Canva launches an AI image model built to get ad text right, and the agent integration it promised is still coming

Canva released Craft, its first in-house AI image-generation model, on October 22 — built to plan a picture's layout and text before it renders a single pixel, instead of generating pixels first and hoping the words come out readable. For a design platform whose core users are marketing teams, that ordering is the whole pitch: the model is aimed squarely at the broken-text problem that has kept AI image generators out of finished campaign work.

What changed

Most image generators produce a picture directly from a prompt, then leave typography and object placement to chance. Craft splits the job into two stages. A planning model first works out what should appear in the image, where each element sits, how large it is relative to everything else, and what any on-image text should say. A separate rendering model then turns that plan into a finished image, up to 2K resolution, in more than 30 aspect ratios. Canva says the sequencing is what lets Craft hold a layout together and keep text legible, rather than scrambling letters the way diffusion models often do when asked to render words.

The training data is the other half of the story. Canva said it trained Craft primarily on millions of its own templates, spanning event posters, pitch decks, menus, infographics and other text-dense layouts, rather than on photographs, which is what most general-purpose image models lean on. Canva's designers and researchers then ran more than 25,000 head-to-head image comparisons against rival models during evaluation, with Canva saying it deferred to its designers' judgment whenever automated scoring and human reviewers disagreed, according to Unite.AI. Text generation works in 55 languages.

Who it affects

Craft is live now inside Canva's Dream Lab for every user, and developers can already call it through Canva's public API, which TechCrunch reports opens a path for other marketing software to plug image generation into ad builders and ecommerce tools. That is the audience this matters most to: software marketing teams that lean on Canva for campaign graphics, social creative, event collateral and pitch decks, and who have mostly treated AI image output as a draft to be redone by hand once the text came out wrong.

One thing Craft is not yet is omnipresent. Canva has said connecting Craft to third-party AI assistants through Canva's MCP server is still coming, not already shipped, so marketers should not assume the model is reachable from ChatGPT or Gemini today just because Canva has talked about agent access. Cheaper, faster variants called Craftfast and Craft-ultrafast are also described as upcoming releases, with no date attached yet.

The launch also lands while Canva is under some pressure to show its AI investment pays off: the company was on pace for $4 billion in revenue last year, but after it cut its 2026 growth forecast, its valuation reportedly slipped from $42 billion to $35 billion, per outside analysts cited by TechCrunch. Owning its own foundation model, rather than licensing one, gives Canva more control over the inference costs behind a platform used by hundreds of millions of people — which is also a reason to expect the free and paid tiers to keep shifting as that economics settles.

What a software marketer should do differently

Treat the DesignArena ranking and Canva's own benchmarks as a starting point, not a verdict. Canva's own numbers are self-reported, and a leaderboard placement says nothing about how a model handles your brand's specific fonts, claims language or regional variants. Run your own test before you let Craft into a live campaign.

• Generate a text-heavy asset — a promo banner or pricing graphic, not a clean hero shot — and check whether every word survives unedited
• Test a multi-format run: the same concept resized across a paid social ad, a landing page header and a slide, to see how the layout plan holds up
• Push a non-English version and confirm the 55-language claim holds for the languages your campaigns actually run in
• Open the output in Magic Layers and see how many of the elements come through as separately editable, versus flattened

What to watch next

Two things will tell you whether Craft is more than a launch-day demo. The first is whether the promised MCP connection actually ships, and whether that makes Craft something your own product or martech stack can call directly, the way Canva's API already allows. The second is pricing: Canva hasn't said what Craftfast and Craft-ultrafast will cost relative to Craft itself, or whether owning the model end to end eventually shows up as a lower-cost tier instead of just better margins for Canva. Until both of those resolve, the practical move is to pilot Craft on low-stakes creative and keep your editable source files as the fallback, not the exception.