Canva has built its own image model, Canva Craft, and made it available to all Canva users through the Dream Lab tab. Developers can reach it through Canva's API. The unusual part is what it learned from: human-designed templates such as posters, presentations, infographics and menus, not photographs.
A menu is a layout before it is a picture
Most image generators learn from enormous collections of photographs and artwork. They are good at faces, landscapes and light. Ask one for a poster with a headline, a date and three lines of small print, and it has to guess where the words go and how large they should be, because photographs rarely teach that.
Craft splits the job in two. In the first step it reads the request and writes a plan: where each element sits, how big it is compared with the others, which one the eye should see first, and what the text should say. In the second step it draws the image, at up to 2K resolution and in about 30 shapes, from a square post to a tall story to a wide banner. A designer does the same when they sketch boxes on paper before opening any software. Because the plan is a separate step, a change to one part does not have to redraw everything.
That is also why training on templates fits. A template is a finished layout that someone already made well, so the model sees what good hierarchy looks like in the exact formats people ask for.
What it changes for people who make their own marketing
Imagine a shop owner who needs a sale announcement as an Instagram post, a story and a printed A4 sheet. With a photo-trained generator, she typically gets a nice picture, then fixes the text somewhere else, then repeats the work for each shape. With a layout-aware model she can ask once for each shape, get something near finished, and spend her time on the part only she can judge: whether the offer reads right and the price is correct.
Canva already had the next step in place. Magic Layers turns a flat image into separate pieces, with text you can retype and elements you can move. Craft's images are meant to go straight into it, so a generated poster does not end up as a picture you can only re-request. I would expect the second half of that to matter more than the first. The usual time sink with AI images is not the first draft but the fourth change to it.
Why more software will train on its own users' work
Until now Canva's image tools ran on a mix of models, including ones from Leonardo, the start-up it bought in 2024. Owning the model lets Canva control quality, speed and what each image costs it. It also gives Canva a way into the chatbots people already use: the company plans to bring Craft's image generation into ChatGPT and Gemini through the connector it already offers there.
The larger point reaches beyond Canva. A company that holds years of finished work, such as slides, ads, invoices or floor plans, has something a general model lacks: examples of what "done" looks like in that field. Expect more tools to train their own models on that kind of material. The result for the reader is that the AI inside the tools you pay for will start to know your formats, not just draw pictures.