Canva launches Craft, its first in-house image model, in Dream Lab and API
Canva unveiled Craft, the first image foundation model trained from scratch by Canva Research. A two-stage plan-then-render design targets up to 2K images across 30+ aspect ratios and on-image text in 55 languages. It is available globally in Dream Lab and via the Public API. DesignArena ranking claims are attributed to Canva.
Canva’s newsroom introduced Canva Craft as the first state-of-the-art image generation model built and trained from scratch by Canva Research, aimed at design work. The company says Craft is available now in Canva and through the public API, generating images at up to 2K resolution across more than 30 aspect ratios, with an emphasis on control and editable outputs.
Unlike models that map prompts straight to pixels, Craft uses a two-stage architecture: an image-planning model drafts composition (elements, spatial relations, scale, hierarchy, and required text), then a rendering model turns that plan into a high-fidelity image. Canva says this improves complex layouts, legible text, and object placement, and feeds Magic Layers for editable design elements. Prompts may be in any language; on-image text covers 55 languages.
Availability: Canva says Craft is available starting today to all users globally via Dream Lab (bottom-left panel), plus the Public API and a Canva MCP Server (coming soon). Safety is described under Canva Shield with input/output moderation. Ranking claims (attribute to Canva): by DesignArena crowdsourced scores, Canva Research is called the world’s second-most-advanced image-generation lab and Craft the fourth-ranked model, with subcategory #1/#2 claims for portraits and landscapes—not independent Shift evaluations.
Sources
- Introducing Canva Craft — Canva