When a software company that builds on other firms’ AI models decides to train its own, the reasons say a lot about where the technology is heading. On October 22, design platform Canva released Canva Craft, its first image generation model, to all users. This case study looks at what Canva built, how it made the model different from general image generators and what other product teams can learn. It is based on reporting by TechCrunch and The Verge and on Canva’s own claims, which we have not independently tested.
The situation
Canva is a design platform that has relied on several third-party image models inside its product. TechCrunch notes that using many external models costs money, and that a model of its own is likely to help the company save on API costs, while also giving it more control over quality and speed. The company said as much, according to the report: beyond costs, the model gives it better quality and speed control.
The business backdrop matters too. TechCrunch reports that Canva was on track for $4 billion in revenue last year, but cut its growth forecast for 2026, and that outside analysts reportedly saw its valuation fall from $42 billion to $35 billion. Those valuation figures come from external analysts, not from Canva.
What Canva built
Canva says Craft works in two stages. First, it interprets the prompt and makes a plan that includes how elements relate in space, their relative scale, visual hierarchy and information needed for any text. Second, it generates an image of up to 2K resolution in more than 30 aspect ratios for different platforms. The Verge’s summary says the planning step allows “greater precision, consistency and control.”
Canva argues this approach produces accurate object details and legible text, and lets users edit precisely without changing other parts of the image. Text rendering is a long-standing weak point of image generators, so this is the claim to test.
The key choice: training data
The most distinctive decision is in the data. According to TechCrunch, Canva trained the model on designs, using human-made templates from its own library, such as posters, social posts, presentations, infographics, menus and typography-heavy layouts, instead of photographs. The Verge cites “millions of human-made templates.”
That has a logical consequence. Because the training data are structured designs, images produced by Craft could be converted into layered designs with Canva Magic, the company’s image-to-design tool, according to TechCrunch. The model fits the product, not the other way round. Canva said researchers from Leonardo.ai, which it acquired in 2024, contributed to the work.
Distribution
Craft is available to all users on the platform, reachable from the Dream Lab tab. It will also be available through Canva’s API, and Canva plans to support image generation in third-party chatbots such as ChatGPT and Gemini through the Model Context Protocol (see our glossary entry on MCP). Two cheaper versions, called Craftfast and Craft-ultrafast, are planned for later release. The company says Craft includes input and output moderation, privacy and safety controls and reporting tools.
How it measures up
The Verge reports Craft ranks fourth on Design Arena’s leaderboard for AI image generation, a crowdsourced benchmark for designs, behind three OpenAI models: GPT-Image-2.5 Sunburst, GPT-Image-2.5 Flare and GPT-Image-2. Fourth place on a public leaderboard from a first-generation model is a solid debut, though crowdsourced rankings reflect voter preferences and shift as models are added.
Lessons for product teams
- Own the model when it is a core cost or control point. Canva generates a huge number of images; reducing per-image cost and latency has compounding value, as our explainer on how AI pricing works shows for token-based billing.
- Domain data can beat general data. Training on templates aligns the model with what users actually make, such as readable posters, instead of generic art.
- Planning before generating improves control. Splitting a task into plan and render is a pattern seen across AI systems, from images to code.
- Keep outputs editable. Layered output through Canva Magic suits users who want to adjust a result, not regenerate it.
- Use standard protocols. Offering the model through MCP and an API extends reach beyond the product.
Open questions
Canva has not published, in the coverage we reviewed, a technical report with details of model size or evaluation. It is unclear how the company licensed or consented the human-made templates used for training; we have no information on that point. And claims about text accuracy and precise edits are the company’s; independent tests will show how well they hold up. You can try the model in Canva yourself, and we would suggest testing prompts with long text and specific layouts first.
What to watch next
- Release of Craftfast and Craft-ultrafast, and their pricing and quality trade-offs.
- Independent comparisons of text accuracy against other leading image models.
- Movement on Design Arena’s leaderboard as new entries arrive.
- Details of how Canva sourced and credits the template data.
- Whether the MCP integration reaches third-party chatbots and how it is priced.
Key takeaways
- Canva Craft is Canva’s first image model; it plans a layout before generating an image up to 2K in over 30 aspect ratios.
- It was trained on human-made design templates, not photographs, which suits layered, editable output.
- A model of its own can cut API costs and give Canva control over quality and speed.
- It ranks fourth on Design Arena’s leaderboard, behind three OpenAI models.
Frequently asked questions
Who can use Canva Craft?
All Canva users, through the Dream Lab tab, according to the reports. It is also planned for the API and, through MCP, for third-party chatbots.
Why train on templates rather than photos?
Templates contain structured layouts, text and hierarchy, which helps the model produce legible, editable designs and convert images into layers.
Is it better than other image generators?
On Design Arena’s crowdsourced leaderboard it ranks fourth, behind three OpenAI models. Results vary by task.
