Plan first, then render
Canva released Canva Craft on October 23, the first foundation model trained from scratch by Canva Research, its in-house AI lab. Unlike most image models, which generate pixels directly from a prompt, Craft splits the job in two. In the first stage a planning model turns the prompt into a detailed composition plan listing which elements appear, how they relate in space, their relative scale, the visual hierarchy and any text that needs to appear. In the second stage a rendering model follows the plan to produce an image at up to 2K resolution in more than 30 aspect ratios.
Canva's argument is that working out composition and typography first gives complex layouts clearer text and more accurate object placement, and makes it easier to change one element without disturbing the whole image. The same structure fits Canva's Magic Layers, which can split a generated image into an editable, layered design. Other official specs: prompts can be written in any language, and text in images is supported in 55 languages.
Training data that leans toward design, not photos
A large share of Craft's training data is millions of human-designed templates from Canva's content library, including posters, social posts, presentations, infographics, menus and typography-heavy layouts. Pre-training was followed by supervised fine-tuning and reinforcement learning. For internal evaluation, Canva's professional designers worked with dozens of researchers to compare more than 25,000 image pairs against leading models, and where automated evaluations and human judgment disagreed, the designers won.
On results, Canva cites the DesignArena crowdsourced rankings: Craft debuted fourth overall, which makes Canva Research the "second most advanced lab" for image generation on that board. In subcategories it ranks first for people and portraits (87 models) and landscapes and nature (85), second for abstract and patterns and fourth for products. The Verge adds that the three models ahead of it overall are all OpenAI GPT-Image models.
How to use it, and how cost is kept down
Craft is available now to all Canva users worldwide via Dream Lab in the bottom-left panel. Developers can call it through a public API, and access for agents through Canva MCP is marked "coming soon". Canva says more than 400 designs are created on the platform every second; to handle that load, the team used model distillation to cut inference steps to a fraction of the original model's and wrote a dedicated inference engine for the planning stage, making it an order of magnitude faster than the first internal version. Two cheaper variants, Craft-fast and Craft-ultrafast, have not been released yet.
For the design-tool market, this means Canva is no longer just a wrapper around outside image APIs. DesignArena, though, is a crowdsourced preference vote: the ranking reflects which image people found better looking, and claims about legible text and precise editing still need more independent testing.
via: Canva Newsroom: Introducing Canva Craft, The Verge: Canva's first AI image model plans designs before generating them