Canva Unveils Canva Craft, an In-House AI Image Model That Plans Layouts Before Generating
Canva has introduced Canva Craft, a proprietary image generation model featuring a two-stage architecture that plans layouts prior to generation and delivers images at up to 2K resolution.
Canva is expanding its footprint in generative technology with the launch of Canva Craft, its first in-house AI image generation model designed to compete directly with offerings from OpenAI, Microsoft, and Google.
The new model relies on a two-stage architecture that maps out and plans a design layout before executing the final image generation. According to Canva, this approach delivers "greater precision, consistency and control," making it particularly adept at handling complex compositions that require legible text and precise object placement.
High-Resolution Output and Flexible Formats
Canva Craft is capable of generating images at up to 2K resolution across more than 30 aspect ratios.
In terms of creative range, the model is built to produce marketing graphics and social media content, as well as digital art, realistic scenes, and environmental imagery.
To power these capabilities, Canva trained the model on millions of human-made templates sourced from its own design library. This training dataset includes everyday graphic assets such as social media graphics, infographics, menus, and business presentations.
Benchmark Performance and Ecosystem Integration
Following its release, Canva Craft entered Design Arena's crowdsourced benchmark leaderboard for AI image generation models in fourth place. It currently trails three models from OpenAI: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare, and GPT-Image-2.
The launch of Canva Craft follows a broader push by the platform to integrate native AI tooling into its creative suite. In April, Canva introduced prompt-based editing tools that allow users to generate and modify designs through text descriptions, as well as a layering feature that lets creators edit isolated sections of AI-generated images while preserving the rest of the canvas.
