Canva releases Craft, its first in-house AI image model

Canva has released Craft, its first proprietary image-generation model, built and trained in-house by Canva Research. The company made it available worldwide to Canva users, positioning the model around design work rather than general-purpose image creation. Craft is available through Canva’s Dream Lab and a public developer API. It can generate images at up to 2K resolution in more than 30 aspect ratios, accepts prompts in any language and supports generated text in 55 languages. The model splits composition from rendering. A planning model first interprets a prompt to map elements, their placement, relative scale, visual hierarchy and text; a separate model then produces the finished image. Canva says behind-the-scenes prompt improvement is also used before generation, with the aim of producing more complete layouts and allowing individual elements to be changed without remaking the entire image. That structure builds on Canva’s Magic Layers technology. JSON bounding boxes help turn a generated image into separate editable design elements, rather than leaving the result as a single flat image. Craft is intended to provide accurate text rendering and precise object placement, two areas that matter for posters, presentations, product imagery and other typography-heavy work. Canva trained Craft on millions of human-made templates in its own content library, including social-media posts, infographics, menus and posters. The training material also included licensed and public databases, as well as data from users who opted in. Professional designers contributed feedback during supervised fine-tuning and reinforcement learning, and Canva designers assessed more than 25,000 image pairs against leading models. The company said it favoured designer judgment when it conflicted with automated scoring. Canva’s earlier creative-AI product, Magic Media, debuted in 2023; the development of Craft was also aided by Canva’s acquisition of Leonardo.Ai and an expansion of its internal AI research, according to CNET. On DesignArena, Craft ranked fourth overall at launch, behind OpenAI’s ChatGPT Images but ahead of Google’s Nano Banana and Microsoft’s MAI-Image. Canva said it ranked first for People and Portraits among 87 models and for Landscapes and Nature among 85, while placing second in Abstract and Patterns and fourth in Products. CNET found it performed strongly on portraits and product-focused images but trailed in the marketing category. Canva said the ranking placed its research group second among AI labs for image generation. More than 250 million people use Canva and create more than 400 designs a second, according to the company. The launch comes as AI image training remains under scrutiny. The U.S. Copyright Office has said generative-AI systems draw on massive data sets that can include copyrighted material, and that training on such work has prompted concerns over consent and competing content. The office reported dozens of pending U.S. lawsuits involving copyright and generative-AI training. Canva Shield provides input and output moderation, privacy and safety controls, continuous testing and reporting tools for Craft. Canva plans faster Craft-fast and Craft-ultrafast variants, improvements to fine print and technical drawings, and future advanced editing and reference-image features. It also plans to connect the model to AI assistants and automated workflows through the Canva MCP Server.