Canva just entered the image model race with something other than a me-too diffusion clone. The company announced Canva Craft, its first foundation model built from scratch by Canva Research, and the pitch is simple: models trained on photos make pretty pictures, but models trained on templates understand design.
That’s a smart bet. Whether it holds up outside Canva’s own framing is the real question.
What Craft does differently
Most image generators go from prompt to pixels in one pass. Craft splits the job in two. A planning model first writes out the composition: which elements appear, where they sit, relative scale, visual hierarchy and any text. A separate rendering model then turns that plan into an image at up to 2K resolution, across more than 30 aspect ratios.
The payoff should be better text, cleaner layouts and edits that don’t scramble everything else. Text rendering has been the weak spot for image models for years, so this is the right problem to attack. It also fits neatly with Magic Layers, Canva’s tool for turning generated images into editable elements. Craft uses JSON bounding boxes to make that conversion faster and more accurate.
The training data is the moat
Canva trained Craft on millions of human-made templates from its own library: posters, menus, infographics, presentations. Midjourney, OpenAI and Google can’t easily copy that. Templates encode how hierarchy and typography fit together, not just what a finished image looks like.
Canva also tuned the model with feedback from its professional designers. Designers and researchers compared more than 25,000 image pairs against leading models, and when automated scores disagreed with human judgment, the designers won. Good call.
The rankings, read carefully
Canva says Craft debuted fourth on DesignArena, a crowdsourced benchmark, which it translates into the world’s second most advanced image lab. That second claim is a lab-level ranking, and the model itself sits fourth. Both can be true, but the headline is doing some work.
The category results are stronger. Craft ranks first for People and Portraits (out of 87 models) and for Landscapes and Nature (out of 85). It’s second for Abstract and Patterns, and fourth for Products. One benchmark, though, is one benchmark. Developers should test on their own prompts.
Speed and cost
With more than a quarter of a billion users creating over 400 designs per second, Canva can’t ship a slow, expensive model. It used distillation to cut inference steps sharply and built a custom inference engine for the planning stage, which it says is an order of magnitude faster than the first internal version. Two cheaper variants, Craft-fast and Craft-ultrafast, are coming. Pricing wasn’t detailed, so the cost claims remain unproven.
Availability
Craft is live now for all Canva users globally through Dream Lab. It’s also available through the public API, aimed at marketing platforms, e-commerce sites, ad builders and print services. Canva MCP support is listed as coming soon, so AI agents can generate ready-to-use images as one step in a larger workflow. Prompts work in any language, and the model can render text in 55 languages.
Why it matters
Canva is moving from model customer to model builder. Adobe has Firefly, and Figma and others rely on third-party models. Owning the model gives Canva control over cost, quality and editability, which are the things that matter for production design work.
The API is the interesting part for developers. If the planning-then-rendering approach delivers reliable text and layout, Craft could be a practical choice for ad and e-commerce pipelines, even if it never tops a general image leaderboard. That’s a narrower ambition than beating Midjourney, but it’s a more defensible one.
