Canva has officially released Canva Craft, its first in-house image generation model. Developed by the company’s internal research team alongside talent from Leonardo.ai, which Canva acquired in 2024, the new model positions the design platform to compete directly with generative systems from Google, OpenAI, Meta, and ByteDance.
Rather than relying exclusively on third-party AI providers, Canva’s proprietary model is tailored to meet the specific requirements of graphic design workflows, with an emphasis on compositional structure, text legibility, and post-generation editability.
A Two-Stage Synthesis Pipeline
Standard diffusion models typically generate an entire image in a single pass, which can lead to common issues such as garbled typography, disproportionate elements, and haphazard layouts. Canva Craft addresses this challenge using a two-stage generative process.
In the first phase, the model interprets the user prompt and generates a structured composition plan. This blueprint establishes spatial relationships, relative sizing, visual hierarchy, and designated text areas. In the second phase, the model executes the plan, rendering images at resolutions up to 2K across 30 supported aspect ratios suited for common social, print, and presentation formats.
This separation of structural planning from pixel synthesis allows for more precise details and enables localized adjustments without requiring a complete regeneration of the image.
Training on Design Data Instead of Photographs
A key distinction of Canva Craft is its training methodology. While conventional image generation models are largely trained on large-scale collections of standard photographs and digital art, Canva trained its model primarily on structured, human-designed templates from its existing library.
The training dataset encompassed typography-heavy layouts, posters, social media posts, multi-page presentations, menus, and infographics. Because the model understands foundational graphic design principles, outputs align closely with standard layout practices.
In addition, this architecture allows images generated by Canva Craft to integrate with Canva Magic, the platform’s image-to-design tool. Flat outputs can be converted into layered, editable project files where individual elements, such as text fields and shapes, remain fully adjustable.
Distribution and Model Variants
Canva Craft is available to all platform users through the Dream Lab interface and also available via the Canva API. The company intends to support generation across third-party AI assistants, such as ChatGPT and Gemini, utilizing Canva’s Model Context Protocol (MCP) integrations.
To address different latency and compute requirements, Canva is developing two optimized versions alongside the flagship model.
| Model Variant | Key Specification | Primary Application |
| Canva Craft | Full fidelity, up to 2K resolution | Detailed visual layouts, print assets, and typography-rich graphics |
| Craftfast | Balanced speed and compute overhead | Routine design iteration and day-to-day web content |
| Craft-ultrafast | Low latency, optimized throughput | High-frequency prompts, quick previews, and chatbot integrations |
Strategic Implications and Cost Control
The rollout of Canva Craft reflects an important operational shift for the company. Historically, Canva has integrated several external models to power its generative features. Transitioning to in-house architecture offers the platform greater autonomy over service latency and generation quality, while simultaneously reducing third-party API expenditures.
The technological shift also arrives amid changing market conditions. Although the company approached $4 billion in revenue over the past year, adjusted growth forecasts for 2026 led independent analysts to revise Canva’s estimated valuation from $42 billion down to $35 billion. Owning its foundational infrastructure provides Canva with improved margin efficiency and proprietary technical assets as enterprise software markets continue to evolve.
