It’ll be an easy mistake to see everything in Canva’s latest artificial intelligence (AI) push, just through the lens of new features. However, what underlines the Canva AI 2.0 refresh announced at the annual Create keynote is more than jus that, with conversational AI, connectors with popular web apps, memory, web research, enhanced coding product and agentic capabilities, all part of the progression. However, a more interesting story is unfolding away from the spotlight. Cameron Adams, who is Canva’s co-founder and Chief Product Officer, gave us the clearest view of the company’s approach to developing AI models.
There are reasons why Canva invested in their own AI models. “We build our own models when we see that there is real competitive advantage for us to be able to do so, and also when we see customers’ needs not really being met by what’s out there,” he explains there are times when they realise certain use-cases and tasks can only be fully completed if they develop their own models. He points to the background removal feature that’s part of the editing array on Canva, as an example. “We know there’s a key job to be done when you’re creating visual content, and that’s an area where we’ve invested a lot of time in building a model.”
Adams points out that there are many different models that have been produced by the teams at Canva. The company’s foundational model capability focus really sharpened after the acquisition of Leonardo.ai a couple of years ago. There are now more than 100 dedicated AI researchers at Canva. More to the point, a lot of the functionality that’s part of Canva AI 2.0 portfolio has been built on the Canva Design Model that was released in October, alongside the Creative OS. At the time, Canva added a number of video editing features as well, including a new timeline, magic video edits and on-trend templates.
The company’s first generation of models took over two years to develop. In the present context, new models are being trained, evaluated, and deployed in as little as a month. This shortening of the window is underlined by advances in training.
Adams calls this a “really critical unlock” which the internal team put together. The Canva Design Model, unlike generative models that can generate static images, also understands finer aspects of design. This encompasses the chain, from structure to layering to hierarchy, with the additional elements of branding, with completely editable content in seconds. While the company hasn’t detailed specific such as parameters in play, it has been designed to work as a parallel with an human at the wheel.
Adams believes that as far as model training is concerned, Canva has a particular advantage that other AI companies simply don’t have.
“We have access to 14 years of design knowledge and expertise through the product that we’ve built. There’s massive data from all of the designs that everyone’s been creating and all the actions they take in the editor. We’ve got a massive content library of 1000000s of templates, images, illustrations, videos, fonts, and colour that we can bring to understanding design,” he points out, before adding that thinking about what is actually a design and what entails to generate or edit it, is a more complex reality than many realise.
“That fully layered design capability is something that text and code LLMs just aren’t tuned for,” he says. A genetic AI model from another company can generate something that looks plausible, but it may not be entirely usable. A one-shot output tends to have limited edit-ability, which further restricts adaptation and reuse.
Costs, free users and on-device AI
Cliff Obrecht, who is co-founder and Chief Operations Officer, points out that they don’t have to be good at just creating design models. “We need to be good at creating a variety of models at scale, because Canva has a large, free user base,” he says, adding that Canva intends to be generous to the users on the free tier as well with efficient models. “Our fission models run at costs between five to thirty times less expensive than frontier models, and that allows us to power a lot of our workspace for the free users,” Obrecht explains.
The Canva Proteus model, which enables the Style Transfer capability, is 2x faster and 23x more cost effective than comparable frontier models. The Canva Lucid Origin, the basis for all image generation in Canva, is said to be 5x faster and 30x less costly, when compared with frontier models. The Canva I2V, which handles image-to-video generations, is 7x faster and 17x less costly to use, in comparison.
There is a shifting momentum regarding AI compute, where preference is towards on-device models, as much as possible. This is not only more cost effective as long as the required hardware such as a neural processing unit (NPU) is available in a computing device or a smartphone, but is effectively a more private and secure method as well compared to sending data on the cloud. Obrecht insists Canva has that perspective in view.
“A lot of on-device model usage is coming, and we’re way ahead of where that is. When that does breach maturity, we can provide all of this power to our huge free user base for very, very affordable prices,” he says. “Our foundational model development has been a core competency we’ve needed to develop, and it’s really just starting to pay dividends.”
The approach is clear, and while it may look like a v-shaped herringbone, it is in fact something closer to an interlinked hero shrew spine architecture. The more control Canva has over the underlying capabilities defined by the AI models in play, the better it’ll be able to shape the features a user sees and interacts with. The fact that the company is already planning for an on-device era, makes clear that ambitions are matching vision and effort.
Canva doesn’t intend to battle in the same ring as OpenAI or Google or Anthropic, but instead, a focused approach to make specific workflows more refined. This is not the loudest AI strategy, but certainly a cost effective one.
