Canva released Canva Craft on Thursday, its first in-house image generation model. It works in two stages: first it builds a plan for the design — element placement, relative scale, visual hierarchy and the text content — and only then renders an image at up to 2K in more than 30 aspect ratios. The model is already available globally to all users.
Key takeaways
- Two-stage architecture: the design plan first, then the render
- Up to 2K resolution and more than 30 aspect ratios
- Trained on millions of human-made templates, not photographs
- Fourth on the Design Arena leaderboard, behind three OpenAI models
- Available in the Dream Lab tab and through the Canva API
Templates instead of photos
The training material was not photographs but designs from Canva's library: posters, social posts, presentations, infographics, menus and typography-heavy layouts. The company says this makes the model more accurate on object detail and legible text, and lets a user fix part of an image without touching the rest.
The first stage settles element placement, relative scale, visual hierarchy and what text should appear. Only the second stage renders. Canva notes the model ships with input and output moderation plus reporting tools.
The output can also be converted into a layered design through Canva Magic, the company's image-to-design converter. Researchers from Leonardo.ai, the generative platform Canva acquired in 2024, contributed to the model.
Fourth place, with three OpenAI models ahead
On Design Arena, a crowdsourced benchmark for image generators, Craft sits fourth — behind GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2. Canva also plans two cheaper and faster versions, Craft-fast and Craft-ultrafast, plus access for third-party chatbots including ChatGPT and Gemini through its own MCP server.
Why it matters
Splitting planning from rendering answers a specific weakness of image generators: a correct picture with incorrect text and a broken layout. For Canva, though, the bigger point may be that it stops paying for someone else's API and regains control of quality, cost and response time. Fourth place also shows an in-house model does not have to lead its category to pay for itself for the platform that owns it.
What's next
- Craft-fast and Craft-ultrafast are announced as cheaper variants, with no date given
- The model is to reach third-party chatbots through Canva's MCP server
