It would be an easy mistake to view everything in Canva’s latest artificial intelligence (AI) push only through the lens of new features. However, what the Canva AI 2.0 refresh announced at the annual Create keynote highlights is that conversational AI, with connectors to popular web apps, memory, web research, advanced coding products, and agentic capabilities, is part of the progress. However, away from the headlines, a more interesting story is emerging. Cameron Adams, who is Canva’s co-founder and chief product officer, gave us a candid insight into the company’s approach to developing AI models.
There are several reasons why Canva invested in its own AI models. “We create our own models when we see that there is a real competitive advantage for us in doing so, and when we see that customers’ needs aren’t really being met by what’s out there,” he explains, adding that at times they realize that certain use-cases and functions can only be met 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 an important job to do when you’re creating visual content, and this is an area where we’ve invested a lot of time in building a model.”
Adams explains that many different models have been created by Canva teams. The company’s focus on fundamental modeling capabilities really intensified after the acquisition of Leonardo.AI a few years ago. Canva now has over 100 dedicated AI researchers. Notably, much of the functionality that is part of the Canva AI 2.0 portfolio is built on the Canva design model, which was released with Creative OS in October. At that time, Canva also added several video editing features, including a new Timeline, magical video editing, and on-trend templates.
The company’s first generation model took more than two years to develop. In the current context, new models are being trained, evaluated and deployed in as little as a month. This shortening of the window has been underlined by advances in training.
Adams calls this a “really important unlock” that the internal team has come up with. The Canva design model, unlike the generative model, which can generate static images, also understands finer aspects of design. This includes everything from structure to layering to hierarchy, with added elements of branding, all with fully editable content in seconds. Although the company has not given detailed information about the parameters, it is designed to work parallel to a human sitting at the wheel.
Adams believes that as far as model training is concerned, Canva has a distinct advantage that other AI companies do not have.
“We have access to 14 years of design knowledge and expertise through the product we’ve built. We have huge data of everyone who is creating a design and what actions they take in the editor. We have a huge content library of 1000000 templates, images, illustrations, videos, fonts and color that we can bring to bear on understanding the design,” he explains, before even thinking about what a design actually is and what’s needed to create or edit it. Are. A more complex reality than many people realize.
“Fully layered design capability is something that text and code LLMs are not designed for,” he says. Another company’s genetic AI model may produce something that looks plausible, but it may not be completely usable. One-shot output has limited editability, which restricts customization and re-use.
Cost, free users, and on-device AI
Cliff Obrecht, co-founder and chief operating officer, points out that they don’t have to be good at just building design models. “We need to be good at building different types of models at scale, because Canva has a large, free user base,” he says, adding that Canva intends to be generous to users with a free tier as well as efficient models. “Our fragmentation models are five to thirty times less expensive than Frontier models, and this allows us to power a larger portion of our workspace for free users,” explains Obrecht.
The Canva Proteus model, which enables style transfer capability, is 2x faster and 23x more cost-effective than a comparable Frontier model. Canva Lucid Origin, which is the basis for all image creation in Canva, is said to be 5x faster and 30x less expensive than the Frontier model. Canva I2V, which handles image-to-video generation, is 7 times faster and 17 times less expensive to use.
There is a dynamic of change with respect to AI compute, where on-device models are preferred as much as possible. This is not only more cost-effective as long as the necessary hardware such as a Neural Processing Unit (NPU) is available in the computing device or smartphone, but is also effectively a more private and secure method than sending data to the cloud. Obrecht emphasizes that Canva has that perspective.
“There are a lot of on-device models being used, and we’re way ahead of that. When it hits maturity, we can provide all this power at very affordable prices to our huge free user base,” he says. “Our fundamental model development has been a core competency we needed to develop, and it’s really starting to pay dividends.”
The approach is clear, and although it may look like a V-shaped herringbone, it is actually closer to an interlinked hero through spine architecture. The more control Canva has over the underlying capabilities defined by the AI model, the better it will be able to shape the features a user sees and interacts with. The fact that the company is already planning for the on-device era makes it clear that the ambitions match the vision and effort.
Canva does not intend to fight in the same ring as OpenAI or Google or Anthropic, but instead, take a focused approach to making specific workflows more sophisticated. It’s not the most vigorous AI strategy, but certainly cost-effective.
