AI is helping humans generate design is really changing. And so there has been a lot of like, I would say siloed innovation where like, okay, I have an image model, the image model creates an image. Then I have a next step, an LLM looks at it and thinks, oh, we need to change something. Goes back to the image model and says, hey, can you make something different?

And like in that back and forth, you lose control and you're also throwing away information. And so we have, with Craft, we have created an image model that most of the work actually is creating a scene graph of what the design should look like. And then the conversion to pixel is just the last step that we train from scratch. But at the layout, let's call it, at the design layout stage, the LLM can already intervene and make changes before the pixels are generated. And so we are seeing that there's incredible product differentiation by integrating the LLM world and the diffusion model world, if you want, in a more cohesive way where, like, the iterations can happen earlier and you have all non-destructive modifications with humans in the loop instead of having to bake the result, show it, iterate, and then to basically throw away in that process a lot of information every single time. So those were, I think, the two main factors that drove our decisions.

And also, if you look, the top models are no open source or open weights on the image generation space. I think video is a little bit of a different ecosystem. Actually, this will be probably a net result on it. So how the ecosystem for video model and image generation model is completely different, sometimes upside down. It's kind of fascinating. In what sense? Without making it a full episode, in what sense? Something happened about, I think, four months ago where some of the companies developing video models started releasing open-weights versions. And that had a fantastic domino effect where it forced all the competitors to do it. And so now you have a little system where a lot of the image models are not open-weights, where the video models are. And then because of that, I think, also because of that, you see the cost per second of AI video generated really plummeting

because there's a whole community that is trying to figure out how to crack that problem. And because there's a lot of models that are open weights, more people can apply to that problem. And so it's a fascinating different dynamics on the two ecosystems, if you want. With craft, did you start from an open-weight model and train it, tune it, or is it a from-scratch train? So for craft, we started training from scratch because we realized there were several breakthroughs that we believe were going to unlock a much more powerful model, and those were not just there in the existing model. So that one we trained from scratch. For Canva Brush, which is our vision language model, we started from an open source and open weights model, and then we built on top of that. So you're using whichever of those makes the most sense given the time or given the particular use case.

And it sounds like with craft, I have covered this idea of a kind of two-stage generation with layout and then rendering. It was in a conversation with Qualcomm, who has some papers in that area. But I don't believe, to my knowledge, there are no popular open weights versions of that. Are the two parts of a model like that, are they trained independently or end-to-end? Or what's the system approach to training a model like that? Yeah, that's a great question. So the answer is both. So we train them individually, but then there's always a time where we're trying to look at the end-to-end. But I would say that the steps are actually three now. First, we take all the input and the prompt or what the context may be, and we try to amplify and make it as powerful as possible. Then there's a process of taking that and converting into basically a structured form of data,

like more like a scene graph, if you want. And then the last step is generation of the T-cells, where our model training from scratch has shown really, really good performance. And by breaking down into these steps, you're also allowing for a better integration with agents, with LLMs, because there could be some components of that pipeline that could be useful for some tasks. If I say, hey, can you just make me 10 different layouts of this single page, right? I don't need DLN to generate pixels. You can use, you know, the craft design planning, can create different plans, can have a human in the loop to confirm, and then can do all the generations, for example. So there's actually, we are seeing new workflows coming out from breaking down into these steps. It seems like that would also help with editability. Yeah, historically, you have a scene and you say, oh, make the person a little bit bigger.

And the thing you get back is totally different. Like, we're a lot better now. Like, the GPT 2.5 image models are really, really good at consistency between revisions. But it does strike me that if you could change the scene graph and move the thing over or make the thing bigger and not, you know, have to re-render, that that would help with consistency as you're trying to move things around? Like 100%. We believe on this so strongly that we release magic layers at the beginning of the year, which is basically taking a flat image, decomposing in layers and elements so that people can really have that fine control. I mean, how many times have you tried to say to any image model generation, oh, can you move that thing a little bit to the right? Or like two feet, but in that image, there is no information to know what is two feet, you know, or two inches or centimeters. And so it's a lot easier for a designer to just move that element to the right and then regenerate.

And so our belief is with craft, basically, the humans and the designers can edit elements, can make them bigger, smaller, move them to the left, to the right, and then regenerate the pixels later. and we also are extending that capability to images that are generated outside Canva using magic layers. So actually the majority of images that we process with magic layers come from outside Canva and people generate them somewhere else and then they come into Canva and they want to edit And so there we can decompose into layers and elements and people can start moving things around and have that fine level of control. They always want, I mean, people want consistency and control, right? Those are the two things that any creative has asked AI for the last, I would say, two years. And we came quite far along, but it's still, you know, We don't want to provide slot machines where like, hey, if you generate 20 times, you might get lucky, right?

We want to actually make sure the time investment that people are putting into it is really controlled and the result is predictable. You kind of referenced earlier two dimensions that were important, quality and efficiency. Talk a little bit about each of those. We can start with quality. the model performs well, like on Design Arena is one of the number four now. Is that right? Yeah. What you brought to bear to create a high-quality model, what went into that result? I think we underestimated the power of like a decade of design generated and design data that we had. I think we were even ourselves a little bit surprised to rank that high because it was also the first submission that we did. So we have, of course, internal evals and we do a lot of also human evaluation internally, but you never know until you actually get into the real world arena.

But yeah, it was basically a lot of work over the last year really collecting data that makes sense for designers and really harvesting data that Canva has been created for over a decade now, including also data from users with their permission or data that we license from the outside. And I think we put together a pretty good body or a pretty good training set, and that, I think, delivered a lot of value. And also, we had several breakthroughs about how to use the data and how to train the models. So I think that was another, you know, the ability to run, you know, hundreds of experiments and really figured out what worked and what provided the quality. I think that was a great, you know, several breakthroughs, but one on top of the other that I think made a difference. Can you give us some examples of the kinds of things you learned that really moved the needle for you? Without getting too much into the weeds,

we underestimated how detailed of a thin description you can give to a diffusion model. Actually, if the diffusion model is above some number of parameters, can handle like incredible detail of the scene. So you can, the size of our scene description really grew in the last six months significantly as we realized, oh, actually, you can be a lot more accurate, a lot more precise. And so we expanded basically not only the initial prompt, of course, but like the description of the scene is now to a level of detail that I think no human will want to actually go and look through and edit, you know, unless it's like visually editing, right? And so there's so much detail. And by increasing a little bit the size of our diffusion model that we trained from scratch, we were actually able to capture all that detail, all those nuances, and really, you know, squeeze the maximum quality out of it.

And is that a unique property of your particular diffusion model that you trained, or is that just a general observation about diffusion models? I think it can be generalized for more cases. In our case, I think because we understand so well the coupling between the generation of the scene and the diffusion model, we really work to make sure we didn't over-generate on the scene description or under-train on the diffusion model. And so the two are perfectly calibrated to be the ideal input and output model of each other. Are the scene descriptions only unstructured text, or are there also some degree of structure, coordinates, bounding boxes, sizes, things like that? Yeah. Oh, they are absolutely structured. Yeah, it's a fully structured... is still text-based, but it's fully structured

with all the details that you just said, bounding box, relative position, absolute position, like all that. We have found that the more detail we can provide to a composition, really, the better, and we have the ability to basically render all that detail in the diffusion model. Also, one thing that I want to mention is, which is great for the creative industry overall, I think, there is this race to build the most capable LLM out there, right? And now we are already in the tens or hundreds of trillions of parameters. But for creative tasks, I think we have found that, let's say, 100 billion, it's already a pretty large number that can, if properly tuned and post-trained and fed the right data, a model of that size can already achieve so much. And we are seeing also a lot of other models,

also video models that leverage LLM as an input to create that structured data, like the Minimax H3 is another model that has a similar approach, even though it's a video model. But we are seeing the size of the basically vision language model or LLM is the input, you know, not growing 10x every three months anymore, but being kind of stable. In an industry where, like, everything grows, you know, 10x every month, like, that number has shown that, like, you know, you don't need more than 100 billion parameters to do most of the creative tasks, I would say. You know, probably, like, video understanding maybe needs just a little bit more because it's probably the most complex type of task, but for everything else, below 100 billion parameters, you can already achieve wonders. And so to be clear on what you're describing there, in this two-stage model where you have an LLM-like thing or an LLM period

generating a scene description, you're saying there's kind of diminishing returns beyond 100 billion, and smaller models do just as well as larger models and kind of nailing that scene description and providing the right input to the diffusion model. That's correct. And while being, you know, a lot faster and a lot cheaper, right? And so just the economics, they don't work. If you use, like, the latest frontier model to do that task, then, you know, it will be like a dollar per image generated or something like that. And so this, I think, together coupled with the ability of people to train those smaller models in very efficient ways, I think really unlocked a lot of value and a lot of usefulness of AI in the creative workforce. To what degree did you use synthetic data in the training process? Is that a big part of your recipe? Yeah, it is. I think like in every data set of this size, we also want to make it as diverse as possible.

And so we have all kinds of data coming in from, you know, stock images and that kind of stuff to synthetic data, to user-generated data, to designs that we created. We have, you know, millions of designs that Canva created or licensed over the last decade. And so that was also, of course, a big part of it. It kind of helps also with the taste, right? Because we have seen AI slope and AI generation where, like, everything looks very similar. And so I think we put so much effort on diversity, which also aligns with our company value. And so we are a company that wants to empower everybody to design. We are global. We serve every market, every language. And so we wanted to really push the diversity to make sure that the output also of the model is as diverse as possible. Talk a little bit about the way you think about benchmarks for image generation. You know, there are well-established benchmarks, but I always wonder, you know, how well they capture kind of the human perspective and human judgment and taste, as you just referenced.

And I'm wondering, you know, someone who works on this constantly, like how do you think about, you know, the applicability of the popular benchmarks, how you supplement them, that kind of thing? That's a great question. I would say the benchmarks overall are lagging a little bit behind to where the power of AI in design is going to be. Because they are mostly focused on single shot, the first generation. And you have benchmarks for image editing, which is great. But design editing is different than just image editing. And so as of today, there is no established benchmark for how good a model is editing a design, for example. There's some public benchmark for generating, let's say, a presentation like this, present bench. I mean, one has like 200 examples of like presentation-related slide-related tasks. And you can test it, but again, it's new generation. and so there's not a lot of benchmarks

that really can assess the effectiveness of editing a design, the speed and the cost also the goal with that. So I hope there's going to be more. Of course, we have internally, we have a lot of those but yeah, we haven't found an industry established benchmark for the ability to actually edit a design because the industry is focused on, oh, AI is so magic it's going to single shot everything which we know is not the case. And so I think now that understanding is finally landing, and I hope we're going to see more benchmarks that really start from, like, a given set of pre-existing designs and defines rubrics for how those designs can be modified, and then does the comparison and the benchmarking on that. That kind of implies that you can use an LLM or a VLM as a judge to look at an image and assess it visually, you know, from any number of perspectives.

Are we there yet? Or, you know, how good are VLMs at kind of articulating the qualities that you might want to optimize an image for? I think they're pretty amazing. and we have also to keep in mind that if you structure the rubrics really well, they are really good at defining, I don't know, is this design readable? Like what's the score of readability? What's the score of layout and composition? So we have very well articulated rubrics about design quality and then we can use VLMs to judge that and we get quite far along. Of course, we always do also human evaluation, and some of the arenas are strongly leveraging that because it's the ultimate test. But I think, yeah, we came a long way. It forced us also to really define what a good design means, right? And so if humans are not able to articulate that, then BLMs are not going to be able to store it, right?

And so I think we had to really go back and really understand what makes the design really good to be able to leverage the VLMs to do it. But I think now we are getting to a place where it's quite reliable. When you think about this two-model approach in craft, what are the qualities of an image come from having the two separate systems or models, and what just come from having better training data? Yeah, that's an interesting question. We have seen just different behavior depending on different tasks, for example, which we are always trying to understand. Is the limitation the diffusion model or is it the, you know, the lens side of things? And sometimes it's hard to know. We have seen, for example, the model perform really well on like portraits and people and also nature landscapes. we have seen the model doing not so great on the cartoon type of content, for example.

Is it not so great? Yeah, I mean, it's not ranking in the top three. So, like, all that is together, that is counterintuitive to me because, you know, a portrait, for example, seems like it would kind of null the value of, like, the planning and the scene component, like it's a portrait. In a cartoon, for example, you've got all these things in a cartoon, that would lean a lot on a scene-generating component. But the two results that you're seeing suggest that that model shouldn't be as strong as it is or that that's not the source of the strength of the model, right? Am I thinking about that correctly? Yeah, it's interesting. So the number one thing for people in portraits that you want is basically to be as photorealistic as possible, right? Like you don't want to have that AI skin effect and all that stuff. And we put so much effort in this because we were like we got an alien It should not look like AI generated I think that effort may have bled into the cartoon side where instead you weren very stylized

And so it turns out that the images, the cartoon images generated with craft initially were a little bit too realistic and too 3D looking. But again, this is something that we discover when we put it on the arena. because probably they had, like, a different population of designers that would test it. But at the same time, it's something that is more about the taste of the model, and that is a lot simpler, a lot easier, a lot faster to tune and correct, and so that we are in the process basically of doing exactly that. And, you know, tuning the taste of the model and the style of the model is more of a fine-tuning and RL type of task. You don't have to pre-train, right? In the pre-training, we, of course, had a lot of cartoon content. Just now that capability didn't surface to the full extent. And so now we are like teaching the model, okay, this is a style that resonates more with humans.

And we want, you know, if this cartoon needs to be flat shaded and these lines and so on. I mean, it's fascinating, honestly. it's a journey of like learning as much as the model is learning we are learning how to train the model and how the model behaves sometimes you find this maybe a little bit weird but sometimes I have like two small kids and sometimes I learn things about learning as a high level concept that I learn by training the models I'm like oh this I'm sure applies also to humans And so it's kind of fascinating, especially on pre-training. If you see pre-training, it's basically like teaching a human as many words as possible about vocabulary, right? But it has nothing to do with the style of talking performance. Or like if that human becomes, I don't know, a stand-up comedian, how many words it knows has nothing to do with the style of the delivery, right?

And so I think the concept is so true for pre-training and post-training and the eye model, and it's also true for, like, humans, where, like, in the early years, you want as much variety and as much exposure to different type of activities, sensory information, experiences, languages, and so on, right? And so the parallel actually is fascinating. I think I'm sure there's books written about it. Let's talk a little bit about brush. is brush, what's the best way to think about it? Is it an editing model? Is it a design agent? A little bit of both of those things? Yeah, we call it like a vision language model, but it's basically an agent that can create design, it can also edit design. So it has a deep understanding of what's going on in the layout of your design and can really take very fine-grained directives from humans and then operate on the design itself.

So what is a design in this context? It's not an image. It's, think of it like a slide or a Canva document, multiple layered composition of things. Yeah, it could be multiple layers, could be multiple elements, could also different type of elements. I can have some vector graphics and some image mixed up together and then some text as font. So it's basically this heterogeneous composition of different elements. They have, you know, particular features and particular type of details that humans can tweak and also the agent can tweak. And then they all come together to create that unique design that is the final output. And so it's a little bit like what craft generates in the first step, right? It's basically all the underlying structure of a design. Brush really understands that very well and can really act on the design and then relay out, can move parts,

can make it more readable, more professional, can take high-level inputs like, hey, make this thing pop. Or make it cool. I'm like, make it Gen Z. These are directives that will be extremely hard for a designer that is not maybe a Gen Z designer to really understand. So those are high-level directives that can really be applied by brush in a quite successful way. I'm kind of thinking about it like, so with craft, you've got this scene creator, and you've got this generator, and the generator generates one kind of one image. And brush, you have the same scene layout thing, but maybe it has different tools and one of those tools is generating images and it's got layout tools and the output is like a composition of sorts. Is that a reasonable way to think about it?

Or where does that way of thinking about it break down from what's actually happening? Yeah, it's not that far. I would say right now Brush can do more general things than just operating a layout description of the scene. But basically, it operates a set of tools, like a large set of tools that are all design tools. It can make decisions on design. I think it goes beyond the image generation and editing, right? It can also create, like, you know, a 30-slide deck. It can make sure that deck is cohesive with, like, branding. So I think it operates, like, a higher level of abstraction than craft, then craft, which is really about image, and soon we'll be also doing more of image editing. And so, yeah, brush is basically a full agent that you can ask for more complex work and more abstracted type of instructions. And then we are constantly iterating on, I mean, it's a lot is like having basically an LLM in the loop

using a bunch of tools and being able to see what it generates. And that loop is really what has changed everything because until, I don't know, a year ago, LLMs would just do like, we're blind and we'll just do a single shot and then be blind to the result, right? And instead now, not only that single shot becomes an iterative process with human in the loop, but LLMs now can see what they generate and can iterate. and the faster you make that iterative loop, the more effective everything is because now an ILM can do maybe 10 iterations before it even shows the result to a human and before every iteration would take, you know, a minute and so that was now, you know, feasible. But now, you know, that's why speed is not just a nice-to-have. Speed is like a complete game-changer because it can really increase the quality, it can increase the user experience, it can really make it like an iterative process. Sounds like there's a simpler way to think about it.

Craft is a model and brush is kind of an agent. And so in creating brush, I'm imagining part of the task here is like a custom harness. I'm wondering what all that entailed. How did you approach that harness development? This is a great question because I think people are, usually they are all focused on the model and not the harness. But I think the importance of the harness cannot be overstated. And for us, this probably was the hardest part and the part where we had the most iterations. In fact, we already had our third generation of harness for brush. And really, the high-level goal is basically to give as much freedom to the LLM to really run free of constraint and really be able to tap into the tools and have as much insight on what is able to generate.

And so we have really focused on making the harness thinner and smaller and more agile and more simple. And then at the end, you know, brass can write code to operate tools and make design operations. And so we were at the same time trying to leverage how good these LLMs are writing code, and at the same time, simplifying the harness in the most basic possible way to really unleash the full power of the LLM. And so that's where we did a lot of iteration, and actually it was more about removing the human knowledge and just letting the LLM run free to some extent. And when you talk about the code generation, what kind of code are we talking about? I can imagine a lot of what Brush might want to do is running a Python code, but I could also imagine you creating some design DSL that more natively represents the things that you care about.

Yeah, it can be that, or it can be just writing code also to operate API calls and layout operations and that kind of stuff. So it's primarily code or a little bit of both? It's definitely both, yeah. Like, the capabilities are quite broad. What's an example of something that a user would do in brush that uses, you know, either code or some kind of DSL or abstraction? I mean, a lot of the editing operation that you can do on an editor, they can be expressed as, like, editing to an editing API, right? And the problem is there is there's, like, infinite ways of describing that design operation. let's say I take an object, I move it to the left, and I rotate, and I make it bigger. These three operations can be encoded as a format change. You can have a format that describes the whole state, or you can describe them as this,

and they can be achieved through an editing API. And so that is the complexity that we have to navigate, trying to understand what was more natural for an LLM, which actually has to do with what is the most available data out there that LLMs are trained on, right? So this way you kind of leverage a groove that has been already created by an enormous body of data instead of trying to teach something new or like that is underrepresented in the training set or is auto distribution in the training set of an LLM. And so I think that's where, you know, you try to align with the wind, right, and the capabilities as opposed to trying to, like, fight it. Earlier I mentioned, like, kind of these two North Stars, you had quality and efficiency, and I think we double-clicked on quality but didn't really double-click on efficiency. Recap your kind of vision for efficiency and, you know, talk through in the case of craft and brush,

What were the levers that you manipulated to drive that efficiency? Yeah, efficiency in inference specifically really enables affordability of AI. And the majority of our users are actually free users of Canva, and we love that. Like students, nonprofits can use Canva for free. If we can bring the cost of AI inference low, we can also allow them to use more of AI. And so we really wanted to make sure AI was something that was within reach for everybody. And so the optimizations, what we have learned is we are overall relatively early in the journey of optimizing AI inference. If you can squeeze 10x performance improvement in a month, means you're very early, right? Like, we are so far from the plateau right now, and that's very exciting because it's also very satisfying for our AI engineers to be able to work on a problem for, like, two weeks

and be like, oh, okay, it's five times faster than it was two weeks ago, right? And so we have put a lot of effort, and you can only make it really efficient if you own the full stack, right? because there are optimizations you do at the GPU level, at the hardware level, the optimization you do at the kernel level, this optimization reducing the number of steps the model does, the distillation. And so going from basically high level to low level hardware, and how you survey, how you batch it, how you cache it, there's like a pile of things that you can really do. And then if you tackle all of them, you're going to squeeze significant gains out of each step. And that's how we have seen, for example, the cost of generation of a second of 720p video in a little bit over a year go down two orders of magnitude. Right now, you can generate sub-cent, where like a year ago, it was like half a dollar per second of generation.

And so there's still a lot to do there. I'm not going to forecast that AI inference is going to be as cheap as storage but we have seen what happened with storage at the beginning we were paying by the gigabytes and it was so expensive and you had to do your math and I don't even know how many terabytes I have to storage in my Dropbox account I don't know I don't think I would ever have the problem to think about it and so I don think AI inference will get to that level but it getting so much better than it was you know a year ago And I think given the first derivative is definitely, you know, still very healthy. I think we're going to see probably another order of magnitude going down. And when you get to like 0.1 cent for a generated image or generated second or video, I think it's another breakthrough. And all of a sudden, you know, the democratization and affordability go really beyond what we see today. In building that full stack and operating that full stack,

are you operating the full stack? Do you have hardware in a data center, or do you delegate part of that to the cloud, but you're running everything on top of cloud providers, clusters? Like, how far down do you go? Yeah, we are as far down as we can go, other than owning the actual hardware ourselves. So we run those models on our cloud, we optimize them, we manage how we optimize the inference, how we serve it, all that. So the raw thing that you're getting from someone else's GPU instances and you're clustering them, you're loading them with whatever your inference stack is, you're tuning LLVM and all the bits and pieces to batch sizes and all the stuff you talked about. Yeah, exactly. Talk about how has your knowledge about that space evolved? Were there things that you found surprising

once you started to work on these new models or did it kind of roll out as expected? Yeah, it's interesting because it was an industry that was completely dominated by AI researchers, if you want, where you had, like, graphics engineers or, like, cloud engineers coming in and being like, oh, my God, there's so much more juice to squeeze here. And so it's been really, like, phenomenal to see how. And here in San Francisco, I saw, like, a whole wave of startups actually focusing on inference optimization. I mean, because there is so much money left on the table right now. And so there's been this massive wave. I think I can probably name a dozen of companies that do that because people that have a different background realize, okay, these models are really not run efficiently. We should rethink about how we feed all this data and these jobs into the GPU.

So there's been really a flurry of interest because it's also a very simple value proposition to explain. This thing costs you now like a billion dollars, and if I do this optimization, it's going to cost you one-tenth of that. Everybody gets it. It's very easy to explain. And it's a very easy ROI to put in place. And we are seeing now, as AI adoption is still skyrocketing, we are seeing a lot of pressure on reduction of cost and making it scalable. And so it's been fascinating to see a lot of those startups and also in big companies bring in knowledge that is more about graphics. And graphics has been optimizing things for graphic cards for decades. And now that knowledge is brought to, combined with cloud scalability knowledge and AI models, all those three components together are really delivering wonders. It's quite exciting. And are there prerequisites

to being at the steep part of the optimization curve? meaning do you have to have a certain degree of scale so that you can manipulate batch sizes or do you have to build your own models such that you can take advantage of deep levels of control? Where do you think it makes sense to own that full stack? Yeah, it was really like an interdisciplinary type of endeavor where we actually put together different teams that had different skill sets. And when we saw the combination of those come together, that's where the best games came. And yeah, absolutely. If you have to host, you have to self-host your model, and then you have to have access to the weights of the model. If you're using an open weights model, that's good. But in some cases, you don't have access to that. So in that case, you build your own model.

But when you're self-host and when you have access to the weights, there's a lot more that you can do and a lot more juice that you can squeeze out of those models. And again, we found that by bringing together a very interdisciplinary team with different backgrounds, we were able to really tackle the problem with a different perspective. And again, it goes back to diversity, right? In this case, it was a very diverse mindset from the AI researcher that we brought in, and I think it really delivered. Going back to craft and that kind of two-part model, does having a unique model like that get in the way of some types of optimizations like distillation? No, distillation is something that we do pretty regularly on every model that we develop. And so for Kraft specifically, we have announced that there's going to be a quick follow on Kraft Fast and Kraft Ultra Fast.

And those are basically models where we are doing also the distillation among many other optimizations. And again, by having all the models in-house, we can also make sure that the distillation that we do on the LLN side and the one that we do on the diffusion model side are also tuned to work optimally together. So again, it goes back to the concept of making sure that the input and output are really tuned for the best quality. And so we're going to have more distilled versions of craft coming out in the next few weeks. Because again, we want to make it even more affordable. And we want also to leave the choice to the user, whether it is a developer using an API or a designer actually using it through the UI. or to MCP, we launched it. I think now it's also available on MCP as of yesterday, Kraft. And that's another place where, like, you know,

if I'm building a prototype app, I don't want to run on the most expensive model, of course. And so the ultra-fast and faster version of the model I was designed, for that they may become, you know, kind of the default, but then when you want quality, then you switch to the higher-stilability model. So the MCP is a Canva MCP or a Craft MCP? Through the Canva MCP, you will be able to also access Craft. Okay. There's basically a bunch of AI capabilities that we are exposing through MCP, and Craft is being added. And then as we add more variants of Craft, we're going to keep adding to MCP, which, by the way, we are seeing the usage of MCP on camera, like a really skyrocket. I think people are finding a lot of value, so we want to give more capabilities, access to more capabilities to MCP going forward. Do I understand correctly that beyond the MCP access to craft,

you are also making craft and some of the other models directly available to folks? kind of, I don't know the specifics, but I'm envisioning like via an open router or via Canva directly, but just per use access to the model? Yeah, absolutely. So we released it through our API called Canva Connect. And so now it's available. You can access through the API as well as MCP. So for developers that want to do that. That feels like a very different business, like running a super low latency, very narrow model inference service as opposed to this is an API for our product. Is this the first model that you're hosting in this way? We have exposed APIs for various capabilities. I think also you can access magic layers and some of the image generation models that we had before, like Lucid Origin. But I think now the level of interest

we are getting is a lot higher. And yes, you're right. It is a type of different business, if you want, right? But we are getting so much demand coming from mostly like two categories. So enterprise customer, our enterprise business is the fastest growing business that we have in the company. Really, you know, triple digit growth year over year. And we are getting a lot of demand also from those enterprise customers to access the capabilities to, you know, an API. because they may build their own custom workflow. Now that people can vibe code anything, right, there's going to be more demand for that. And then we are getting also a lot of demand from people using, I don't know, Cloud or ChargeGPT to MCP. They now want to access more capabilities, right? So the agents can go there and say, oh, you want to generate an image? I can use Canva. And then they can always then open in Canva for further editing. So we are seeing these two type of use cases really growing.

And sometimes those users are already using Canva and they just want to use it in a different way or they just want to build scalable applications using some of the Canva technology. So we have seen a lot of demand. And yes, it's a new and exciting business and we're still learning a lot about how to best serve it. But I think this we can say is not unique to Canva. It's across the industry, right? There's a lot of companies that all of a sudden are like, oh, this is, like, we are basically building not only for humans, but also for agents, right? Even though, you know, humans is our focus even for AI features, but there's going to be in the future just more agents using some of those capabilities. I think that's unavoidable. What do you think those agents need that's different from what humans need? I think they need more structure. They need documentation in a certain, defined a certain way so they can digest more things. They need the ability to try to preview things without having to subscribe.

And so a lot of things that basically are designed to remove friction, finding information more easily, streamline workflows, because ultimately many times it will be the agent that will decide what solution to use. as opposed to the human. And so we have to basically make sure, for example, the agent is able to like spin up with all friction and not a lot of burden on like initial credentials or credit or purchase. And we can spin up something that the agent can test, verify it's working. And then, you know, go back to the human and be like, hey, I created this like a prototype app using this API. I tested it, it works. You want to give it a shot? Okay, now it's approved. then now we can go for the real scalable thing. So that's the kind of approach. Just to kind of wrap things up, where do you see Canva evolving with, you know, in terms of research in particular? It's a world where like, you know, typically you plan like a year ahead and now it's completely impossible.

It's like, it's like just a month ahead is already like a 3D, you know, pretty foggy kind of thing. But at the high level, you know, we have a very clear mission and vision. That one is not changing. So we want to make AI more accessible. We want to make it more part of the iterative creative process. And we want to remove any barrier. And we want to reimagine all those workflows. So that's the direction that we've been pursuing. And what we're actually doing, what models we work on, specifically at Canva, yeah, this is going to change, you know, every few months. And so I, you know, I told my team basically, don't fight, you know, this ever-changing, constantly changing ecosystem, but just like embrace it. And so I think it's really like a mind shift of like, everything is going to change, which means also like, what is important today may not be important tomorrow,

which opens up also for opportunities, right? it's a double-edged sword to some extent, but at the same time, we see that as a big opportunity. And so we are constantly analyzing our business, see where there are opportunities, where there's bottlenecks, where there are technical parts of our workflow that we can automate, and then we prioritize those and we work on them. Like a year ago, none of the projects, the models that we ship now was underway. you know so it's logic to think that in one year we will probably release models that we haven't even started working on it today and that's you know it could be terrifying or could be very exciting so I think I think researchers are probably the most exposed to the speed at which AI is changing and so it's it takes really like some major adaptation and mindset shift but at the same time it's quite exciting well Steph thanks so much for taking the time

to share a bit about what you and your team are up to. It's very cool to hear more about it. Thank you, Sam. It was a true pleasure and honor to be on your podcast. Thank you so much. Thank you.