Canva has built its own image generator. On October 22 the Australian design company released Craft, the first foundation model trained from scratch by its in-house lab, Canva Research, and made it available to every Canva user and to developers through its public API. Until now Canva's image features ran largely on models from outside providers.
Craft splits the job in two. A planning model first reads the prompt and writes out the composition: which elements appear, where they sit relative to each other, how large they are, what should dominate, and any words that must be shown. A separate rendering model then turns that plan into a picture as large as 2K, in more than 30 aspect ratios, or a custom one. Canva says separating layout from rendering produces more legible text and more precise object placement, and lets users change one element without disturbing the rest.
The training data reflects the company's catalogue. Rather than relying mainly on photographs, Canva says Craft learned from millions of templates made by people in its own library, from posters and social graphics to slide decks, menus, infographics and type-heavy layouts. The model was then tuned with supervised fine-tuning and reinforcement learning using feedback from Canva's professional designers, who also judged more than 25,000 head-to-head image pairs against leading models. Craft accepts prompts in any language and can render text in 55 languages, according to the company.
Canva is leaning hard on one benchmark. It says Craft entered DesignArena, a crowdsourced site where people compare images from identical prompts without seeing which model made them, in fourth place overall, which Canva presents as making its lab the second strongest in image generation. It also claims first place in the People and Portraits and the Landscapes and Nature categories. Capital Brief, an Australian business outlet, reports that the announcement was brought forward after those preliminary rankings placed Craft ahead of models from Google, Microsoft, Meta, xAI and ByteDance, and that OpenAI's lab sits above it.
Speed and cost are part of the pitch. Canva says distillation cut the number of inference steps sharply and a custom engine for the planning stage made Craft an order of magnitude faster than its first internal version. Two lighter variants, Craft-fast and Craft-ultrafast, are promised soon. Outputs are designed to feed Magic Layers, Canva's tool for turning a flat image into editable design elements, and Canva says a connection through its MCP server is coming so AI assistants can generate images as part of larger tasks. TechCrunch notes that researchers from Leonardo.ai, which Canva bought in 2024, contributed to the model.
The business context explains the urgency. TechCrunch reports that Canva has relied on a mix of third-party image models, so its own model could reduce API spending and give it control over quality and speed. The Australian Financial Review reports that the company shelved some heavily promoted AI products in August over token costs and cut its revenue outlook, although outlets differ on how far its valuation has since moved.
The rankings deserve caution. DesignArena positions are preliminary and shift as votes accumulate, Canva's internal comparisons have not been published in detail, and claims about speed and cost have not been tested independently. What is clear is that a design platform with a very large user base now runs a model it controls.