Canva has shipped its first in-house image generation model. It’s called Craft, it was built by Canva’s own research team, and it’s live for every Canva user right now in the Dream Lab tab. That’s free and paid accounts alike. The model is also coming to the Canva API, and Canva says it plans to let people generate images with it from inside ChatGPT and Gemini through its MCP connector.
On paper Craft is a challenger to the image models from Google, OpenAI, ByteDance and Meta. In practice it’s aimed somewhere else. Those models learned mostly from photographs. Craft learned from layouts.
Trained on Posters and Menus
The training set is the interesting part. Canva didn’t feed Craft a pile of photos scraped from the web. It used the human-made templates in its own library: posters, social posts, presentation slides, infographics, restaurant menus and type-heavy layouts. Years of people designing things for other people to read.

A beach volleyball player lines up her serve beside the net post, the ball resting on an open palm and her other hand drawn back at the wrist, while the surf breaks white behind the court. A real photograph of a real moment, which is exactly the kind of picture Craft was not trained on.
That choice shows up in what the model is good at. Ask a photo-trained generator for a flyer and you usually get a nice picture with mangled lettering somewhere in the middle. Text has been the weak spot of image AI since the start. A model that has seen millions of menus and event posters has a much better idea of where a headline sits, how big the date should be and what goes in the corner.
Craft also works in two steps. First it reads the prompt and draws up a plan, covering where each element sits, how big things are relative to each other, which parts should grab the eye first and what the text says. Then it renders the image, at up to 2K resolution and in about 30 aspect ratios matched to the usual platforms. Canva says the planning step is why the lettering comes out legible and why you can change one part of an image without the rest of it shifting.
Because the output starts life as a design, Canva Magic can turn a Craft image into a layered Canva file. That’s the image-to-design tool Canva built on the design model it released last year. So the generated poster isn’t a flat picture you have to live with. It splits back into text boxes, shapes and images you can move around. For anyone who has tried to fix one typo in an AI-made graphic, that’s the feature that matters.
Why Build Its Own
Dream Lab isn’t new. Canva launched it in 2024, a few months after buying Leonardo.ai, and it ran on Leonardo’s Phoenix model. Canva has leaned on a mix of outside image models across its apps too. Researchers from the Leonardo team worked on Craft, so in a sense the acquisition is finally paying out as a home-grown model.
Money is a big part of the story. Every image a Canva user generates on someone else’s model is an API bill. Canva says Craft gives it more control over quality and speed as well as cost, and it’s already working on faster, cheaper versions of the model that it says are coming soon.
The timing fits. Canva was on track for around $4 billion in revenue last year, then cut its growth forecast for 2026. Outside analysts reportedly marked its valuation down from about $42 billion to $35 billion after that. AI costs were part of the reason. Owning the model is how you stop paying rent on it.
The Chatbot Angle
The MCP plan may end up mattering more than the Dream Lab launch. Plenty of people now start a design job by typing into ChatGPT or Gemini, not by opening Canva. If Craft can answer those requests from inside the chatbot, Canva gets its model in front of people who never visit canva.com, and every result can be handed off to Canva as an editable design.
That’s a different bet from the big labs. Google and OpenAI want the best general picture maker. Canva wants the one that makes the thing you were going to print, post or put on a slide, with words that are spelled right and layers you can fix.
Whether templates beat photographs as training material will come down to the outputs, and people will judge them fast. Canva has more than a decade of users who know exactly what a good flyer looks like. They’ll tell it.