Canva spent years collecting millions of design templates, and this week it launched Craft, an AI image generator whose training data includes them and that debuted fourth on a public leaderboard (Canva). The launch raises a question for any software company: could the expert work your product has collected improve a model for one specific task?

Craft writes a layout plan before it renders the image

Text inside images has been a weak spot for many image models. Canva said Craft works in two stages: a planning stage writes out the composition, including which elements appear, where they sit, how large they are, and what text goes in, and a rendering stage then produces the image. Canva said Craft writes text in 55 languages and produces more than 30 aspect ratios.

Canva said the training data includes millions of human-made templates from its content library, among them posters, menus, social posts, and presentations, which show how designers arrange headlines, pictures, and spacing. A photograph shows only the finished result. A template lists its parts, each with a position, size, and text, which may help explain why templates teach layout well. Canva said its professional designers gave feedback during training and compared more than 25,000 pairs of images against leading models in its own evaluations.

Canva needed years and a research lab to get here

Canva Research has more than 160 scientists and engineers. Canva bought Leonardo.ai in 2024 for its image model expertise, and people make more than a billion designs a month on the platform, a figure that covers all designs and not only AI images. Stef Corazza, Canva's head of AI research, wrote that “even a variation of a few cents per generation makes an enormous difference” at that scale.

I argued in April that Canva's cost structure is its strategy. Canva describes Craft as the first foundation model built and trained from scratch by Canva Research.

Customers should test whether Craft saves editing time

The potential upside appears in the first draft: words in the right place, many aspect ratios, and Magic Layers, a separate Canva tool that can split a Craft image into pieces you can move and edit, so designers adjust the image instead of rebuilding it. Canva said Craft itself produces an image, and Magic Layers does the splitting afterward.

Generation time and price decide the budget. WebProNews, citing Canva's developer forum, reported that a Craft image takes about 50 to 60 seconds through Canva's API and uses AI credits (Deffry). A minute is a long wait for a quick thumbnail and may suit a finished layout, depending on how much editing the image saves. Canva announced two faster variants, Craft-fast and Craft-ultrafast, and prices and dates are still to come.

Craft debuted fourth, behind three OpenAI models

DesignArena is a public site where visitors vote between two anonymous images. WebProNews reported that all three models ahead of Craft come from OpenAI, and SmartCompany reported that Craft placed ahead of entries from Google, Microsoft, xAI, and ByteDance. Canva reports first place among 87 models for People & Portraits.

The ranking reflects voter preferences and does not show how much editing a team needs before approving an asset. Canva's claims about layout and typography rest on tests by its own designers.

Adobe builds on licensed content and helps customers train their own models

Adobe, a major competitor in design software, has its own image models, called Firefly. Its approach to training data differs from Canva's.

Adobe says it trained its Firefly models on licensed content, including Adobe Stock, and public domain material. The announcements emphasize different things: Canva describes layout planning and template training, while Adobe emphasizes licensed sources and customer-specific models.

Adobe's Foundry service, announced in October 2025 as Adobe AI Foundry, retrains Firefly on a customer's own material. VentureBeat reported that the customer owns the output, its data stays out of Adobe's base model, and The Home Depot and Walt Disney Imagineering were early customers. Firefly Custom Models, in public beta since March 2026, let customers train on their own images to match a style.

Firefly also offers more than 30 image and video models from Adobe and other companies, including Google and OpenAI (Adobe). Canva uses partner models too, and Craft is available to outside developers through its API.

Some company data can teach a model one narrow job

Canva and Adobe show two ways to put a company's own data to work. Many software products record how experts do their work, and a company that can organize those records and has the right to use them may be able to train a model for one job. Four questions test whether your company, or a vendor you depend on, has that asset:

• Does the data show how people reached a result, or only the final product?
• Is it organized well enough for a model to learn from?
• Does the company have the right to train on it?
• Would better quality, control, or operating cost justify training and maintaining a model for one task?

A sales software vendor might examine whether outreach records linked to outcomes could improve a narrow task, and an accounting vendor might do the same with categorized transactions.