Canva launches Canva Craft, its first in-house AI image generation model

Canva has unveiled Canva Craft, its first proprietary image generation foundation model developed by Canva Research. Unlike many competitors that focus on photographic realism, Craft is specifically engineered for design tasks such as creating posters, social media graphics, presentations, and infographics.

The model utilizes a unique two-stage architecture. In the first stage, it interprets user prompts to create a detailed compositional plan, defining object placement, scale, visual hierarchy, and text requirements. In the second stage, a rendering model converts this plan into a final image with up to 2K resolution and support for over 30 aspect ratios.

To achieve high precision in layout and typography, the model was trained on millions of human-designed templates from Canva’s own library rather than primarily on photographs. This approach allows for more legible text and the ability to perform targeted edits on specific elements without altering the entire image. Furthermore, generated outputs can be converted into editable, multi-layered designs using Canva’s Magic Layers technology.

Canva Craft is currently available to users via the Dream Lab interface and through a public API. The company also plans to integrate the model into third-party chatbots like ChatGPT and Gemini using the Model Context Protocol. To improve cost-efficiency, Canva is also developing two lighter versions, titled Craft-fast and Craft-ultrafast.