Canva launches Craft, its first AI image model that plans designs before generating

Canva's first in-house image model

Canva has released its first in-house AI image generation model, a step that puts the design platform in direct competition with the image generators offered by OpenAI, Microsoft and Google. According to The Verge, the model is named Canva Craft and was announced on Thursday.

Craft's distinguishing feature is a two-stage architecture. Instead of producing an image in a single pass, the model first works out a plan for the design and then generates the picture. Canva says this planning stage gives the model greater precision, consistency and control than a conventional approach.

What Craft can do

The model generates images at resolutions of up to 2K and supports more than 30 aspect ratios, The Verge reports. Its intended uses extend beyond the social media posts and marketing content Canva is known for, covering realistic scenes, digital art and environmental imagery as well.

Canva claims the two-stage system is especially effective at complex images that require legible text and precisely placed objects — two areas where general-purpose image generators have historically struggled.

Trained on Canva's own templates

Rather than drawing only on broad image datasets, Canva says it trained Craft on millions of human-made templates from its own content library, including social media posts, presentations, infographics and menus. The underlying bet is that a model steeped in real design layouts will be better at producing design-ready output.

The launch builds on a broader AI push by the company. As The Verge notes, Canva introduced prompt-based editing in April, letting users create or change a design simply by describing it, and added layers for AI-generated images so that one part of a picture can be edited while the rest stays untouched.

Where it stands on the leaderboard

Early benchmark results are respectable but not leading. Canva Craft currently ranks fourth on Design Arena, a crowdsourced leaderboard for AI design models, placing it behind three OpenAI entries: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2.

Safety, agents and faster variants

Canva says the model ships with robust input and output moderation, privacy and safety controls, and reporting tools. The company also has two developments on its roadmap: support for connecting external AI agents to Craft through the open Model Context Protocol, and more efficient versions of the model called Craft-fast and Craft-ultrafast.

Craft is available now to all users globally. According to The Verge, it can be tried through the three-dot option in Canva's left-hand menu by selecting Dream Lab.

Why it matters

Most design and productivity platforms have so far bolted on third-party AI models rather than building their own. Canva developing Craft in house, and training it on its own library of human-made templates, signals that application companies increasingly see model ownership as strategically important rather than optional.

The technical claim matters too. Legible text and precise object placement are the biggest gaps between typical AI image output and something a business can actually publish. If Canva's plan-then-generate approach delivers on those fronts, AI image generation moves closer to being genuinely useful inside real design workflows, where Canva already has distribution.

The fourth-place Design Arena ranking shows OpenAI still leads on raw benchmark performance, but Canva does not necessarily need to win the leaderboard to win its market. Its users need usable designs, not gallery pieces, and a model tuned for layouts, aspect ratios and text may serve them better than a general-purpose rival. The planned Model Context Protocol support also positions Craft for a near future in which AI agents, not just human designers, commission visual content.