Canva's Craft Plans the Poster Before It Paints It

Canva shipped its first in-house image model this week, and the pitch is a sentence most image models cannot say: it writes out a plan before it draws anything.

The model is called Craft. Developed by Canva Research, trained on the company's own library of human-made templates, available to every Canva user and through the public API. It generates at up to 2K in more than 30 aspect ratios and renders text in 55 languages.

The two-stage design is the substance. A planning model reads the prompt and writes a detailed composition plan: which elements appear, how they sit relative to each other, their relative scale, the visual hierarchy, and any text the image needs. A separate rendering model turns that plan into pixels. Because the plan is structured, Canva says users can change a single element without the rest of the image moving.

Anyone who has generated a poster with a general image model knows why this matters. You ask for a headline, three icons, and a product shot in the lower right. You get something that looks like all three, arranged almost correctly, with the headline spelled wrong and the icons floating at slightly different heights. The composition was never really a decision the model made. It was a side effect of sampling.

What the training data changes

Most image models are trained mostly on photographs. Craft's training set includes millions of templates for posters, social posts, presentations, infographics, menus, and layouts where text is the point rather than a decorative element.

That is a real difference in prior. A photograph-trained model has seen billions of images where text appears on signs, packaging, and screens, usually small, often blurry, rarely the main subject. A template-trained model has seen millions of images where the typographic hierarchy is the entire reason the image exists. If your job is making a menu or a course announcement, those are not the same starting point.

The rankings, read carefully

Canva says Craft debuted fourth on DesignArena, which it describes as the largest crowdsourced assessor of AI design. It also says Craft ranked first in People and Portraits (out of 87 models) and first in Landscapes and Nature (out of 85), second in Abstract and Patterns, fourth in Products. And it says this makes Canva Research the world's second most advanced lab for image generation.

That last sentence is doing more work than the data supports. Fourth overall plus two category wins is a strong debut for a first model. Calling the lab second in the world is a claim about labs, not about a model, and DesignArena is not the same arena as the text-to-image leaderboards most people quote. Treat the two category wins as the meaningful part. Portraits and landscapes are exactly where template-trained models might be expected to lag, so winning them is the surprise.

There is also a detail in Canva's methodology worth quoting, because it is unusual: designers worked with researchers to assess more than 25,000 pairs of images in head-to-head comparisons against other models, and when automated evaluations disagreed with human judgment, the company sided with the designer. Whether you trust the resulting numbers or not, that is a stated tie-breaking rule, which is more than most launch posts offer.

Why the planning stage is also a speed trick

Beyond quality, the architecture buys Canva something a straight diffusion model cannot easily get: the expensive reasoning happens once, in text, and the rendering stage can be distilled hard.

Canva says it used model distillation to cut the inference steps to a fraction of the original, and built a custom inference engine for the planning stage that makes Craft an order of magnitude faster than its first internal version. It also says two variants are coming, Craft-fast and Craft-ultrafast, at a fraction of the cost of leading frontier models.

The scale context explains the urgency. Canva says more than a quarter of a billion people use the platform, creating over 400 designs every second. At that volume, a model that is 20% better but twice as expensive per image is a business problem, not a feature.

Where it stops

The list of what Craft does not do yet is more informative than the feature list. Canva says future versions will handle fine print and technical drawings better, and that more advanced editing and reference image features will follow.

That means no reference images for now. No "put this product in this scene." No matching an existing brand asset. For a model sold into marketing and e-commerce, the absence of reference image support is the single biggest limitation, because most commercial work starts from something that already exists: a product photo, a logo, a person.

The bridge Canva is offering instead is Magic Layers, which turns a generated image into editable design elements inside Canva. That is a smart answer to a real problem. If the model cannot take your product shot as input, at least the output can be taken apart and reassembled around it. It is a workaround, and it is the workaround most designers already do by hand in three other tools.

Who this is for

If you already make things in Canva, Craft is close to a free upgrade, and the languages and aspect ratios mean it will fit most social and print formats without a resize pass. The MCP server matters for a different reason: it lets an agent generate an image as one step in a larger task, which is how a lot of marketing work is actually being assembled now.

There is a second-order effect of the planning stage that Canva does not advertise and that matters more than any leaderboard position: a plan is editable in a way a latent is not. Once the composition exists as a structured description, a brand's rules can be applied to that description rather than to the pixels. Logo in the lower right, minimum cap height, two brand colours plus black. Those are constraints a planning model can be told about. A diffusion model can only be asked politely. That is the same argument Adobe has been making about template-driven generation, and it is the reason a design company shipping a model is a different event from a lab shipping one. Canva's moat is not the checkpoint. It is the million templates and the brand kit that every one of its users already filled in. If you need photoreal product renders, or you need to place a specific object into a specific scene, this is not the tool yet. Canva says reference images are coming. Until they ship, Craft is strongest exactly where its training data is strongest: layouts where the words are the design, and the layout has to be right the first time.

Which, given that Canva's entire business is layouts, is not a coincidence. It is the most defensible place for a design company to plant a model.