Plan first, render second
Most image models go straight from a prompt to pixels, and the familiar failures follow: text that dissolves into shapes, a headline placed where the brief did not put it, one element edited and the rest of the picture redrawn. Craft splits the job. In the first stage, a planning model interprets the prompt and writes a detailed plan of the composition: which elements appear, their spatial relationships, relative scale, visual hierarchy and any text that needs to appear in the image. In the second stage, a dedicated rendering model turns that plan into an image at up to 2K resolution, across more than 30 aspect ratios or a custom size.
That is a design decision aimed at design work rather than photorealism contests. Canva says Craft's training places a much greater emphasis on design: its training data includes millions of human-made templates from Canva's own content library — posters, social posts, presentations, infographics, menus and typography-heavy layouts. The argument is that a template teaches layout, hierarchy and typography in a way a finished photograph does not.
The output is also structured for what happens next. Craft's underlying structure suits Canva's Magic Layers, which transforms generated images into fully editable design elements, and Canva says JSON boundary boxing in the output makes that conversion more accurate and faster. A generated poster that arrives as separable elements is a different product from a flat image you can only re-prompt.
Two rankings, two different units
Canva's launch headline is a lab claim. Craft, Canva says, "makes Canva the world's second most advanced AI lab for image generation", as measured by rankings from DesignArena, which Canva describes as the largest crowdsourced assessor of AI design. Read the same announcement for the model's own result and a different number appears: Craft "debuted as the fourth most advanced image generation model on DesignArena", and Canva adds that this "also makes Canva Research the world's second most advanced lab for image generation".
Canva reports two different DesignArena results: Canva Research ranks second among image-generation labs, while Craft debuted fourth among individual image models. Those are different ranking units, so the figures are not necessarily contradictory. However, Canva's announcement does not explain how the lab ranking is calculated or how individual models contribute to it. DesignArena's own published methodology documents how its model leaderboards are computed — crowdsourced pairwise preference votes, fitted with the Bradley-Terry model — and says nothing about a lab ranking. The second-place lab claim should therefore not be presented as Craft's individual ranking, and a reader who takes only the headline away will quote the wrong number.
The category figures underneath are stronger and more specific. According to Canva's reporting of DesignArena results, Craft ranked first in People & Portraits among 87 models and first in Landscapes & Nature among 85. Canva also reports second place in Abstract & Patterns and fourth in Products, without specifying category denominators in the announcement. All of these are Canva-reported results from a crowdsourced preference ranking: they measure which output voters pick, not accuracy against a fixed test. Preference rankings are useful evidence about taste. They are not a benchmark with a right answer.
Canva cites external crowdsourced preference rankings and a separate internal designer evaluation. The announcement does not provide the underlying DesignArena votes, a reproducible independent test of Craft's design-specific claims, or numerical API pricing and speed comparisons.
Whose judging, whose savings
The other evaluation in the announcement is internal. Canva says its professional designers, working alongside dozens of researchers, assessed more than 25,000 pairs of images in head-to-head comparisons with leading models during development, and that when automated evaluations and human judgments differed, Canva sided with the designer. That may be a useful internal product-development process, but it is not an independent benchmark. The announcement does not provide enough methodological detail to assess sampling, blinding, model selection or the distribution of results — and the 25,000 figure counts image pairs evaluated, not wins for Craft.
The business reason for building the model is stated plainly enough by what Canva has done before. As TechCrunch notes, Canva has used a range of other companies' image models inside its platform for years, and in 2024 it acquired Leonardo.ai, whose researchers TechCrunch reports contributed to Craft. TechCrunch's read is that an in-house model should cut the API costs Canva pays others and give it direct control over quality and speed. That is TechCrunch's inference about Canva's costs, not a demonstrated saving, and any saving would be Canva's first. Whether any of it reaches users depends on the two variants Canva announced without numbers. Canva says Craft-fast and Craft-ultrafast are in development and will offer lower-cost generation relative to leading frontier models. The announcement provides no numerical cost comparison, API prices, benchmark methodology or release dates for these variants. Lower production costs for Canva would not automatically mean lower prices for developers, and as of the 22 October announcement there is no published Craft API price to compare against the models it replaces.
What to test first
Canva says Craft is available to users through Dream Lab, and to developers through its public API; support through the Canva MCP server is listed as coming soon. The announcement does not establish that generation is unlimited or free across account tiers. Open access makes the useful test cheap, though. Give Craft the job image models usually fail: a poster or menu with several lines of real text, a clear hierarchy and one element you will want to move afterwards. Check whether the text survives, whether the layout follows the plan you described, and whether Magic Layers gives you elements you can actually edit. If those three hold, the two-stage approach has earned its claim. The lab ranking can be left where Canva put it, in the headline.
Sources
- [1] Canva, "Introducing Canva Craft" (Canva Newsroom, 22 October 2026) — primary source, read in full for this article: first in-house model, two-stage planning/rendering architecture, up to 2K and more than 30 aspect ratios, template-based training, Magic Layers and JSON boundary boxing, text in 55 languages, DesignArena figures (lab #2 claim, model debut #4, category ranks #1 of 87 and #1 of 85, #2 Abstract & Patterns and #4 Products with no denominators stated), the 25,000-pair internal designer evaluation, availability in Dream Lab and via public API, MCP listed as coming soon, and the Craft-fast / Craft-ultrafast variantsRead source
- [2] TechCrunch, "Canva releases its first image generation model" (22 October 2026) — independent report of the launch: the two-stage description, training on designs rather than photographs, availability to all users via Dream Lab plus API and planned MCP support, the inference that an in-house model should reduce the API costs Canva pays others and give it better quality and speed control, Leonardo.ai researchers' contribution after the 2024 acquisition, and the two cheaper variants in developmentRead source
- [3] DesignArena, "Design Arena Methodology" — the ranking operator's published methodology, read for this article: model leaderboards are computed from crowdsourced pairwise preference votes using the Bradley-Terry model. It documents the model-ranking calculation only; it does not explain how a lab ranking is calculatedRead source