Canva introduced Craft, an AI image generation model built to compete directly with offerings from OpenAI, Microsoft, and Google. The two-stage architecture plans a design before generating an image, enabling more accurate generation of complex visuals requiring legible text and precise object placement—a significant operational difference from single-pass competitors.

The model supports image generation up to 2K resolution across more than 30 aspect ratios. Canva trained Craft on millions of human-made templates from its content library—social media posts, presentations, infographics, and menus—creating a proprietary dataset advantage that competitors cannot easily replicate.

Canva Craft ranks fourth on Design Arena's leaderboard for AI design models, behind OpenAI's GPT-Image-2.5 variants and GPT-Image-2. The initial positioning places it within the top tier but acknowledges market leadership remains elsewhere.

The company built the model with enterprise requirements in mind. Craft includes robust input and output moderation, privacy controls, and reporting tools—addressing data governance and security concerns critical for corporate deployment. Canva plans to release faster variants (Craft-fast and Craft-ultrafast) and enable AI agents to integrate through the open Model Context Protocol, expanding the model's utility across workflows.

The launch extends Canva's AI roadmap. In April, the company introduced prompt-based editing features and integrated layers into AI-generated images, allowing partial edits without regenerating entire designs. These incremental capabilities stack toward a coherent vision: positioning Canva as an AI-native design platform rather than a design tool with AI bolted on.

Canva Craft is now available globally through the Dream Lab option in its design interface. The timing reflects intensifying competition in image generation—a category where architectural innovation (two-stage planning versus single-pass generation) may begin to matter more than raw scale as differentiation between models narrows.