Canva Craft: Canva launches its first image generation model, trained on its own layout models

Rather than photos, Canva dragged Craft into human-created templates in his library. The model plans the composition before generating, for legible texts and localized retouches.

Canva enters the field occupied by Google, OpenAI, ByteDance and Meta with Canva Craft, an image generation model designed by his research team, with the input of Leonardo researchers. bought in 2024. The operation is done in two stages: the model starts by translating the set into a plan (relative position of the elements, scales, visual hierarchy, text to be displayed), and then produces an image up to 2K in one of the 30 formats proposed for the different platforms.

Designs rather than photos

The training corpus is made up of posters, social media publications, presentations, infographics or menus from Canva templates. The images obtained can therefore be decomposed into editable layers thanks to Canva Magic, the home tool that converts an image to design. Canva also promises targeted editing without altering the rest of the image.

Craft is accessible to all users from the Dream Lab tab and will be accessible via the API; Canva plans to open it to ChatGPT or Gemini via its MCP server. Two more economical variants, Craft-fast and Craft-ultrafast, are announced. Having its own model must also reduce API fees paid to third-party suppliers, while the company has lowered its growth forecasts for 2026.