Canva’s AI model Craft promises sharper AI images and editable designs

Canva has introduced Craft, its first image generation model developed internally by Canva Research. The company is making the model available to Canva users through Dream Lab, as well as to developers through its public API.

Craft is designed for visual work that often challenges image generators: layouts, typography, object placement, and images intended for further editing. Canva says the model can produce images at up to 2K resolution in more than 30 aspect ratios. Users can also enter a custom format for uses such as social posts, presentations, banners, and print materials.

Unlike many image models that move directly from a text prompt to pixels, Craft uses two stages. First, a planning model interprets the prompt and builds a description of the intended composition. This includes the objects in the image, their size and position, visual hierarchy, and any text that should appear. A separate rendering model then turns that plan into the final image.

According to Canva, separating planning from rendering should improve control over complex images. It is intended to help users make focused changes, such as adjusting one object or visual detail, without substantially changing the rest of the composition. The company also says the approach improves text rendering, an area where many AI image tools still produce errors.

A model trained for design work

Canva says it trained Craft with a stronger focus on human-made design assets than on photographs alone. Its training material includes millions of templates from Canva’s content library, including posters, presentations, social posts, menus, infographics, and typography-heavy layouts.

The company argues that this material gives the model a better grasp of practical design principles, including alignment, spacing, hierarchy, and the relationship between text and images. After initial training, Canva says it refined the model using supervised fine-tuning and reinforcement learning, with input from professional designers.

Canva reports that its designers and researchers reviewed more than 25,000 image pairs in comparisons with other models. The company also cites DesignArena rankings that place Craft particularly highly in categories for people and portraits, landscapes and nature, abstract images and patterns, and product images. These rankings are external benchmarks, while the company’s broader performance claims come from Canva.

Craft can generate images from prompts in any language and can create text in 55 languages, Canva says. The model is positioned for a range of work, from marketing materials and social graphics to editorial imagery, illustrations, portraits, product visuals, and architectural scenes.

Editable outputs and developer access

A central part of Canva’s pitch is the connection between generated images and its existing editing tools. Craft’s outputs are designed to work with Magic Layers, Canva’s feature for turning an image into separate editable elements. Canva says the model uses JSON boundary boxes to identify elements more accurately, which could make it easier to isolate and edit parts of a generated visual.

Developers can also access Craft through Canva’s public API for use in third-party products, such as marketing platforms, ecommerce services, ad builders, and print tools. Canva plans to add access through its Model Context Protocol server, or MCP, which would let AI assistants and automated workflows request images that can be used in Canva designs.

Ivan Mehta reports for TechCrunch that Canva also plans to support image creation through third-party chatbots, including services such as ChatGPT and Gemini. The report notes that researchers from Leonardo.ai, which Canva acquired in 2024, contributed to the new model.

Canva says it is developing two faster and lower-cost versions, called Craft-fast and Craft-ultrafast. The company says model distillation and a custom inference engine have reduced the number of steps needed to generate images. It has not yet provided a release date for those variants.

Sources

• Introducing Canva Craft – Canva Newsroom
• Canva launches Craft, its first image generation model, trained on human-designed templates, available to all users, with plans to support third-party chatbots – TechCrunch