IA · 23 October 2026 · 4 min read
Canva Takes on AI Giants: Craft Debuts to Plan Layouts Before Rendering Pixels
In brief: Canva has introduced Canva Craft, its first image generation model developed entirely by its in-house research team. Moving away from standard text-to-image systems, the model uses a two-stage approach that plans spatial layout, visual hierarchy, and typography before rendering up to 2K resolution across more than thirty aspect ratios. This structure allows generated designs to be converted into fully editable layers rather than static, unalterable flat images.
by Team Mocchi's
For years, digital design platforms have relied on external AI vendors to provide generative capabilities to their user bases. Canva is now charting a distinct course with the launch of Canva Craft, its first proprietary image generation model built from the ground up by its internal research division. The move is aimed squarely at competing with flagship models from OpenAI, Google, and Meta, approaching the medium through the lens of structured design rather than unconstrained pixel synthesis.
A Two-Stage Pipeline: Planning Before Rendering
The fundamental innovation behind Canva Craft is its distinct two-stage execution architecture. When an operator inputs a prompt, the system does not immediately start drawing pixels. Instead, the first phase parses user intent and generates a formal layout plan, mapping out spatial relationships, element proportions, visual hierarchy, and typographic parameters.
Only in the second phase does the rendering engine synthesize the visual output, delivering images at resolutions up to 2K across more than thirty different aspect ratios tailored for print and social media formats. As reported by The Verge, this staged workflow resolves long-standing issues in generative imagery, such as warped text rendering and spatial incoherence, while making it possible to modify individual scene components without regenerating the entire asset.
From Flat Pixels to Editable Layers
Canva Craft also diverges from conventional models in its underlying training data. Rather than crawling generic web photography, Canva trained the model on structured templates produced by human graphic designers across its platform catalog, encompassing corporate presentations, infographics, restaurant menus, promotional flyers, and typography-heavy compositions.
Because the model understands semantic structure, generated assets can be directly imported into Canva Magic and converted into layered designs. Users receive an editable composition where background art, icons, and text boxes remain discrete, reconfigurable entities. According to TechCrunch, Canva Craft is immediately accessible to all users through the platform's Dream Lab, will be exposed via the Canva API, and is slated for integration with third-party conversational assistants such as ChatGPT and Gemini via the Model Context Protocol (MCP).
Escaping Third-Party API Dependency
The introduction of an in-house model is equally a defensive financial maneuver. Serving generative features to a user base exceeding 190 million monthly active users on top of third-party APIs incurs staggering inference expenses over time.
By leveraging the foundation acquired through its 2024 purchase of Leonardo.ai, Canva achieves operational sovereignty over inference latency, unit economics, and moderation guardrails. The release illustrates a broader industry maturation: vertical software platforms are moving away from merely wrapping third-party foundational models, choosing instead to train specialized architectures tailored to the exacting demands of their domain workflows.
Mocchi's take
Canva's architectural choice highlights a critical turning point for enterprise software: in real production environments, static flat assets produced by generative prompts often create more manual rework than they save. Deconstructing visual generation into a planning phase followed by layered rendering proves that practical AI adoption is about structural integration rather than raw parameter count. For engineering teams building enterprise tools, leveraging open interoperability standards like MCP to connect structured generative engines into operational workflows represents the most viable path toward automated, brand-compliant digital production.