Canva Craft plans the layout before rendering the image

The most annoying failure of an AI image tool is sometimes neither poor realism nor a misspelled word. It is the way a perfectly good image falls apart when the user asks for a small correction. Move the title above the product. Make the cup larger. Keep the background exactly as it was. If the model treats every revision as permission to repaint the entire scene, a designer spends the time supposedly saved by generation rebuilding the composition.[1]

Canva’s answer is Craft, introduced this month as the first foundation image model developed and trained from scratch by Canva Research. The release generates images up to 2K resolution, supports more than 30 aspect ratios and is available in Canva’s Dream Lab and through a developer API. The more consequential claim is architectural: rather than passing a prompt directly into one opaque image-making step, Craft first writes a composition plan that establishes elements, scale, placement, hierarchy and text, then uses a rendering model to turn that plan into an image.[1]

Treat the layout as a decision, not an accident

This is closely aligned with how graphic designers work. An effective promotional poster establishes what the reader should see first, how typography interacts with a product image and what visual space must remain clear for the message. Canva says its model learned heavily from millions of human-made templates in its own content library, which contain explicit layout and hierarchy rather than just finished photos. Professional designers also contributed preferences to training and evaluation. The company reports evaluating more than 25,000 pairs of images head to head against competitors during development.[1]

Canva says Craft can render text in 55 languages and use its composition structure to make conversion through Magic Layers more accurate and faster. Magic Layers is a separate step that turns a generated image into editable design elements; the announcement does not establish that every image is born as a complete, layered document. The benefit to test is whether a title, object or background can be revised without rebuilding the composition.[1]

A leaderboard is evidence of preference, not universal fitness

Canva reports that Craft debuted fourth on DesignArena and placed its research lab second by the platform’s rankings. Those preference-based results describe one evaluation setting. They leave a different workflow question for designers: how accurately does the model follow a brief through several rounds of changes?[1]

Craft also changes Canva’s position in the AI supply chain. Rather than relying solely on partner models or buying externally produced outputs, the platform can modify its own model architecture, inference path and operating costs. Canva says distillation reduced the number of sampling steps and that a new inference engine made Craft an order of magnitude faster than an earlier internal version. That is an internal comparison, not a tenfold speed guarantee against a competitor. Faster Craft variants are planned, not part of the present launch.[1]

Craft is available in Dream Lab and through a public API. Canva’s detailed availability section marks MCP support as coming soon, alongside planned faster model variants. For existing Canva users, the immediate opportunity is straightforward: generate a composition inside the design environment, then see whether it survives the corrections that a finished piece of work requires.[1][2]

Report an error or suggest a correction