Canva Research unveiled Canva Craft on October 22, 2026, the company’s first in-house image generation model and a pointed break from the industry’s photo-trained orthodoxy. Craft was trained on millions of Canva’s own human-made templates rather than on scraped photography, and it shows in what the model is good at: readable text, stable layouts, and multi-element compositions that survive editing.

The architecture is two-stage. A planning model drafts a compositional plan; a rendering model then produces the final image at up to 2K across more than 30 aspect ratios. That deliberation costs time (roughly 50 to 60 seconds end-to-end), but it buys something detector-trained diffusion models still fumble: typography that doesn’t melt and hierarchy that doesn’t collapse. Craft generates text in 55 languages and accepts prompts in any language.

Preliminary Design Arena results placed Craft fourth on the image-generation leaderboard, ahead of models from Google, Microsoft, Meta, xAI, and ByteDance, with only OpenAI’s lab ranking higher. By category, Canva reports Craft ranks first globally for People & Portraits and for Landscapes & Nature, second for Abstract & Patterns, and fourth for Products.

What makes Craft structurally interesting isn’t the leaderboard, it’s the handoff. Outputs flow into Canva Magic, which converts a generated image into a layered, editable design file. Generation stops being a terminal act and becomes the first step of a workflow that already owns the small-business visual stack, from Instagram tiles to menu boards to event posters.

The timing isn’t subtle. One day earlier, Google turned Gemini into a marketing hub pointed at small businesses. Two of the largest horizontal software companies are now racing to be the default surface where a founder generates a week of branded content before coffee.

Developer access runs through the Connect API, still labeled preview, and consumes AI credits. Canva says cheaper variants, Craft-fast and Craft-ultrafast, will follow.

A model trained on design templates competing against models trained on the open web is a legible bet about where generative imagery is actually used. Most of it, in the end, is marketing.

Sources