Canva builds its own image model after AI costs cut its forecast

Canva released Craft, its first image generation model, on 22 October. Canva Research, the company’s in-house AI lab, built and trained it from scratch. It is available to every Canva user through the Dream Lab panel and to developers through Canva’s public API, and Canva lists a connection through its MCP server, which would let AI assistants make images as one step in a wider task, as coming soon.

The launch followed by a day the lab’s first public account of itself. In a post on 21 October, Stef Corazza, Canva’s head of AI research, wrote that Canva Research has more than 160 AI scientists and engineers inside a broader AI unit of 400 people, and that it was a third of its current size when Corazza joined more than a year ago.

“To deliver AI at Canva’s scale, we knew we needed the flexibility to develop and run our own models alongside those from our partners,” Corazza wrote. With more than one billion designs created on Canva every month, the post added, “even a variation of a few cents per generation makes an enormous difference.”

Canva leads its Craft announcement with leaderboard rankings. The case for the model sits in the cost line underneath them: Canva wants more of the AI inside its product to run on models it owns, alongside those from partners, so that less of what it pays per image depends on another lab’s prices.

The August reset

In a second-quarter update to shareholders, reported by Startup Daily on 5 August, chief executive Melanie Perkins said Canva had been “relying too heavily on frontier models” and that its average cost of serving an AI task was too high. Canva cut its 2026 revenue growth forecast from 30% to 20%, and Perkins told shareholders that its pricing, consumption model and usage controls had not kept pace with demand, according to the same report.

Startup Daily reported at the time that Canva had cut the cost of serving an AI task by 90%. Corazza’s post describes two of the in-house models built with cost in mind. Canva’s image-to-video model started from open-weight models, was improved with Canva’s own data and now produces 70% more video generations per GPU hour. Lucid Origin, a model from Leonardo, the image AI company Canva bought in 2024, is described as delivering “unprecedented cost savings” for high-definition images.

Canva says Craft is the first foundation model Canva Research has built and trained from scratch, rather than adapted from another model. Canva says model distillation cut the number of inference steps Craft needs to a fraction of the original, and a custom inference engine made it an order of magnitude faster than its first internal version.

A model trained on templates

Canva says most image models are trained mainly on photographs, and that it trained Craft on millions of human-made templates from its own library, including posters, social posts, presentations, menus and typography-heavy layouts. Its designers compared more than 25,000 image pairs against leading models during evaluation. The model writes a plan of the composition before it renders the image, produces images up to 2K resolution in more than 30 aspect ratios and renders text in 55 languages.

That training is aimed at the work business users already do in Canva: flyers, ads, social posts and presentation visuals, where the words in the image have to be legible. Craft’s output can also be split into editable layers with Canva’s Magic Layers tool, so a generated image can be changed as a design rather than generated again.

Canva says Craft debuted fourth among image generation models on DesignArena, a crowdsourced design leaderboard, and ranks first in its portraits and landscapes categories. Those placings come from a public leaderboard as Canva reports them, not from an independent test of the model against a company’s own brand work.

The price that is not published

Two faster variants, Craft-fast and Craft-ultrafast, are on the way. Canva says they are designed to deliver high-quality image generation “at a fraction of the cost of leading frontier models”. Neither post gives a price for Craft through the API, or says how its use will be metered on paid Canva plans.

For mid-market firms that buy Canva seats for their marketing teams, that is the part of the launch that touches the bill. Every image a customer generates inside a seat price is a cost Canva carries, and Canva told its shareholders in its second-quarter update that its usage controls had not caught up with demand. A model it owns gives Canva more room to keep image generation inside the seat price than a model rented from another lab by the call. The launch does not say whether Canva will use that room or meter Craft through those usage controls.

Corazza hopes Canva Research will be known in 10 years as a team that made AI “humble, affordable, and playful”. Until Canva publishes what Craft costs through its API and how it is counted on paid plans, the second of those words rests on Canva’s own description.