Canva has released Craft, its first image generation model trained entirely in-house. The system was built by Canva Research using millions of design templates instead of standard photographs. This approach allows the tool to plan composition, proportions, and text before rendering images at up to 2K resolution. The output supports more than 30 aspect ratios and renders text in 55 languages. Users can separate results into editable layers via Magic Layers, a feature that distinguishes the workflow from generic image generators. On the DesignArena benchmark, the model ranked fourth overall and first in specific categories like portraits and landscapes. Leonardo.ai researchers, acquired by the company in 2024, also contributed to the project. Two cheaper variants named Craft-fast and Craft-ultrafast are expected to arrive soon. The platform is currently accessible through Dream Lab and a public API, with plans to integrate Model Context Protocol for chatbots like ChatGPT and Gemini later. This move gives Canva greater control over generation speed and quality while reducing reliance on external APIs. Lower costs for the company and a tailored output format are the primary drivers behind this development.
- Training data consists mainly of design templates rather than photos
- Leonardo.ai researchers contributed to the model’s development
- Craft-fast and Craft-ultrafast variants are scheduled for release
