Canva has launched Craft, its first proprietary image model. The distinguishing claim is architectural: Craft plans a design before rendering it, rather than generating an image in one pass. Canva says this buys "greater precision, consistency, and control". Output runs to 2K across more than 30 aspect ratios, available through Dream Lab and an API.
The plan-then-render approach is aimed at the failure mode that makes general image models frustrating for actual design work. A general model asked for a poster produces a picture of a poster - text in the wrong place, hierarchy invented, a layout that cannot be edited. A model that decides on composition first is trying to produce something with structure, which is what a design tool needs to hand back to an editor.
It also makes Canva a model vendor rather than a model customer, competing with OpenAI, Microsoft and Google on a surface they all want. Canva's advantage is not capability, it is that it already owns the editor the output lands in.
Sources
BetaNews · Technology.org