Canva has released Canva Craft, its first in-house image generation model. Until now the platform relied on third-party models, and now it has its own option to compete with offerings from Google, OpenAI, ByteDance and Meta.
The key difference is the training data: instead of photographs, the model learned from layouts in Canva's library, such as posters, presentations, social posts, infographics and text-heavy menus.
Plan first, then the image
Generation happens in two steps. First the model parses the prompt and builds a plan: where elements sit, how large they are relative to each other, what matters most visually and what text should appear in the image.
Then the image is rendered from that plan at up to 2K resolution in one of 30 aspect ratios for different platforms. According to Canva, this yields accurate details, legible text and the ability to edit one fragment without touching the rest.
From image to editable layout
Since the model learned from designs, its output can be turned into a layered layout with Canva Magic, the tool that converts an image into an editable design. The model is available to all users in the Dream Lab tab and through the Canva API.
Support for generation inside third-party chatbots such as ChatGPT and Gemini via Canva's MCP server is also planned. Two lighter and cheaper versions, Craft-fast and Craft-ultrafast, are in the works. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the model.
Commentary
For website work, the interesting part is not yet another image generator but the link to layers and the API. A banner or cover produced by the model can be refined like a normal layout instead of being redrawn from scratch, and MCP access makes it possible to plug this into AI agent workflows. Details in the TechCrunch article.