Canva announced this week Canva Craft, its first image generation model developed and entirely trained by its own research team. Unlike other tools in the space that focus on photorealism or artistic styles, Craft was designed specifically for the context of graphic design: it understands visual hierarchy, typographic composition, and the logic behind professional templates —from social media posts to presentations and infographics.
What makes Craft architecturally interesting is not just what it produces, but how it produces it. Conventional image generation models —such as DALL·E, Midjourney, and Stable Diffusion— map a text prompt directly to pixels without any intermediate layer of reasoning about image structure. Craft, on the other hand, employs a two-stage architecture: first, a planning model interprets the prompt and writes a detailed composition plan —which elements will appear, their spatial relationships, relative scale, visual hierarchy, and any text that needs to be rendered. Only then does a second stage translate that plan into a high-fidelity image at up to 2K resolution, supporting over 30 aspect ratios.
This separation between planning and rendering is not merely an academic curiosity. It translates into results that matter practically: legible text positioned correctly in the layout, objects placed in exact positions, and consistent visual relationships. In traditional models, text generated within images is notoriously problematic —letters are distorted, words are partially erased or entirely incorrect. Craft addresses the problem at its root, treating typography as a structural component planned before rendering, not as a byproduct.
The model's training also follows a different path. Instead of training on millions of photographs —which is the industry standard— Canva trained Craft on human templates from its own library: social media posts, presentations, infographics, menu cards, flyers, and typography-rich layouts. The effect is that the model learns not only visual aesthetics but the underlying logic of professional design: alignment, spacing, hierarchy, and how elements relate to one another.
For Canva, this decision has financial implications just as significant as technical ones. The platform operates with 240 million monthly users and near US$3 billion in annual revenue, but had relied on third-party models for image generation capabilities since the beginning. That meant paying significant API costs and depending on competitors —Google, OpenAI, Meta, ByteDance— who all have image generation tools. With Craft, Canva takes full control of the supply chain: quality, speed, and costs are now internal.
The company has already announced that it is developing faster and more cost-efficient versions of the model —called Craft-fast and Craft-ultrafast— that promise to generate high-quality images at a fraction of the cost of frontier models. This is consistent with a broader industry trend: while frontier models continue to scale in size and complexity, specific applications like design benefit from lean, use-case-optimized versions rather than oversized generic models.
Craft also integrates strategically into Canva's existing ecosystem. Generated images can be converted into layered, editable designs through Canva Magic, the image-to-design transformer launched the previous year. This bridge between generation and editing is exactly what differentiates Craft from competing tools: while tools like Midjourney deliver a static ready-made image, Craft delivers a living project that can be edited, resized, rearranged, and refined —all within the same tool.
The company also plans to integrate Craft via MCP into external AI assistants like ChatGPT and Gemini, enabling AI agents to use image generation as one step within larger workflows —for instance, an agent that creates a launch graphic, a product shot, or a social media post as part of a broader task.
On the DesignArena ranking, which measures image generation quality through crowdsourced evaluations by designers, Craft debuted at fourth position overall among all image generation models —positioning Canva Research as the second-most advanced lab in image generation by image. The model also ranked first in People & Portraits (among 87 models) and Landscapes & Nature (among 85 models), and second in Abstract & Patterns.
Craft's launch comes at an interesting moment for the creative tools industry. Canva cut its 2026 growth projections, leading analysts to reduce its valuation from US$42 billion to approximately US$35 billion. Nevertheless, the company continues to invest consistently in artificial intelligence as its primary competitive differentiator —Craft is the latest and most ambitious of those investments, and the first generative image model developed internally by the company. Researchers from Leonardo.ai, which Canva acquired in 2024, contributed directly to the model.
What remains to be seen is whether this planning-before-rendering architecture can generalize beyond image generation —and whether other creative sectors, such as video and 3D, will adopt the same reasoning. Craft demonstrates that, for certain applications, the secret is not about making the model "imagine" more pixels, but about making it think more about what it's going to do.
Sources: The Verge, Canva Newsroom, TechCrunch
✓ Independent sources cross-checked and verified before publishing
