Full Breakdown
Canva Unveils In-House AI Image Model “Craft”
By Drooid · · How we work
Launch of Canva Craft
On October 22, 2026, Canva announced the release of Canva Craft, the company’s first foundation model built and trained from scratch by its internal AI unit, Canva Research. The model is offered globally through Canva’s Dream Lab interface and via a public API for developers. Craft employs a two-stage architecture that first creates a detailed compositional plan and then renders a high-fidelity image, supporting resolutions up to 2 K and more than 30 aspect-ratio options.
Background and AI Lab Development
Canva’s AI efforts date back to 2017 with machine-learning tools for search and recommendation. The firm expanded its capabilities through acquisitions of Vienna-based Kaleido (2021) and Leonardo (2024), later consolidating its research teams into Canva Research in 2025. Additional hires from MangoAI and Simtheory brought expertise in reinforcement learning and AI agents. By the time of the Craft launch, Canva Research comprised over 160 scientists and engineers within a broader AI organization of roughly 400 staff.
Model Capabilities and Performance Metrics
Craft is trained on “millions of human-made templates” from Canva’s library, covering posters, social posts, presentations, infographics, menus, and other typography-rich layouts. Canva reports that the two-stage design enables accurate text rendering in 55 languages and precise object placement, addressing limitations of models trained primarily on photographs. In internal head-to-head evaluations of more than 25,000 image pairs, Craft ranked fourth overall on DesignArena’s leaderboard, holding the top global rank for People & Portraits and Landscapes & Nature, second for Abstract & Patterns, and fourth for Products. Canva also notes that its platform serves more than a quarter of a billion users, generating over 400 designs each second.
Official Statements and Safety Measures
The company emphasized that brand safety, cultural context, and potential harmful content are evaluated before any model release. Canva also highlighted that designers’ feedback guided supervised fine-tuning and reinforcement learning to improve instruction following and typography accuracy.
Future Plans and Upcoming Variants
Canva indicated that faster versions—Craft-fast and Craft-ultrafast—are in development to deliver high-quality generation at lower computational cost. The firm plans to expand integration with AI agents via the Model Context Protocol and to add more efficient inference engines for the planning stage. These enhancements aim to broaden Craft’s use in marketing platforms, e-commerce sites, ad builders, and print services.