The short version
- Canva has released its first proprietary AI image model, named Craft, which utilizes a two-stage architecture to plan designs prior to generation.
- The system is trained on millions of human-created templates and currently ranks fourth on independent benchmarking leaderboards behind OpenAI models.
- The tool is available globally to all users and includes safety controls, with plans for faster variants and agent integration via open protocols.
Canva has introduced its first in-house artificial intelligence image generation model, marking a significant shift in the company’s strategy to compete directly with major technology firms such as OpenAI, Microsoft, and Google. The new system, branded as Canva Craft, represents a departure from relying solely on third-party AI providers by offering a native solution built specifically for the platform’s design ecosystem. This launch underscores the growing pressure on creative software companies to develop proprietary AI capabilities that can offer greater control and consistency than general-purpose models.
The core innovation of Canva Craft lies in its two-stage architecture. Unlike many existing image generators that produce visuals directly from text prompts, this model first plans the design before generating the final image. This preliminary planning phase is intended to provide users with higher precision and better control over complex compositions. According to company announcements, this approach helps address common challenges in AI-generated graphics, such as ensuring legible text and precise object placement within a layout.
Canva states that the model was trained on millions of human-made templates sourced from its own content library. These training materials include a wide variety of design assets such as social media posts, business presentations, infographics, and restaurant menus. By leveraging this specific dataset, the company aims to tailor the AI’s output to the practical needs of marketers and designers who require structured, professional-looking results rather than abstract artistic interpretations.
In terms of technical specifications, Canva Craft supports image generation at resolutions up to 2K. It also accommodates more than thirty different aspect ratios, allowing for flexibility across various digital platforms and print formats. The model is designed to handle a broad range of visual tasks, including the creation of realistic scenes, digital art, environmental imagery, and standard marketing content. This versatility suggests an attempt to consolidate multiple design workflows into a single AI-driven interface.
Independent benchmarking provides some context for the model’s current performance capabilities. On Design Arena, a crowdsourced leaderboard that ranks AI image generation models, Canva Craft currently holds the fourth position. It trails behind three versions of OpenAI’s GPT-Image-2.5 series, including Sunburst, Flare, and the base GPT-Image-2 model. This ranking indicates that while the new tool is competitive, it has not yet surpassed the leading general-purpose image generators in overall quality or versatility as measured by community benchmarks.
The release of Canva Craft builds upon a series of AI enhancements the company has rolled out earlier this year. In April, Canva introduced prompt-based editing features that allow users to create or modify designs through natural language descriptions. The platform also added layer support for AI-generated images, enabling editors to adjust specific parts of a visual without altering the entire composition. These previous updates laid the groundwork for the more sophisticated planning and generation capabilities found in the new model.
Security and compliance remain central to the deployment strategy. Canva emphasizes that the model includes robust input and output moderation systems, along with privacy safeguards and reporting tools. These measures are designed to prevent the generation of harmful or inappropriate content while protecting user data. As AI integration deepens within creative workflows, such controls become increasingly important for enterprise clients and individual users concerned about brand safety and regulatory compliance.
Looking ahead, Canva plans to expand the accessibility and speed of its AI offerings. The company intends to allow external AI agents to connect with Canva Craft through the open Model Context Protocol, facilitating broader automation possibilities. Additionally, more efficient versions of the model, labeled Craft-fast and Craft-ultrafast, are scheduled for future release. These updates aim to reduce processing times while maintaining quality, addressing one of the primary friction points in current AI design tools.
Canva Craft is now available to all users worldwide. To access the tool, individuals can navigate to the three-dot menu option on the left-hand side of the Canva interface and select Dream Lab. This immediate global rollout suggests confidence in the model’s stability and utility. As the competition among AI image generators intensifies, Canva’s focus on design-specific planning and template-based training may offer a distinct advantage for users prioritizing structured visual communication over open-ended artistic generation.
The introduction of this proprietary model signals a broader trend in the tech industry where platforms are moving away from dependency on external AI providers. By developing its own engine, Canva seeks to differentiate its product through specialized features that align closely with the needs of graphic designers and marketers. Whether this two-stage approach will significantly outperform competitors in real-world usage remains to be seen, but it represents a notable evolution in how AI is integrated into creative software.
Sources behind this briefing
Go to the original reporting
- The Verge↗Canva’s first AI image model plans designs before generating them
