Canva Unveils Advanced AI Model for Image Generation
Canva has introduced its inaugural artificial intelligence (AI) model, capable of generating images with resolutions reaching up to 2K and offering over 30 aspect ratios. Named Craft, this innovative tool has been meticulously developed and trained from the ground up by the company’s in-house research laboratory, which has now been recognised as the second most advanced globally in the field of image generation, according to recent assessments by DesignArena.
The company emphasises that its new AI model is the culmination of years of research and investment aimed at providing a deeper understanding of design. It promises results that align more closely with user imagination and affords the freedom to effortlessly edit and refine all creations.
Versatile and Intuitive Image Creation
In practical terms, Craft enables users to generate photorealistic images that seamlessly blend intricate details, depth, and smooth transitions, all while adhering to fundamental design principles such as alignment, spacing, defined edges, and precise text placement. Notably, the system processes user prompts in “any language” and can incorporate text in images across 55 different languages.
Canva asserts that these features render the tool intuitive to use for both professionals and novices alike, facilitating the creation of high-quality and faithful results. Users can produce a variety of materials, including advertisements, posters, brochures, presentations, photographs, social media posts, illustrations, digital art, portraits, multi-person scenes, landscapes, and conceptual compositions that merge realistic and stylised elements.
Innovative Two-Stage Image Generation Process
A distinguishing aspect of Craft is its unique two-stage architecture, which analyses user requests and devises a composition plan before generating the final image. In the initial phase, the model employs a secondary algorithm to identify essential elements, their relative positions and sizes, spatial relationships, visual hierarchy, and the texts to be included. During the second phase, a specialised rendering model transforms this schematic into a high-fidelity, high-resolution image.
The development team highlights that by explicitly modelling composition and typography prior to rendering the final image, Craft can produce complex designs with greater accuracy, including legible text, precise object placement, and coherent visual relationships. This architecture also facilitates more detailed control over individual elements, enabling specific alterations without unnecessarily affecting the remainder of the design.
Future Developments and Security Features
Canva asserts that this platform lays the technical groundwork for the development of advanced editing features, particularly those focused on incorporating reference images, which may be integrated into future iterations of Craft. The developers note that the training process was pivotal in achieving the model’s capabilities. They explain that the data utilised encompassed “millions of human-created templates from the content library.” Through these examples, Craft learned to understand the relationships among element distribution, visual hierarchy, spacing, and typography, rather than merely recognising the appearance of a finished image.
The model underwent fine-tuning through supervised adjustments and reinforcement learning techniques that incorporated feedback from professional designers at Canva. The goal was to enhance its ability to follow instructions accurately, generate text with precision, and maintain visual coherence in compositions.
Global Availability and Future Aspirations
In terms of security, Craft is supported by Canva Shield, an internal mechanism designed to manage aspects such as prompt and output moderation, alongside privacy controls. It also features tools for reporting potentially harmful or misleading content. These safeguards assess factors including brand protection, cultural context, and other elements that might lead to detrimental content.
Craft is now available globally to Canva users through the Dream Lab platform. The model can also be accessed via a public API, allowing developers and product teams to integrate it into their own tools and applications.
“This initial version of Craft is merely the beginning, and we are already working on faster Craft models. From this point onwards, we will continue to elevate the standards of quality, detail, and control, enhancing aspects such as handcrafted effects and technical drawings,” the company concluded.